The next patient, the next update.
Software can never be tested on every path, and an AI model learns only from the data its makers chose. What evidence shows either will give the right answer for the next patient, and what keeps that true after an update?
Four hours of training.
The people holding the camera had attested that they had never taken a picture of the inside of an eye. They worked in ten primary-care offices spread across the United States, and each had been given one standardized training session of four hours, with no refresher afterward. Their task was to photograph the retinas of adults with diabetes enrolled at those offices: two pictures of each eye on a Topcon NW400, one centered on the optic disc, where the optic nerve leaves the eye, and one on the fovea, the small pit at the center of sharp vision. About three participants in four needed no drops to widen the pupil. The other quarter were dilated and photographed again.
The pictures went to a program that had been locked before the first participant arrived: its decision rules were fixed and did not change during the trial. Within seconds it returned one of three answers: more than mild diabetic retinopathy detected, not detected, or image quality insufficient. Diabetic retinopathy is damage to the blood vessels of the retina caused by diabetes, and it can threaten sight. Nobody at the clinic interpreted the photographs. Between January and July 2017, 900 people enrolled.
Every participant who completed the study was photographed a second time, by photographers certified by the Wisconsin Fundus Photograph Reading Center, with a far more demanding protocol: four overlapping stereo pairs of each retina, and a cross-sectional scan by optical coherence tomography of the macula, the central area of the retina. Three experienced readers graded the photographs without seeing the program's answer, and their majority decided which participants the trial would count as having the disease; the macular scans were graded separately. With input from the FDA on its endpoints, the trial's designers had fixed its test before it began. They planned the number of participants around an expected sensitivity of 85 percent and specificity of 82.5 percent, and the program would pass only if the lower end of each result's confidence interval cleared a floor: 75 percent for sensitivity and 77.5 percent for specificity (Chapter 13). Sensitivity is the share of participants with the disease that the program flags; specificity is the share without the disease that it clears.
Of the 892 participants who completed the study, 819 had reference images that could be graded and a usable answer from the program. On the reference images, 198 of the 819 had more than mild disease, and the program flagged 173 of them, or 87.4 percent. It cleared 556 of the other 621, or 89.5 percent. Among participants whose reference images could be graded, it gave a usable answer for 96.1 percent. The developer, IDx, funded the trial. The trial's first author was Michael Abràmoff, an ophthalmologist at the University of Iowa who had founded the company, and the paper discloses the funding and his financial ties to IDx. On 11 April 2018 the FDA granted the system a De Novo authorization, its route for a new kind of device of low or moderate risk with nothing comparable on the market. The FDA described it as the first device authorized to provide a screening decision "without the need for a clinician to also interpret the image or results."
The trial measured one locked program, run by those operators on that camera, for those patients. It said nothing about the same program in a clinic whose computers send images differently, after the camera is replaced or the model retrained, or when someone sets out to make it fail. A device that will be used for years on people who were never in its trial needs evidence of a different kind, and each later version needs more of it.
Before we start.
Software in a medical device does what it was written, or trained, to do, on whatever input arrives. A trial like the one above shows how one program behaved for 819 analyzable people. A maker, a regulator and a hospital need something more: a reason to believe the next person will get the right answer too, including people the trial never saw, and that the reason will survive the next release. That is the question this guide answers: software can never be tested on every path, and an AI model learns only from the data its makers chose; what evidence shows either will give the right answer for the next patient, and what keeps that true after an update?
Four terms carry the argument. A device software function is software that meets the legal definition of a medical device, whether it runs inside a pump or on a server by itself. An algorithm is the procedure the software follows. In this guide, an AI model is an algorithm whose decision rules were fitted to example data rather than written line by line. And evidence means records that someone outside the team could check: requirements, test results, study data, monitoring reports.
The answer comes in four parts. Testing samples what software does and can never cover all of it, because even a modest program has more paths, and may meet more inputs, than any test campaign can try. Confidence therefore comes from how the software was built: requirements derived from the harm a fault could do and traced to tests, an architecture that keeps the dangerous part small enough to test thoroughly, a full account of the code someone else wrote, and proof that each change broke nothing already verified. Agile teams and automated pipelines can produce that evidence, provided each increment leaves its records.
An AI model adds a second limit. Its accuracy is a measurement on a test set. That measurement holds only for patients like those in the set, scored against the reference someone chose, and only for the exact model that was locked and fed the inputs it was specified for. The evidence that the model will be right for the next patient is a test kept apart from training and drawn from the setting of use, results reported for each group of patients with their uncertainty, a check at each hospital that the specified conditions hold, and monitoring once the model is in use.
Someone can also make a device fail on purpose. New ways of attacking it appear while its code stays the same, so a device must be built to be patched and watched for years, starting from a list of the components inside it. What keeps all of this true after an update is change control: every change is judged against the claims it could move, and, for AI, a plan agreed with the regulator in advance says what may change and how the new model must be tested.
One fact shapes every chapter. Software does not wear out. A pump's motor and a camera's sensor age and fail at rates that can be measured; software does not age at all. Every way it can fail was built in on the day it shipped or arrived with an update, and it waits there for the input that triggers it. For an AI model, the training data is part of what was built.
How to read this guide
The guide runs through 29 chapters in eight parts. Parts I and II explain why software fails differently from hardware and what a software lifecycle does about it, including when much of the code comes from elsewhere and when the team ships every two weeks. Parts III and IV turn to models learned from data: how they are built and scored, and why they do worse on patients unlike those they were tested on. Part V follows a model into a hospital and through its later versions, and its last chapter, Chapter 21, closes with the answer to the central question. Part VI covers attack and defense, Part VII surveys the products and makers of October 2026, and Part VIII sets out what the FDA, the EU and India require, as of that month. IDx-DR is the guide's running case and returns throughout, and the recalled insulin pumps return in the security chapters. A reader who builds devices can start with Chapters 4 to 9, 21, 23, 27 and 28; a reader who buys or deploys them, with Chapters 17 to 20, 24 and 29. Every chapter works through one example with real numbers. Plates are numbered in their own sequence, so that each can be cited by its number alone. These boxes recur:
Blue edge. Carries the structural point of a section, or a calculation worked once with real values and stated assumptions.
Orange edge. Names a common misreading, a trap, or the limit of a claim.
Green edge. Maps the idea onto the reader's own work: specifying, building or testing device software, evaluating an AI product, deploying it in a hospital, or planning the evidence for a submission.
- Each chapter closes with what it established, in four lines.
No formula needs more than arithmetic. Proportions such as sensitivity are given as percentages, and uncertainty as a range. Three calculations size a test, each worked once with numbers: a run of failure-free trials puts an upper bound on a failure rate (Chapter 3), the precision wanted from an estimate sets how many patients a study needs (Chapter 11), and the power wanted from a test against a floor sets how many a local check needs (Chapter 19). Counting paths uses doubling, and large numbers are written in words or in full, such as a billion or 1,000,000,000. The guide uses US spelling.
Colors in the plates follow a fixed key. Requirements, claims and acceptance criteria are gold; software the maker wrote, and a model running as software, is teal, with third-party code hatched in the same teal; data of every kind, from training sets to the images a device receives, is violet; an attacker, an attack path or a vulnerability is red; and a defect, a wrong output or a drift in performance is magenta. The output reported to a clinician or patient is always blue, and one orange mark in each plate points at the detail that matters most. A key strip under every plate lists only the colors that plate uses.
This is a reference for understanding, not a development procedure, a regulatory opinion or clinical advice. Products, authorizations, guidances, standards and laws are stated as of October 2026 and will date; several of those named here were drafts or under revision in that month, and the text says which. Where a figure comes from a company about its own product, or from a paper whose authors work for that company, the text says so. Named companies and products are examples of a principle, not recommendations. Cost, review times, fees and make-or-buy decisions are outside the guide, although the chapters on documentation, classification, test design, change plans and third-party code describe what drives them.
+ Part I · Built in on the day it shipped
Why software fails differently.
Between 1992 and 1998 the FDA counted 3,140 recalls of medical devices, and 242 of them, 7.7 percent, were caused by software. Of those 242, 192 were caused by defects introduced when the software was changed after it had first been distributed. Both counts are evidence of the same peculiarity: software is not a part that ages, so its failures are mistakes made by people, in the first version or in a later one. The three chapters of this part set out what software in a device is trusted to do, why it fails when it does, and why no amount of testing can prove that it will not.
What the code is trusted to decide.
+ The questionWhat does the software in a medical device actually do, and when is the software itself the device?
A camera, a client and a server
The system in the opening scene has three pieces, and only its software is the device the FDA authorized. A commercial retinal camera, the Topcon NW400, takes the photographs. A computer attached to the camera runs a program the maker calls the client, which lets the operator pick two images of each eye and sends them "over a secure internet connection" to a server in a data center. There a second program, the analysis, examines the images and returns one of three answers to the client screen: more than mild diabetic retinopathy detected, not detected, or image quality insufficient. The FDA set out its review in a public decision summary, which describes this architecture and names the software version reviewed in 2018, IDx-DR 2.0.0.
The camera is a separate product, outside the authorization. What the FDA classified in April 2018 was software. The FDA sorts devices into three classes by risk, from class I, the lowest, to class III, the highest. In a 510(k), a maker shows that its device is substantially equivalent to one already legally marketed, its predicate, and the FDA clears the device. A De Novo is the route for a novel device of low or moderate risk with nothing comparable on the market. The FDA grants a De Novo rather than clearing it, and in granting it creates a new device type; for a class II type it also sets special controls, requirements that every device of that type must meet. The highest-risk devices need premarket approval.
IDx-DR was authorized through a De Novo. The FDA created a new category for it, the retinal diagnostic software device, in class II, the class for moderate risk, under a regulation written for this authorization. Nothing in that device touches the patient, emits energy or moves. Its effect on health passes entirely through a decision: whether a person with diabetes is referred to an eye specialist now or told to come back in a year.
When the software is the device
Most medical devices with software carry it inside: the firmware of an infusion pump, the program that runs a ventilator's valves, the code in an imaging scanner. That software is part of the device and is judged with it. A growing share of devices are software alone, running on ordinary computers, phones or cloud servers. Regulators call these software as a medical device, and newer documents describe them simply as medical device software that stands alone. IDx-DR's analysis program is an example.
What turns code into a device is not anything in the code. It is the intended use: what the maker says the software is for, in its claims and in its labeling, which covers the label and any written material that accompanies the device. A program that draws a graph of blood glucose readings for a person's own interest is not a device in the US; the same graph with a rule that tells the patient how much insulin to take is. The US definition, the EU's rules and India's 2026 guidance differ in detail on where the boundary lies, and Chapter 26 sets those differences out. They agree on the principle: the claim makes the device, and the claim also fixes what the evidence must show.
Two companies can ship identical image-analysis code, one to help researchers count lesions in a study and one to tell clinicians which patients to refer. Only the second makes a medical claim, and only the second needs a medical device's evidence. Changing a sentence of labeling can therefore change the regulatory status of software that has not changed at all.
Inform, drive, diagnose or treat
Once software is a device, the evidence it needs rises with the answers to two questions. The first is how serious the patient's situation is: whether a wrong answer could lead to death or irreversible harm, to a serious but treatable deterioration, or to something minor. The second is how much the software's output decides: whether it treats or diagnoses directly, drives the next step of care, or only informs a clinician who still decides. Regulators from many countries and the EU agree common approaches through the International Medical Device Regulators Forum. In 2014 the Forum set out these two questions as the axes of a grid for software as a medical device. India's regulator uses the same grid in its 2026 guidance to assign device classes.
IDx-DR sits high on the second axis. Its output is the screening decision itself; no clinician interprets the image. Diabetic retinopathy threatens sight but rarely kills, and a person told to rescreen in a year will usually be seen again, so the situation is serious rather than critical. By contrast, a stroke-triage program that alerts a specialist to a suspected blocked artery informs a clinician who still reads the scan, but in a situation where minutes count. The two products sit in different cells of the same grid, and Chapter 16 returns to the difference between a program that decides and one that advises.
Take four functions. A pump's bolus calculator computes a bolus, a dose of insulin given all at once, from a glucose reading and the carbohydrates entered. It sets the treatment itself, and an error of a few units can cause dangerously low blood sugar, or hypoglycemia, so it sits at the top of both axes. A triage program examines a CT angiogram, a scan that shows the blood vessels, and flags a suspected large-vessel occlusion, a blockage in one of the large arteries that supply the brain. It informs a stroke team in a critical situation: lower on the second axis, highest on the first. A logbook app that stores and displays glucose readings without interpreting them generally falls outside the US device definition, because a US statute excludes the display of device data from that definition. A step counter for fitness is a wellness product, not a device. The same phone could run all four; the evidence each needs ranges from none to a clinical trial.
The claim is the unit of evidence
Because the claim defines the device, it also defines what can go wrong. The FDA's decision summary for IDx-DR lists three risks to health: a false positive leading to an unnecessary referral, a false negative delaying evaluation and treatment, and an operator failing to capture images good enough to analyze. Each special control the FDA required answers one of those three risks. The controls cover clinical testing; software verification and validation built on a hazard analysis, a study of how the device could cause harm; training and human-factors testing for operators; labeling; and a protocol stating which changes could significantly affect the device's safety or effectiveness. Chapter 21 shows how that last requirement anticipated the plans for changing AI models that the FDA finalized in 2024, six years later.
Both lists are short because the claim is narrow. IDx-DR is indicated for adults with diabetes who have not been diagnosed with retinopathy. It is to be used by health care providers, with images from the Topcon NW400. It does not look for glaucoma, an eye disease that damages the optic nerve, and at the time of authorization it was not labeled for patients younger than 22 or for pregnant patients. Each limit removes a population or a condition from what the evidence has to cover. The rest of this guide is about the evidence inside those limits and about what happens at their edges, which real use tends to reach.
A useful intended-use statement names the decision the software supports, the patients it applies to, the users, the setting and the inputs it accepts, including the devices that produce them. Draft it before architecture begins, and treat each change to it as a design change, because the statement fixes the class of the device, the hazards to analyze, the population a clinical study must sample, and the edges where use outside the claim is most likely.
- IDx-DR is a software device: a client and a server-side analysis that turn photographs from a named camera into one of three screening answers.
- Software becomes a medical device through its intended use, not through anything in its code, so a change of claim can change its status.
- The evidence a software device needs rises with how serious the patient's situation is and with how much the software's output decides.
- The claim defines the risks and the controls; narrow indications shrink what the evidence must cover, and real use tends to test the edges.
Built in on the day it shipped.
+ The questionIf code never wears out, where do its failures come from?
Four conditions at once
In May 2024 Hamilton Medical began correcting the software of its HAMILTON-C6 ventilators, and the FDA classed the action as the most serious type of recall. The fault needed four things to happen together. A clinician pressed the key that briefly raises the oxygen supplied and disconnected the breathing tube to suction the patient's airway. During that disconnection a sensor error occurred, for example because the tubing of the flow sensor was kinked. The ventilator entered its sensor-fail mode. And the patient was reconnected while that mode was still active. In that sequence, and only in it, ventilation might not restart. The FDA's notice lists one injury and one death.
Each of the four events is ordinary in an intensive-care unit. Their combination is rare. The fault that the combination triggered was present in three versions of the software, 1.1.4 to 1.1.6, and in every ventilator that ran them, until version 1.2.3 removed it. No part had worn and no component had drifted. The ventilators that could fail this way were exactly as they had been on the day the software was installed.
A fault waits for its input
In 2002 the FDA issued a guidance on software validation. A guidance states a regulator's current thinking on how to meet the law; it does not bind a maker, but it describes what reviewers expect to see. This one is still in force apart from one section that was replaced in 2025. It puts the difference plainly: "Unlike hardware, software is not a physical entity and does not wear out." A hardware part fails for physical reasons that accumulate: a bearing wears, a capacitor dries, a solder joint cracks under thousands of heating cycles. Its failures are random in the engineering sense: each unit fails at its own moment, and a rate per hour of operation describes the population well.
Software fails only when an input reaches a fault: a flaw in its logic or data that has been there since that version was built. Engineers call these failures systematic. They occur every time the same conditions recur, in every copy of the same version. The time to failure depends on when the triggering input arrives, which depends on how the device is used, not on how long it has been running. A rarely used function can carry a fault for years before any input triggers it, and a busy hospital can trigger it on the first day.
A version that has run in the field for years without a reported failure has shown that the inputs it met so far did not reach a fault. A new hospital, a new workflow or a change in another device can supply the input that the earlier users never did. Field history is useful evidence about the common paths and weak evidence about the rare ones.
Why reliability arithmetic does not transfer
The difference between random and systematic failure changes how safety is argued. For hardware, two independent parts that each fail once in many thousand hours rarely fail in the same hour, so a second part fitted as a backup makes a failure of both far less likely than a failure of either one. For software, two copies of the same program fed the same input fail together, so copying the code adds nothing against its own faults.
Suppose a pump motor fails at random at a rate of 1 in 100,000 per hour, and a second, independent motor is fitted as a backup. The chance that both fail in the same hour is 1 in 100,000 multiplied by 1 in 100,000: 1 in 10,000,000,000. Now suppose the pump's control program has a fault that a particular input triggers once in 100,000 hours of use, and a second processor runs an identical copy as a backup. The input that reaches the first copy reaches the second, so the chance that both fail in the same hour is still 1 in 100,000. The two copies are not independent, and independence is what made the hardware redundancy work. Protection against a software fault has to come from a design that is different or from a check that does not depend on the same code, which is the subject of Chapter 5.
What do such faults look like? Researchers at NIST, the US National Institute of Standards and Technology, studied the 383 software-related device recalls that the FDA recorded from 1983 to 1997, and they could assign a type of fault to 342 of them. Of those, 43 percent were logic faults, such as a wrong condition or an unhandled case, and 24 percent were calculation faults; the rest were spread across data, requirements, timing, interfaces and other types. None is a matter of wear. Each is a decision written into the program that was wrong for some input. The researchers also found that the recalled devices had caused no reported deaths or serious injuries, so the figures describe how software fails, not how often it harms.
A change is a new chance to fail
If faults are built in, the moment of building matters, and building does not stop at the first release. In the FDA's count from 1992 to 1998, 192 of the 242 software-related recalls, 79 percent, came from defects introduced by changes made after the software was first distributed. A fix or a new feature interacts with parts of the program that were already verified, and anything it breaks in those parts is a new fault.
In March 2024 Tandem Diabetes Care recalled version 2.7 of its t:connect app for iPhone, which works with the company's t:slim X2 insulin pump. The FDA's notice describes the mechanism. The app could crash and be relaunched by the phone's operating system again and again, and the repeated relaunches caused excessive Bluetooth communication with the pump, which drained the pump's battery until the pump shut down early and insulin delivery stopped. The recall covered about 86,000 installed apps, and by mid-April 2024 the company had received 224 reports of injury. The FDA recorded the root cause as a software design change. The pump had not changed at all; a new version of the program that talked to it had.
The two halves of the central question start with these two sources of failure. A program that was right for one set of inputs can be wrong for another that arrives later, and a program that was right can become wrong when it is changed. Chapter 3 asks why testing alone cannot rule out either failure, and Part II describes what the software lifecycle does instead.
When a software failure is reported, assume the fault is present in every unit running that version and in every version that shares the code. Reconstruct the exact sequence of inputs and write it as a test that fails on the faulty version. Keep that test permanently in the regression suite, the set of tests rerun after every change to confirm that nothing already verified has broken, so that no later change can bring the fault back unnoticed.
- Software faults are present from the moment the code is written; a failure occurs when an input reaches one, in every copy of that version.
- Software failures are systematic, so the time to failure depends on use, and years without failure say little about rare paths.
- Identical copies of a program fail together, so redundancy that works for hardware does not protect against a software fault.
- Changes are a major source of new faults: 79 percent of the software recalls the FDA counted in 1992 to 1998 came from defects introduced by changes after release.
More paths than seconds.
+ The questionWhy can't a team simply test every case before release?
Thirty decisions, a billion paths
A routine with 30 independent yes-or-no decisions has 1,073,741,824 distinct paths through it, because each decision doubles the number of paths that came before. A device program has thousands of decisions, many of them inside loops that can run any number of times, and each path can be taken with different data values and at different moments relative to other events. The number of distinct behaviors is therefore not just large. It is too large to count.
Suppose a program had only 100 independent decisions and a test rig could run a billion paths a second. Running every path once would take about 40 trillion years, nearly 3,000 times the age of the universe. Faster computers do not change the conclusion, because each extra decision doubles the work. The FDA's software validation guidance states the consequence without arithmetic: "Except for the simplest of programs, software cannot be exhaustively tested," and "path coverage is generally not achievable."
Presence, not absence
The computer scientist Edsger Dijkstra drew the conclusion that shaped software engineering. In his 1972 Turing Award lecture he said that program testing "can be a very effective way to show the presence of bugs, but is hopelessly inadequate for showing their absence." A test that passes shows that one path, with one set of values, gave the expected answer. It says nothing directly about the paths not taken.
Dijkstra's limit does not make testing useless. It makes testing a sample, and like any sample its value depends on how it was drawn. A test campaign chosen at random from the space of possible inputs would almost never reach the rare combinations that matter, such as the four simultaneous conditions of the HAMILTON-C6 recall in Chapter 2. A campaign chosen from what the software is required to do, and from the ways it could cause harm, reaches far more of them. The FDA's guidance states that testing alone "cannot fully verify that software is complete and correct." Part II therefore describes the rest of the evidence that a software lifecycle produces.
A coverage tool reports which statements of the code a test campaign executed, or which branches, the alternative outcomes of each decision. A campaign can execute every statement and still miss a wrong condition, because executing a line with one value does not test it with others. The FDA's guidance calls statement coverage "insufficient to provide confidence" on its own, and even full branch coverage leaves most paths unexercised.
How many failure-free runs prove a rate?
Statistics gives a precise answer to a narrower question: if a program runs many times without failing, how high could its failure rate still be? The rule of three answers it. When no failures are seen in a number of independent trials, the one-sided 95 percent upper confidence bound for the failure probability is about three divided by that number. That bound is the highest failure probability at which that many trials would still all succeed at least 5 percent of the time.
A team runs a function 300 times under realistic conditions and sees no failure. By the rule of three, the true failure probability could still be as high as 3 in 300, or 1 percent, at 95 percent confidence. To bring the bound down to 1 in 1,000, the team needs about 3,000 failure-free runs; to 1 in 1,000,000, about 3,000,000. A target used for life-critical aviation software is 1 failure in 1,000,000,000 hours. To show that rate, the team would need about 3,000,000,000 failure-free hours of realistic operation, which is about 342,000 years. NASA researchers Ricky Butler and George Finelli made the same point in 1993. Their target was 1 failure in 1,000,000,000 ten-hour missions, or 1 in 10,000,000,000 per hour, and they asked how long a test must run to expect even one failure at that rate: about 1.1 million years on one system, and 114 years even with 10,000 copies tested in parallel.
The bound holds only when the test runs resemble real use, so that the rare inputs of the field appear in the test as often as they do in practice. That condition is a second weakness of statistical testing for safety, after the number of runs it needs: by their nature, the inputs that cause harm are the ones nobody thought to make common in the test.
Choosing the inputs that matter
If testing cannot be complete, the question becomes which inputs to try. Three ideas guide most of that choice. Inputs can be grouped into classes the software should treat alike, with tests at the boundaries between classes, where off-by-one mistakes occur. Combinations of settings can be covered systematically rather than exhaustively. And the hazard analysis of Chapter 4 can name the specific sequences that would cause the most harm, such as the ventilator's four conditions, so that each becomes a deliberate test.
The second idea has unusually direct evidence from medical devices. NIST researchers Dolores Wallace and Richard Kuhn studied FDA recall reports from 1983 to 1997 and found 109 that described their failures in enough detail to show what triggered them. In all but 3 of those 109, testing every pair of parameter settings would have revealed the failure. That is 97 percent of them; the authors give 98 percent. The other 3 needed more than two conditions to occur together. The number of tests needed to cover all pairs of settings grows slowly as settings are added, so pairwise testing is affordable where testing all combinations is not.
The HAMILTON-C6 recall shows the limit of pairwise testing. Its failure needed four conditions, so a pairwise campaign would not have been sure to find it, and only an analysis of what happens when a sensor fails during suctioning would have pointed to that sequence. Testing, then, is necessary and never sufficient. The confidence a regulator accepts comes from testing chosen by requirements and risk, applied at several levels, repeated after every change, and backed by a process that makes faults less likely to be written in the first place. That process is the subject of the next part.
For each hazard the software could contribute to, list the inputs and sequences that could lead to it, including combinations of user actions, sensor states and timing. Make those deliberate tests; cover the remaining settings with boundary values and all-pairs combinations; and record which hazards rely on testing alone and which also rely on safeguards built into the design, so that a reviewer can see where the testing argument is weak.
- Paths through software multiply with every decision, so even modest programs have far more behaviors than any campaign can test.
- Testing can show that faults exist but cannot show their absence; coverage measures what ran, not whether it was right.
- By the rule of three, proving a very low failure rate by failure-free testing would take an impossible amount of realistic operation.
- Useful testing is a deliberate sample: input classes and boundaries, all pairs of settings, and sequences named by the hazard analysis.
+ Part II · Confidence from process
The software lifecycle.
IEC 62304, the international standard for the life cycle of medical device software, was published on 9 May 2006 and amended in June 2015. It prescribes activities and the records each one produces, but not an order in which to carry them out. Since January 2019 the FDA has recognized the amended edition in full, which means it accepts a maker's declaration of conformity to the standard, or to named clauses of it, in a premarket submission. A second edition would widen the standard's scope to health software in general. In August 2026 it reached its second committee draft, and it is not expected to be final before 2028. Because testing cannot be complete, the confidence a regulator accepts comes largely from how the software was made. The six chapters of this part follow that process: requirements derived from risk, an architecture that isolates what is dangerous, an account of the code taken from others, testing at several levels, and the ways agile teams and automated pipelines meet the same obligations.
What it must do, and must never do.
+ The questionWhat has to be written down before the code can be trusted, and why does risk decide how much?
A requirement is a promise that can be checked
A software requirement is a statement of what the software must do, or must never do, written so that a test, an inspection or an analysis can show whether it holds. "The display shall be easy to read" is not a requirement in this sense; "each dose value shall be shown with its unit, in characters at least 5 mm high, and shall remain on screen until the user confirms it" is. The difference is that the second statement can fail a test.
Requirements come from the intended use of Chapter 1. A claim to detect more than mild diabetic retinopathy from two photographs per eye implies requirements about which images are accepted, what happens when an image is too dark or blurred, what the three possible answers are and how each is shown, how results are tied to the right patient, and how long the analysis may take. Some requirements come from users and clinicians, some from standards, some from the hazards the software could cause, and some from the rest of the system the software is part of. Once written down and reviewed, the requirements become the reference for all later work: design is checked against them, tests are written from them, and a change is judged by which of them it touches.
Risk decides how much
Not every requirement carries the same weight. Risk management for medical devices follows ISO 14971. Its 2019 edition was reviewed in 2025 and confirmed without change. The standard describes a chain: a hazard, a potential source of harm; a sequence of events that leads to a hazardous situation, in which someone is exposed to the hazard; and the harm that may follow. The maker judges each chain by the severity of the harm and the probability that the harm occurs. Each unacceptable risk gets a risk control, a measure that reduces the probability or the severity of the harm. The immunoassay guide in this library follows the chain for an analyzer, from a fault to a wrong result to a harm. For software, one part of the judgment changes: the probability.
Chapter 2 showed that software failures are systematic, so there is no failure rate per hour to put into the chain. In a 2025 document on software-specific risk, the International Medical Device Regulators Forum suggests that it can help to start from a probability of software failure of one, as if the failure were certain, and, where possible, to estimate the probability of harm from other factors. In that approach, the risk is judged by its severity and by the controls that act if the failure happens. A risk control may itself be software, a check in another part of the program; or it may be hardware, procedure or labeling outside the software altogether.
Three classes of software
IEC 62304 builds on this risk analysis with three software safety classes. In the amended edition, software is class A if it cannot contribute to a hazardous situation, or if no hazardous situation it can contribute to leads to an unacceptable risk once risk controls outside the software are taken into account. It is class B if, after those external controls are taken into account, it could still contribute to a hazardous situation whose possible harm is a non-serious injury. It is class C if that harm could be death or serious injury. Until a team has classified its software, the standard treats it as class C.
The class sets how much process the standard demands. All three classes need a development plan, requirements, testing of the whole software against those requirements, risk management, configuration management, problem resolution and controlled release. In outline, class B adds a documented architecture and the verification of software units and of their integration; a software unit is the smallest part into which the software is divided. Class C adds a detailed design of each unit and more demanding criteria for verifying each unit. A class can also be set for each part of a program, and Chapter 5 shows how an architecture that separates dangerous functions can put most of the code in a lower class.
An infusion pump delivers fluid or a drug into a patient's body at a controlled rate, and its software computes that rate from a prescription. A wrong rate could kill, so before any controls the worst credible harm is death. Suppose the pump also has an independent hardware circuit, outside the software system, that stops the motor if the delivered volume exceeds a limit the clinician has set. Suppose too that the team's analysis shows that, with this circuit in place, the worst remaining harm from a software error is a non-serious injury. Under IEC 62304 as amended, the software may then be class B, because the external control counts. In its 2023 guidance on premarket software documentation, the FDA instead asks whether a failure "could present a hazardous situation with a probable risk of death or serious injury" before any risk controls are considered. The answer sets the documentation level, Basic or Enhanced, which decides how much software evidence a submission must contain. For the same pump software the answer is yes, so it would need the FDA's Enhanced documentation level. For the development practices in a submission, the guidance accepts a declaration of conformity to named clauses of IEC 62304, such as the development plan, maintenance and configuration management, and explains that a declaration to the whole standard is not needed, because the two schemes classify software differently.
A check that runs inside the same program, on the same processor and from the same code base, can fail together with the function it guards, as the redundant copies of Chapter 2 did. When a team sets the class under IEC 62304, the standard credits the risk controls that are external to the software system. A team that lowers a class by pointing to a software check inside the system has misread the standard and has also left the risk in place.
Tracing the threads
What makes requirements and risk controls auditable is a set of links between records. Each hazard links to the requirements that control it; each requirement links to the design elements that implement it and to the tests that verify it; each test links to its result. This linking is called traceability. With the links in place, a reviewer can start anywhere: from a hazard, find whether its control was implemented and tested; from a failed test, find which hazard is now uncontrolled; from a change, find every requirement and test it touches. Without the links, the same documents give a reviewer no path from a risk to its evidence.
The FDA made the top of such a chain public for IDx-DR in its decision summary: the three risks to health listed in Chapter 1, each linked to the mitigations that answer it. The FDA's reviewers also read the requirements and tests behind that summary, and those are not public. Chapter 7 describes the testing; Chapter 27 lists what the FDA, the EU and India ask to see of it.
Classify the software, and each software item, before the architecture is fixed. Write down the hazards, the external controls credited and the worst remaining harm behind each class. Record the FDA documentation level separately, assessed before risk controls. Revisit both whenever a requirement, a control or the intended use changes, because a class that rested on a hardware limit becomes wrong, without any warning, if that limit is removed.
- A software requirement states what the software must or must never do in a form that a test or inspection can check.
- ISO 14971 describes the chain from hazard to harm; because software failures are systematic, regulators suggest that it can help to assume a software failure will occur.
- Under IEC 62304, software is class A, B or C according to the worst harm after controls outside the software, and the class sets the process required.
- Traceability links hazards, requirements, design and tests, so a reviewer can follow any risk to its control and its evidence.
Keeping the dangerous part small.
+ The questionHow does the way software is divided into parts change what must be proven about each?
A timer that does not trust the program
A watchdog is a timer, often a separate circuit, that the software must reset at regular intervals. If the program hangs, stuck in a loop or waiting for something that never arrives, it stops resetting the timer, the timer runs out, and the watchdog forces the device into a safe state: it stops a motor, raises an alarm or restarts the processor. The FDA finalized its guidance on infusion pumps in 2014. In that guidance, the FDA asks makers to describe their watchdog timer, lists watchdog tests among the safety mechanisms a pump should have, and lists a watchdog failure among the hardware causes of system failure.
The watchdog shows the main idea of software architecture for safety: a function that must not fail is protected by something that does not share its weaknesses. The watchdog does not run the dose calculation and does not need to understand it. It needs only to be reset at regular intervals, which shows that the program is still running, and its own logic is small enough to verify thoroughly. A maker's safety case is its argument, backed by evidence, that the device is safe. The same guidance adds a caution that applies to every safety mechanism: the safety case must also cover the hazards that the mechanism itself could create. One such hazard would be a watchdog that resets a pump in the middle of an infusion.
A program that runs normally while computing a wrong dose keeps resetting its watchdog on time, so the watchdog never acts. Protection against a wrong output needs a different mechanism: an independent check of the result, a limit enforced outside the software, or a second calculation by different means. Naming which failures each safety mechanism detects, and which it cannot detect, is part of the design.
Items, units and boundaries
IEC 62304 describes a program as a software system made up of software items, which are any identifiable parts of it. Items are divided in turn into smaller items, down to software units (Chapter 4). The division is not a filing scheme. It determines whether a fault in one part can affect another. If one item can overwrite another's memory, use up the processor time the other needs, or hand it corrupted data without detection, a fault in the first item can make the second fail, and both must be treated as one.
The standard calls the remedy segregation: any mechanism that prevents one software item from negatively affecting another. Segregation can be physical, with the critical function on its own processor; it can use the memory protection of an operating system, so that one process cannot write into another's memory; or it can rest on checks at the boundary, so that data crossing the boundary is validated. Each mechanism has to be shown to work, because a claimed boundary that a fault can cross is worse than none: it lowers the rigor applied to the code outside the protected part without lowering the risk.
Segregation is what lets the class of Chapter 4 apply to parts of a program rather than to the whole. If the items that could contribute to serious harm are separated from the rest, only those items need the class C activities, and the user interface, the logging and the reporting code can be developed at class B or A. An architecture is therefore a decision about how much evidence each part of the program needs, as much as a decision about its structure.
Suppose a pump's software has 200,000 lines. The dose calculation and the delivery control are the only parts whose failure could cause death or serious injury, and together they take 8,000 of those lines. Without demonstrated segregation, the whole program is class C, and the detailed design and stricter unit-verification criteria of that class apply to all 200,000 lines. Now suppose the 8,000 lines run on a separate processor that the rest of the program cannot write to, and every command crossing the boundary is checked against the prescription. The class C work then applies to 8,000 lines, 4 percent of the program. The remaining 96 percent still needs its own class, its own tests and evidence that the boundary holds.
Where the parts live
The architecture also has to say where each part runs and what it depends on. Much of the code in a modern device was written by someone else: the operating system, the network stack, a database, a library that reads images, a cloud service. Each of these is an item the maker did not write and cannot fully inspect. For each one, the architecture has to say where it runs, limit what it can affect, and state what the device does when it misbehaves. Chapter 6 is about those items.
IDx-DR's public record lists three software items, each with its own version number: the client on the camera's computer, a service on the server that passes images and results between the client and the analysis, and the analysis that grades the images. In June 2021 the FDA cleared a 510(k) for an update that moved the client to version 3.2.0, the analysis to 2.1.1 and the service to 1.1.2. The next update, cleared in June 2022, moved them to 3.5.0, 2.3.0 and 1.2.0.
Because the parts are separate, a maker can change one and argue, with evidence, that the others are untouched. The split also changes what can go wrong. An analysis that runs on a server depends on the network, and the 2021 version added messages that tell an operator whether a submission was lost or is simply not ready yet.
Chapter 21 returns to the 2022 change. That change replaced one model inside the analysis, the classifier that judges image quality, and it changed which patients received an answer. A maker can confine a change to one item, but the architecture does not guarantee that the change's effects stay inside that item. Every change therefore needs regression tests at the level of the whole system.
Draw every software item, including those from third parties, with its safety class, where it runs, and the mechanism that separates it from its neighbors. For each boundary that lowers a class, record the evidence that the mechanism works, such as memory-protection settings or the separate processor, and include a test that tries to break it. For each safety mechanism, record which failures it detects, which it cannot, and what hazards it could create itself.
- A watchdog shows the core idea: protect a critical function with a mechanism that does not share its weaknesses.
- IEC 62304 divides software into items and units; segregation is any mechanism that stops one item from harming another.
- Demonstrated segregation lets the strictest class apply only to the code that could cause serious harm, not to the whole program.
- IDx-DR's three items carry their own version numbers, so each update records which items changed; a change made inside one item, such as the 2022 image-quality classifier, still needs regression tests of the whole system.
SOUP, OTS and the bill of materials.
+ The questionMost of the code in a modern device was written by someone else. How can a maker answer for it?
When the operating system expires
On 8 April 2014 Microsoft ended extended support for Windows XP, twelve years and three months after the system's lifecycle began. Extended support was the last phase of that lifecycle, and after it ended Microsoft stopped routine security fixes for the system. For medical devices built on Windows XP, those fixes stopped on the same date, and many of the devices stayed in service after it. An imaging scanner, a laboratory analyzer or a radiology workstation is expensive, is expected to last a decade or more, and was validated with one exact version of its operating system. Replacing that system is a design change that needs new testing and, for some devices, a new review.
The cost of waiting became visible on 12 May 2017, when the WannaCry ransomware spread across networks running unpatched or unsupported versions of Windows. In England, the hospitals and many other services of the National Health Service are run by public bodies called trusts. The UK's National Audit Office found that at least 81 of the 236 trusts, 34 percent, were affected. It identified 6,912 canceled appointments and reported an estimate of more than 19,000 in all. Bayer confirmed two reports from US customers that devices it made had been infected, and Siemens Healthineers warned that some of its products might be affected.
In 2020 a security vendor surveyed its customers' networks and reported that 83 percent of the medical imaging devices it saw ran operating systems that were no longer supported. That share had risen sharply since 2018, and the vendor attributed the rise to the end of support for Windows 7.
Suppose a device is launched in 2009 on Windows XP, sold until 2012, and used for ten years from the date of sale. Extended support ended in April 2014, five years after launch. A unit sold in 2012 stays in service until 2022, eight years after its operating system stopped receiving security fixes. Any vulnerability published in those eight years remains unfixed on that unit unless the maker can fix it some other way. In its 2026 cybersecurity guidance, the FDA addresses this gap between the end of support and the end of service: it asks makers to state, for each software component, its level of support and its end-of-support date.
SOUP and OTS
Two terms name the code a maker did not write, and they overlap without being the same. IEC 62304 calls such code SOUP, software of unknown provenance. In the standard's definition, a SOUP item is either a software item that is already developed, generally available and not developed to be part of the device, or a software item developed earlier for which adequate records of its development are not available. The FDA calls such code OTS, off-the-shelf software: a generally available software component used by a device maker that cannot claim complete control of the component's life cycle.
The emphasis differs. SOUP is about provenance: a maker's own old code, written before anyone kept records, is SOUP even though nobody bought it. OTS is about control: a commercial operating system is OTS because its maker, not the device maker, decides what goes into it and when. Most third-party components are both. A compiler or a test tool that never ships in the device is not SOUP, because it is not part of the device. Such a tool is validated for its use under the quality-system rules of Chapter 9.
Under IEC 62304, a maker does the same things for every SOUP item. It states what functional and performance requirements the device places on the item, what hardware and software the item needs, and its exact identity and version. It evaluates the anomalies, or known defects, that the component's own developer has published, to see whether any could cause a hazard. It also keeps the item under configuration control, so that the version tested is the version shipped. The FDA last revised its guidance on off-the-shelf software in August 2023. In that guidance it asks for much the same and adds a plan for maintaining support. For software at the Enhanced documentation level, the FDA also asks for assurance about how the component's developer works and what the device maker will do if that developer changes or stops its support.
How much of the code is borrowed
The borrowed share is usually most of the program. A vendor that audits software for company acquisitions reported in 2026 on 947 codebases it had examined. Of these, 98 percent contained open-source code, with a mean of 1,180 open-source components per application, and 87 percent contained at least one component with a known vulnerability. The sample is software being sold in a transaction, not medical devices, but device software is built from the same libraries. An operating system, a network stack, a database, an image library, a cryptography package and a user-interface toolkit can each depend on dozens of further components of their own.
Models and data are now among the borrowed components. A device can include a model pretrained by another company on images the device maker has never seen, or a public dataset used for tuning. Neither is code in the usual sense, but each is a component whose origin, version and known limits matter as much as a library's. In 2023 a machine-learning section was added to CycloneDX, a standard format for listing the components of a program, so that models and datasets can be listed with the code.
The bill of materials
A software bill of materials, or SBOM, is the machine-readable list of every component in a piece of software, with enough detail to identify each exactly. In July 2021 the US National Telecommunications and Information Administration set out seven minimum fields: the supplier, the component's name, its version, other unique identifiers, its dependency relationships, the author of the SBOM data and a timestamp. In mid-2026 CISA, the US cybersecurity agency, published updated minimum elements. They add each component's cryptographic hash, which identifies its exact contents, and its license. They also add the tool that generated the SBOM and the context in which it was generated, and they extend the scope explicitly to AI software and to software delivered as a service. Two formats dominate: SPDX, published as the international standard ISO/IEC 5962 in 2021, and CycloneDX, an Ecma standard since 2024.
A component with a published vulnerability may be present but unreachable: the device may never call the affected function, or never expose it to a network. A component with no published vulnerability may still be flawed. An SBOM lists which components are in the device. Whether a given vulnerability can be exploited in that device is a separate statement, which the industry calls a VEX; it gives the device's status for that vulnerability: not affected, affected, fixed, or under investigation.
In the US, since 29 March 2023 the law has required every premarket submission for a "cyber device" to include an SBOM that covers commercial, open-source and off-the-shelf components. A cyber device is one with software that can connect to the internet, a condition the FDA reads broadly. Chapter 24 follows the SBOM after release, where it serves its main purpose. When a new vulnerability is published in a component, a maker and a hospital can each use the SBOM to find which devices contain that component, far faster than by searching old build records.
For each SOUP or OTS item, record its exact version, its supplier, the requirements the device places on it, the published anomalies reviewed and the conclusion for each, its support level and end-of-support date, and a plan for replacing it. Pin versions so that what was tested is what ships, generate the SBOM from the build rather than by hand, and review the anomaly lists again before each release.
- Windows XP's end of support in 2014 left devices without security fixes for years; in 2017 WannaCry, spreading through unpatched or unsupported Windows, showed what that can cost.
- IEC 62304's SOUP is code of unknown provenance; the FDA's OTS is code whose life cycle the maker cannot fully control; most components are both.
- For each borrowed component a maker must state requirements, exact version and evaluated anomalies, and keep it under configuration control.
- An SBOM lists every component, now including models and datasets; whether a vulnerability is exploitable is a separate VEX statement.
Evidence at every level.
+ The questionWhich tests, at which level, give evidence a regulator accepts, and when is testing enough?
A sequence of steps nobody combined
In February 2023 Elekta began correcting its Monaco radiotherapy treatment planning system, the software a planner uses to design the radiation beams for a patient and to compute the dose they will deliver. Elekta had found that one sequence of user steps could make the system display an inaccurate dose. The problem arose when a planner added contours, outlines of structures drawn on the patient's images, and then re-optimized the plan, so that the system recalculated the beam settings to meet the plan's dose goals. The system computes dose from the density of the tissue, and a planner can force a fixed density value onto a region. If no density value was forced outside the patient's external outline, the system could display a dose distribution that did not match what the plan would deliver.
The correction covered 2,020 installed systems running four builds of version 5.11, and the FDA recorded the root cause as software design. Drawing a contour, assigning densities, optimizing a plan and displaying a dose were each an ordinary use of the software; the fault lay in their combination. Failures of this kind are why device software is tested at more than one level. A test of a single function can show that the function meets its own specification. Only tests that combine functions in the sequences users actually follow can show how the functions behave together, and only tests of the whole system in its intended environment can show whether the device does what its users need.
Levels of evidence
Verification is the FDA's term for objective evidence that the outputs of a stage of development meet the requirements set for that stage: in short, evidence that the software was built as specified. Verification happens at several levels. Unit testing checks each unit in isolation against its detailed design, often automatically. Integration testing checks that units and items work together, including at the segregation boundaries of Chapter 5. System testing checks the complete software against the software requirements, usually on the real hardware. Other verification activities find faults without running the code at all: reviews of requirements and design, inspection of the code by a second engineer, and static analysis tools that search the source for whole classes of mistakes, such as reading a memory location before any value was written to it.
Under IEC 62304, every class needs system testing, class B and above also need unit and integration verification, and each of these activities must leave records. In its 2023 guidance on premarket software documentation, the FDA asks for a summary of unit, integration and system testing at the Basic documentation level, with the protocol and report of the system-level tests. At the Enhanced level it also asks for the protocols and reports of the unit and integration tests. The standard and the FDA both ask for a list of the unresolved anomalies: the known defects the software will ship with, each with an assessment of its effect on safety and effectiveness. A maker can release a device with known defects, provided it has judged each one and none leaves an unacceptable risk.
How much testing is enough
Chapter 3 showed that the answer cannot be "all of it". Coverage tools measure part of the answer: they count how much of the code a test campaign executed. The FDA's 2002 guidance on software validation describes a series of coverage measures. Statement coverage runs every line at least once, and the FDA calls it insufficient to provide confidence. Decision coverage, also called branch coverage, takes each decision both ways. Condition coverage makes each condition inside a decision both true and false. Multiple-condition coverage tries every combination of the conditions in each decision. Full path coverage would run every possible route through the program, and the FDA calls it generally not achievable. The FDA does not prescribe a level; the maker chooses one in proportion to risk and justifies it.
An infusion pump stops when "(line pressure is high AND the occlusion alarm is enabled) OR the door is open". The occlusion alarm warns of a blocked line. Decision coverage needs two tests: one in which the pump stops and one in which it does not. Condition coverage can also be met with two tests: one with all three conditions true and one with all three false, a pair that never shows what happens when pressure is high but the alarm is disabled. Multiple-condition coverage needs every combination of the three conditions: two to the power three, or 8 tests. For one decision the difference between 2 tests and 8 is small, but the number needed for multiple-condition coverage doubles with every condition added to a decision: a decision with ten conditions needs 1,024 tests. Across a program with thousands of decisions, many with several conditions, that growth separates a campaign that can be run every night from one that cannot. That is why the strictest measures are kept for the code whose failure could cause serious harm.
Coverage measures show what ran, not what was checked. A team therefore pairs them with three measures of what was checked: every requirement traced to at least one test that would fail if the requirement were broken, every hazard-related sequence from the risk analysis tested on purpose, and combinations of settings covered by the all-pairs method of Chapter 3. The team has the evidence a reviewer looks for once three things hold: the campaign meets those targets, its tests pass, and every remaining anomaly has been judged. The team has not proved the software correct, and its documentation says so: it lists what was not tested and why.
Regression after every change
In the FDA's count of software recalls from 1992 to 1998, 79 percent came from defects introduced by changes after release, which makes regression testing the most important habit in maintenance. After each change, the team re-runs the tests that cover what the change touched and, for anything beyond a trivial change, the whole system-level suite, so that behavior verified before the change is verified again after it. A suite that runs automatically makes re-running cheap enough to do after every change, and Chapter 9 is about that automation.
IDx-DR, the guide's running case, went through a regression test of a model. In 2020 IDx renamed itself Digital Diagnostics. The renamed company replaced the classifier that judges image quality in version 2.3. To support the 510(k) for that version, which the FDA cleared in 2022, it re-ran the new version on the images from the 2017 trial and compared its answers with those of the original version. Chapter 21 shows what that comparison found, and what it could and could not prove.
Verification and validation
Verification asks whether the software meets its specification; validation asks whether the specification was right. The FDA defines software validation as confirmation that software specifications conform to user needs and intended uses, made by examination and objective evidence. For a software device, validation involves more than testing code: it includes usability studies with representative users and, where the claim requires it, clinical evidence that the outputs are right for patients. IEC 62304 deliberately leaves validation out. Its scope excludes the validation and final release of the device, which belong to the maker's design controls: the quality-system procedures that govern how a device's design is planned, reviewed, verified, validated and changed.
A system test checks the software against requirements the team wrote. If a requirement was wrong, every test can pass while the device fails its users. An image-quality threshold set for a camera the clinics do not use would be such a requirement. Validation needs evidence from outside the specification: real users, real conditions and, for a diagnostic claim, a comparison with an independent reference.
The evidence that IDx submitted for IDx-DR's 2018 authorization included both kinds, and the company funded the trial. The trial of the opening scene validated the claim, with novice operators who had four hours of training, real patients and a reference standard (Chapter 12). A smaller study checked repeatability, whether the system gives the same answer when the same person is imaged again. Each of 24 participants was imaged ten times, by three operators on two cameras of the same model. The system analyzed 235 of the image sets, gave identical answers for 23 of the 24 participants across every repeat, and agreed with itself 99.6 percent of the time. Repeatability is not accuracy, but a screening program whose answer changed with the operator would have failed before the question of accuracy arose.
For each software item, state which levels of testing apply, which coverage measure is targeted and why it fits the item's class, how requirements and hazard sequences map to tests, and which test suites are re-run after which kinds of change. Keep the unresolved-anomaly list with each release and record why each entry leaves no unacceptable risk.
- Device software is verified at the unit, integration and system levels, with reviews and static analysis finding faults without running the code.
- Coverage measures how much of the code ran; a maker chooses the level in proportion to the harm a failure could cause and pairs it with requirement and hazard coverage.
- A release may carry known defects only if each is listed and judged; regression testing after every change guards against the source of 79 percent of the software recalls the FDA counted in 1992 to 1998.
- Verification shows the software meets its specification; validation, including usability and clinical evidence, shows the specification meets the needs of users and patients.
Sprints with a paper trail.
+ The questionDoes IEC 62304 force a team into phases, or can it work in two-week sprints?
A standard without a sequence
IEC 62304 lists activities: planning, requirements analysis, architectural design, detailed design, implementation and unit verification, integration, system testing and release, with maintenance, risk management, configuration management and problem resolution running alongside. Read in that order, the list looks like a waterfall, a process in which each phase is finished and signed off before the next begins. For years many device teams worked that way and assumed the standard required it. The standard requires that the activities be done and recorded, and that the maker's development plan say how; it does not say in what order, or how many times.
Most software outside regulated industries is now built iteratively, by methods usually grouped under the name agile. A team works in short cycles of one to four weeks, each ending with working, tested software; requirements are refined as users see results, and the design grows with the product. The question for a device maker is whether that way of working can produce the evidence of Chapters 4 to 7: requirements traced to hazards and tests, a documented architecture, verification at each level, and a controlled release.
AAMI, the US standards association for medical technology, answers that question in a technical report, TIR45, which describes how agile practices map onto IEC 62304. AAMI first published TIR45 in 2012 and revised it in 2023. The FDA has recognized both editions in full, the second in May 2025.
Four layers of time
TIR45 describes the work in four nested layers. The project or product layer runs for months to a year or more and holds the high-level requirements and the overall architecture. A release runs for one to several months and ends with software that could be delivered. An increment runs for one to four weeks and ends with integrated, tested software; many teams call it a sprint. A story is one small piece of function that a user can see, and takes one to a few days.
TIR45 places each IEC 62304 activity in one or more layers. Planning happens at every layer, at its own scale. High-level requirements and the coarse architecture belong to the project layer. Detailed requirements, detailed design, and implementation with its unit tests belong to each story. Integration and system testing happen at the story, increment and release levels, and the formal release activity sits at the project layer: it is carried out each time the maker ships a release. A user story with its acceptance criteria becomes a design input, one of the documented requirements the design must meet. The end of an increment or a release becomes a point for design review, a documented examination of whether the design meets those requirements. The development plan states which of those reviews are formal and recorded.
Suppose a team plans a 12-month project in two-week increments, with a release to users every three months. That is 26 increments and 4 releases. If the team completes about ten stories per increment, it finishes about 260 stories in the year. Each story then leaves updated requirements, design notes, unit tests, trace links and, where it touches a hazard, an updated risk analysis, provided the team's definition of done, described in the next section, requires those records. The formal system test, the anomaly review and the release approval happen at the end of each of the four releases. Instead of one large documentation effort at the end, the team produces the evidence in about 260 small pieces and reviews each while the work is fresh, so each release is assembled from records that already exist.
Software can move from a developer's machine to a test environment many times a day, and an increment can be demonstrated every two weeks. A release to patients is different: IEC 62304 requires that verification be complete, that known residual anomalies be documented and evaluated, that the released version and how it was built be recorded, and that the software be archived and reliably delivered. Agile methods speed up the work that leads to a release, but they do not reduce what the release itself requires.
Done means verified and recorded
The practice that makes agile work under a standard is a strict definition of done: the list of conditions a story must meet before it counts as finished. A story is not finished when its code runs. It is finished when its requirement is written and reviewed, its design and code are reviewed, its unit and integration tests pass, its trace links are in place, the risk analysis reflects any new hazard or control, and any change to SOUP is recorded. If any of these is missing, the story goes back into the work, exactly as it would if a test failed.
Two further habits make sure the records describe the software as it is. The first concerns requirements: the backlog, the team's list of stories waiting to be built, is not the requirements specification, but it supplies requirements to the specification. As stories are accepted, their requirements enter a controlled baseline, a fixed version of the specification that is approved at each release, so that the release is checked against a stable specification rather than a moving list. The second habit is live traceability: the team links stories, requirements, code changes and tests in the tools it already uses, as the work happens. The trace matrix, a table of the links that Chapter 4 describes, is then a report generated from them rather than a document written afterwards.
What changed in 2023
The 2023 edition of TIR45 kept the core of the first edition and updated its context. According to AAMI, the new edition adapts documentation to digital practice, with streamlined approval processes and signature requirements, and it brings cybersecurity, risk management, design validation and usability engineering into agile work rather than treating them as separate phases. Its references now point to ISO 13485:2016, the amended IEC 62304 and ISO 14971:2019. The FDA accepts declarations of conformity to the 2012 edition only until 2 July 2028.
The same shift appears elsewhere. ISPE, a professional society for pharmaceutical engineering, publishes GAMP 5, the guide that regulated companies use to validate the computerized systems they buy and build. In its second edition, published in 2022, GAMP 5 states that its life cycle is not inherently linear and that it supports iterative and incremental methods; Chapter 19 returns to GAMP 5. In July 2026 VDE, the German electrical standards association, reported that the draft second edition of IEC 62304 itself places agile development in an informative annex, which gives guidance but adds no requirements.
Write the layers, their lengths and the activities placed in each into the software development plan, and state which reviews are formal. Make the definition of done include requirements, design, tests, trace links and risk updates for every story, and approve a requirements baseline at each release. An auditor who opens the plan should find the team's real way of working described there, not a waterfall it does not follow.
- IEC 62304 requires activities and records, not an order of phases, so iterative development can meet it if the plan says how.
- AAMI TIR45, recognized by the FDA, maps the standard's activities onto project, release, increment and story layers.
- A strict definition of done, a requirements baseline at each release and live traceability let evidence accumulate as the work is done.
- Increments can be frequent, but a release to users stays a controlled event with verification, anomaly review, archiving and approval.
Every build a record.
+ The questionWhen code is built and tested automatically on every change, what turns that pipeline into evidence, and where must a person still decide?
Deploying many times a day
DORA, a research program at Google Cloud, runs a survey of software delivery. In its 2024 survey, 19 percent of respondents, the best-performing fifth, formed what DORA calls the elite group. These respondents deployed changes on demand, often several times a day, and their changes took less than a day to go from code to production, the live system that users run. Of their deployments, 5 percent failed, and they recovered from a failed deployment in less than an hour. The survey covers technology workers in general, not medical device makers, and the difference in pace is deliberate: a device maker cannot put a new version in front of patients several times a day, and should not.
What the best general software teams do have is a pipeline: an automated sequence that takes every change a developer submits, builds the software from source, runs the tests, checks the code with analysis tools, and packages the result, with no step done by hand. Device makers increasingly use the same kind of pipeline, not to release faster to patients but to make every internal build a complete, repeatable record. With a pipeline, the regression testing of Chapter 7 runs on every change instead of in occasional campaigns.
The change failure rate is the share of deployments that fail, and for the elite group it was 5 percent. At three deployments a day, that rate produces 0.15 failed deployments a day, about one a week counting every day. In a test environment such a failure costs little: the team finds it within hours and rolls the change back to the previous version. For released device software the same arithmetic would mean a field failure each week. A device maker therefore runs the frequent cycle internally, where failures are expected and contained, and keeps the release to patients as a separate, verified and approved event, as Chapter 8 described.
What makes a pipeline evidence
A pipeline produces evidence when its outputs can be trusted and traced. Configuration management means identifying and controlling every item that goes into a build: the source code, the SOUP versions, the build scripts, and the compiler and its settings. Its purpose is that the same inputs produce the same software, and IEC 62304 requires it at every class. A reproducible build goes one step further: rebuilding a past release from its recorded inputs gives a bit-for-bit identical result, so the maker can show that the version tested and the version shipped are the same.
Each stage then leaves a record tied to the exact build it checked. Unit and integration test results are stored with the commit behind the build and linked to the requirements they verify; a commit is a recorded version of the source code, and the build is made from it. The pipeline also generates the static analysis reports, the coverage reports and the software bill of materials of Chapter 6, so that no one assembles them by hand and they describe the build that exists rather than the one someone remembers. The full system tests may take hours on real hardware, and they run nightly or before each candidate release. When someone proposes a release, the team does not collect the evidence for it; it retrieves records that already exist.
Two decisions stay with people. Someone must judge each unresolved anomaly and decide that none leaves an unacceptable risk, and someone authorized must approve the release itself. The pipeline can check that every required record exists and every test passed. Whether the remaining defects are acceptable for patients is a judgment, and IEC 62304 and the regulators expect a named person to make it.
Trusting the tools
The pipeline's tools are software too: the build system, the test runner, the coverage tool, the static analyzer, the SBOM generator, and the issue tracker that holds the trace links. None of them ships in the device, so none is SOUP. Each is, however, quality-system software, because the maker uses it in its quality system: the set of controlled processes by which a maker designs, makes and monitors its devices. The maker must validate each tool for its intended use, whether the tool was bought, is open-source or was written in-house. A test runner that reports a pass for a test that never ran, or a coverage tool that miscounts, would produce false evidence about a device that might be dangerous.
ISO 13485:2016 is the quality-management standard for device makers, and the FDA's rules have incorporated it since February 2026. It requires a maker to validate software used in its quality system and in manufacturing, before first use and after changes, in proportion to the risk of its use. In 2017 ISO published guidance on doing so, its technical report 80002-2. The FDA sets out a risk-based approach in its guidance on computer software assurance. The agency finalized that guidance on 24 September 2025 and reissued it on 3 February 2026 under the title Computer Software Assurance for Production and Quality Management System Software.
Under the guidance, a function is high process risk if its failure "may result in a quality problem that foreseeably compromises safety". For such functions the guidance suggests more rigorous assurance, such as documented, scripted testing, which follows written steps with stated expected results. For other functions it suggests lighter, unscripted methods such as exploratory testing, in which the tester designs and runs tests while learning how the software behaves. The guidance replaces one section of the FDA's 2002 software validation guidance. It explicitly does not cover the device software itself, which remains under design controls.
Some pipeline failures produce records that look complete: a test step that silently skips tests after a configuration change, a runner that treats a crash as a pass, and an SBOM generator that misses dependencies loaded at run time. The most direct check is to insert known faults on purpose, a practice called seeding, and confirm that the pipeline catches them. That check should be repeated whenever a tool or its configuration changes.
After release: fixes and maintenance
Release starts the longest part of a device's software life. IEC 62304 requires a maintenance plan and a problem-resolution process. Under that process, field reports and internal findings become problem reports, and each report is investigated and classified for its effect on safety. Changes go through the same controlled process as development, and trends across problems are analyzed. The lessons of Chapter 2 apply directly here. In the FDA's count from 1992 to 1998, 79 percent of software recalls came from defects introduced by changes after release, and the Tandem app recall of 2024 followed a change to software that already worked.
Hosted software changes how fixes reach users. When software runs on a maker's server, as IDx-DR's analysis does, a fix can reach every user at once and can be rolled back at once, which is an advantage for safety. The same property is a risk for control: any change, not only a fix, can reach every user at once. A function can also be turned on for some users by a switch, often called a feature flag. Turning a function on this way is a change to the released device, however small the switch, and it goes through change control. Whether that change needs a regulator's review depends on its effect on claims and risk, and Chapter 28 sets out the rules for the US, the EU and India.
Much code in the field also predates the processes now expected of it. The 2015 amendment to IEC 62304 added a route for legacy software: software that is legally marketed but lacks enough evidence that it was built to the standard. The route has four parts: a risk analysis using field experience, a gap analysis against the standard's requirements, a plan to close the gaps that matter, and a documented rationale for continued use.
List the tools in the pipeline and, for each, the records it produces and what a silent failure would do to the device's evidence. Give scripted testing with seeded faults to the tools that produce verification results, coverage, builds and SBOMs; pin their versions and archive the build environment with each release; and use lighter, unscripted assurance for tools whose failure could not compromise a release.
- A pipeline builds and tests every change automatically; device makers use it to make each build a complete record, not to release faster to patients, and a person still judges open anomalies and approves each release.
- Configuration management, reproducible builds and records tied to each build let a release's evidence be retrieved rather than assembled.
- Pipeline tools are quality-system software that must be validated in proportion to risk, the approach of the FDA's computer software assurance guidance.
- After release, maintenance and problem resolution follow the same controls, and legacy software needs a gap analysis and a rationale for continued use.
+ Part III · Learning from data
How an AI model is built and scored.
In 2016 a team at Google trained a neural network to grade photographs of the retina for diabetic retinopathy without writing a single rule about what the disease looks like: no instruction to look for small hemorrhages, no threshold for the size of a lesion. Every rule the network applied came instead from 128,175 images that ophthalmologists had already graded. A program built that way can still be tested, but only by comparing its answers with right answers on data it has not seen. Every step of that comparison involves choices that decide what the result means. The four chapters of this part follow those choices: what a model is, how the data that builds it is kept apart from the data that judges it, what the right answers are and who decided them, and how a score becomes a decision with a measurable error.
Rules nobody wrote.
+ The questionWhat is a machine-learning model, and how is building one different from writing code?
Twenty-five million adjustable numbers
ResNet-50 is one of the most widely used neural networks for analyzing images. The version published with the PyTorch software library has about 25.6 million adjustable numbers. Each is a parameter: most are weights that multiply values inside the network before they are passed on, and the rest are offsets added along the way. A photograph goes in as a grid of pixel intensities and passes through some fifty layers, each of which combines the values from the layer before using its weights, so that a score comes out at the end: for example, the network's estimate that the image shows a particular disease. If the weights change, the same photograph produces a different score.
A machine-learning model is that whole arrangement: a fixed structure, chosen by engineers, with its parameters set from data rather than by hand. The structure is code like any other, written, reviewed and tested. The parameters are not written at all. They are found by training. The model scores a batch of examples whose right answers are known, the training code computes a measure of how wrong the scores were, called the loss, and the code then nudges every parameter slightly in the direction that would have made the loss smaller. Repeated over millions of batches, the nudges settle into values that make the model's scores match the known answers well.
Google's system of 2016 was trained this way, on 128,175 retinal images that a panel of 54 US ophthalmologists and senior residents had graded three to seven times each. The system was an ensemble of ten networks of a design called Inception-v3; an ensemble is a set of models whose scores are combined into one. Its authors describe each of the ten networks as having about 22 million parameters.
Fitting, and fitting too well
Training finds parameters that fit the examples. Whether they fit anything else is a separate question, and the Google system shows why: each of its networks had more parameters than there were images to fit. A model with that much freedom can, in principle, memorize its training set, giving the right answer for every image it has seen and an arbitrary one for any image it has not. Engineers limit memorization in four ways: by the choice of structure, by starting from parameters already trained on more than a million ordinary photographs, by penalties that keep weights small, and by stopping training before the fit becomes too close. None of these guarantees that the model will work on new images; only a test on new images can show whether it does.
Take ten points along a straight line, y equals 2x for x from 0 to 9, each moved up or down by a small, fixed amount of noise between about −1.4 and +1.3. A straight line has two parameters. Fitted to the ten points, it misses them by 0.93 on average. The average here is the root mean square: each miss is squared, the squares are averaged, and the square root of that average is taken, so that large misses count for more. A polynomial with ten parameters, a curve built from x and its powers up to x to the ninth, can pass through every point exactly, so its error on the training points is zero. Now test both on nine new points taken halfway between the old ones, from the same line with fresh noise. The straight line misses them by 0.78 on average. The polynomial misses them by 1.67, more than twice as much, because between the training points it bends to follow the noise. The model with the better training score is the worse model.
The omics guide in this library describes the same problem in another form. A molecular signature is a pattern across many molecular measurements that is used to classify patients. A signature discovered from many more measurements than patients fits its discovery data almost perfectly and fails on the next cohort; only a locked signature tested on independent samples shows whether it is real. An image network differs in scale, not in principle. Its score on the training data describes the past; its value lies in its predictions about data it has not seen.
The data is part of the design
Because the parameters come from the examples, any regularity the examples contain is a candidate for the model to learn. Some regularities are signs of the disease. Others are accidents of how the data was collected: which hospital an image came from, which camera took it, what text was stamped in its corner. A model has no way to tell a sign of disease from a sign of the clinic where diseased patients were photographed, unless the data contains cases that separate the two. Chapter 11 shows a network that learned, in part, to recognize the hospital.
The fact that shapes this guide, that software does not wear out, extends to AI. A conventional program's faults are built into its code; a model's faults can also be built into its data, through what the data contains, what it lacks and how its labels were assigned. A label is the right answer recorded for an example. It is unrelated to a device's labeling, the label and accompanying written material of Chapter 1. The device is the code, the parameters and the data that set them, together with the processing that turns a raw image into the grid of numbers the model expects. Changing any of them changes the device.
Reviewers can read the code of a model's structure, but they cannot read its parameters as rules, and no list of the features a network relies on is complete. Evidence about what a model does therefore comes from its behavior on carefully chosen data. The quality of that data, not how well the code is written, sets the limit on what can be claimed.
Locked and learning models
A model whose parameters are fixed after training is locked: the same input always gives the same output, and the model can be verified like any other software. The paper that reports the 2017 trial states that IDx-DR's analysis was locked before the trial began. A model that keeps adjusting its parameters from new data in the field is continuously learning: its behavior on Tuesday may differ from Monday's, and the evidence gathered before release describes a model that no longer exists.
In 2019 the FDA wrote that the AI devices it had cleared or approved had typically "only included algorithms that are 'locked' prior to marketing." Updates happen, but as new versions: retrained, tested and released through the change controls of Chapter 21. One of the earliest authorized plans for changing an AI device belongs to a program that guides cardiac ultrasound. When the maker modified that plan through a 510(k) in 2020, the public summary stated that all algorithm modifications would be "trained, tuned, and locked prior to release", and the plan excluded continuously learning algorithms. A draft EU guideline on AI in the manufacture of medicines goes further. It was published for consultation in July 2025 and was still a draft as of October 2026. It states that dynamic models, which keep adapting to new data in use, should not be used in critical applications at all. How a model may change after authorization is the second half of this guide's central question, and Part V answers it.
Put the model's structure, its trained parameters, the preprocessing code and an identifier of the exact training data under configuration management as one released item, with a version that changes whenever any of them does. Record the data's origin, size and labeling method alongside, so that a later reviewer can tell which data produced which version of the model.
- A machine-learning model is a fixed structure whose millions of parameters are set by training on examples with known answers.
- A model can fit its training data closely and still predict poorly; only a test on data it has not seen shows whether it generalizes.
- The training data is part of the design, so a model can learn accidents of data collection as readily as signs of disease.
- Locked models give the same output for the same input; in 2019 the FDA wrote that the AI devices it had authorized were typically locked, and such devices change only through new, tested versions.
Keeping the judge away from the builder.
+ The questionWhy must the data that judges a model be kept apart from the data that built it, and how does it leak back in?
A model that recognized the hospital
In 2018 a group at the Icahn School of Medicine at Mount Sinai in New York trained a network to detect pneumonia on chest radiographs and then examined what it had learned. The group used 158,323 images from three hospital systems. It trained on images from two of them and held out the third for external testing. An external test uses data from sites, devices or periods that played no part in building the model. Pneumonia was far more common in one source than in another: 34.2 percent of the images from Mount Sinai Hospital showed it, against 1.2 percent of those from the NIH Clinical Center.
The area under the curve measures how well a model's scores separate cases with the disease from cases without it: 1 means perfect separation, and 0.5 means no better than chance. The curve is a plot of the trade between missed cases and false alarms, which Chapter 13 draws. A model trained on both sources together scored an area under the curve of 0.931 on test images from the same two sources, an excellent result. Scored on each hospital's test images separately, the model reached only 0.805 for Mount Sinai and 0.733 for the NIH.
The difference between the pooled score and the separate scores had a simple cause. A second network identified which hospital system a radiograph came from for 99.95 percent of the NIH test images and 99.98 percent of Mount Sinai's. Cues that helped it included the laterality markers placed on the film to show the patient's left or right side, among them a metal token. Within Mount Sinai, a network could identify even the department, from the portable scanners used on wards and the inverted color scheme of emergency-department images.
The authors also ranked the images by nothing more than the pneumonia rate of their hospital, ignoring every pixel of lung, and that ranking gave an area under the curve of 0.861 on the pooled test set, most of the way to the model's 0.931 on the same images. Part of what looked like pneumonia detection was hospital detection. The authors reported that such cues "only became apparent to us after manual image review."
Three kinds of data
A model is built and judged with three separate sets of data. The training set is what the parameters are fitted to. The tuning set is used to make choices the training itself does not make: how long to train, how large a model to use, which of several versions to keep. Machine-learning engineers usually call it the validation set. The test set is held back until those choices are finished and is used once, to estimate how the finished model will perform. In regulation the word validation means something different: Chapter 7 used it for proof that a device meets users' needs. A submission that uses the word in both senses has to say which it means.
The separation matters because every set that influences the model stops being a fair judge of it. A model scored on its training data is scored on examples it was fitted to; a model scored on the data used to pick it is scored on the data that selected it for doing well there. Only a score on data that played no part in building the model estimates how the model will do on the next patient. In January 2025 the International Medical Device Regulators Forum finalized international principles of good machine-learning practice. Its fourth principle reads: training datasets are independent of test sets.
How the judge leaks into the builder
Independence between the sets fails in ordinary ways. The same patient can appear on both sides of a split: two eyes, two visits, several slices of one scan, each a separate image but all carrying the same anatomy and often the same disease. A model that has seen a patient's left eye in training has partly seen the right eye in testing. The Mount Sinai group split its data by patient: 70 percent of patients for training, 10 percent for tuning and 20 percent for testing. In 2018 the FDA wrote the same concern into the special controls for retinal diagnostic software, which require that analysis of performance make no unjustified assumption that repeated samples from one patient are independent.
Other leaks are quieter. Images can be normalized using statistics computed from the whole dataset, test images included; normalizing rescales pixel values to a common average and spread. Duplicates and near-duplicates can sit in both sets. A source with a high disease rate can make up a large share of both the training set and the test set, so that a model that learns to recognize the source scores well in training and again in testing, as the pneumonia network did with the hospital cues. And a team can look at the test results, adjust the model, and look again, until the test set has become a second tuning set. Each of these makes the reported score higher than the score on the next patient will be.
A test set used once gives an unbiased estimate for the model that was finished before the look. Every later decision informed by that score makes the next score on the same set more flattering: a changed threshold, a retrained model, a different preprocessing step. A sealed test set kept for the final estimate, and a fresh one for each major version, prevents that inflation.
How large a test set
A test set must also be large enough to give a precise score. Because a sensitivity is a share of the diseased patients, its uncertainty shrinks as the number of diseased patients in the test set grows, not as the total grows. A test set with plenty of healthy patients and few sick ones gives a precise specificity and an imprecise sensitivity.
A team expects a sensitivity of about 87 percent and wants its 95 percent confidence interval to extend no more than 5 percentage points either side. A 95 percent confidence interval is a range computed from the data in a way that, over many repeated studies, would include the true value 95 times in 100. By the usual approximation for a proportion, the number of diseased patients needed is 1.96 squared, times 0.87, times 0.13, divided by 0.05 squared, which is about 174; the 1.96 is the factor that corresponds to 95 percent confidence. If the disease is present in 24 percent of the people tested, about 725 participants are needed to include 174 with the disease. The IDx-DR trial was planned for at least 149 participants with the disease and 682 without, to give at least 85 percent power for its prespecified statistical tests, and it analyzed 198 participants with the disease and 621 without. Power is the chance that a trial passes its tests if the device truly performs as well as its designers expected. The trial could plan for fewer diseased participants than 174 because it asked a different question: whether sensitivity cleared a floor of 75 percent, rather than how closely it could be estimated. Chapter 13 describes that test.
The calculation sets a minimum size, not a target. A test set drawn from one hospital, one camera or one population can be large and still unrepresentative, which is the subject of Part IV. Size makes the estimate precise; only the right composition makes the estimate relevant.
Split the data by patient and, where possible, by site, before any training begins. Store the test set where the development team cannot use it for training or tuning, record who accessed it and when, and use it once for the final estimate of each version. Size it from the number of positive cases the claimed sensitivity needs, and record its composition so that a reviewer can compare it with the intended population.
- A pneumonia network scored far better on pooled data than on each hospital's images, because part of what it had learned was to recognize the hospital.
- Training, tuning and test sets must be separate; only data that played no part in building a model estimates its performance.
- Splitting by image instead of patient, shared preprocessing, duplicates and repeated looks at the test set all inflate the reported score.
- Precision depends on the number of diseased patients in the test set; about 174 are needed to estimate an 87 percent sensitivity within 5 percentage points either side at 95 percent confidence.
What the right answer is, and who decided.
+ The questionWhen a model is scored, what is it scored against?
The answer key
For every participant in the IDx-DR trial, readers at the Fundus Photograph Reading Center of the University of Wisconsin decided the right answer, from images the program never saw. Certified photographers took four overlapping stereo pairs of each retina, which covered a far wider field than the program's two pictures per eye. They also scanned the macula by optical coherence tomography, which uses light to image tissue in cross-section; each scan had at least 121 cross-sections.
Three experienced readers graded the photographs on the severity scale of the Early Treatment Diabetic Retinopathy Study, a numbered scale on which higher levels mean more severe disease. They graded without seeing the program's output or the other imaging, and the majority grade stood. A participant counted as having more than mild diabetic retinopathy if the worse eye reached level 35 or higher on that scale or showed macular edema, a swelling of the macula caused by fluid leaking from damaged blood vessels.
That protocol is the trial's reference standard, often called ground truth: the best available judgment of the true state, against which the device is scored. The trial's headline numbers are statements of agreement with the photographic grade: the 87.4 percent sensitivity, the 89.5 percent specificity and the predictive values of Chapter 13. A predictive value is the share of the program's positive, or negative, answers that were correct. A secondary analysis added the macular scans to the reference and gave a sensitivity of 85.9 percent and a specificity of 90.7 percent, figures adjusted for the trial's enrichment. Enrichment means recruiting extra participants who are more likely to have the disease, so that a trial includes enough cases. If the reference were wrong for some participants, the numbers would be wrong in ways the trial could not detect.
Experts disagree
Grading retinal photographs is skilled work, and skilled graders differ. In a 2018 study, researchers at Google and retina specialists from two clinical practices measured how much. Three fellowship-trained retina specialists graded 1,813 photographs independently and then resolved every disagreement in adjudication sessions, which produced a consensus grade for each image. Measured against that consensus, the three specialists working alone detected moderate or worse retinopathy in 74.6, 74.4 and 82.1 percent of the images that had it, while correctly clearing about 99 percent of those that did not. Three board-certified general ophthalmologists did similarly. In other words, five of the six graders working alone missed about one case in four that the adjudicated panel found, and the best of them missed about one in six.
The same study showed how much the reference matters to a model. The authors measured each grader's agreement with the consensus on a weighted scale on which 1 means perfect agreement and 0 means agreement no better than chance, and for the specialists it ranged from 0.80 to 0.91. They also built an improved version of the Google system of Chapter 10, with more training images, a higher image resolution and a newer network design, and tuned it on a small set of adjudicated grades. Against the adjudicated reference, the original network scored an area under the curve of 0.942 and the improved one 0.986. The authors report that the small set of adjudicated grades allowed substantial improvements, although the higher resolution alone raised the score to 0.960. How good a model looks depends on which human answer it is compared with, and a model trained or tested against a single grader inherits that grader's misses.
If a model is trained on one grader's labels and tested against the same grader, a high score shows that the model has learned to agree with that person, including the cases that person gets wrong. Independent, adjudicated or multimodal reference standards cost more, but they are the only way to measure accuracy rather than imitation. A multimodal reference combines more than one kind of test or imaging.
The reference caps what can be measured
An imperfect reference makes a perfect device look imperfect. When the device is right and the reference is wrong, the disagreement is counted as the device's error; when both are wrong in the same direction, the error is invisible.
Prevalence is the share of the people tested who have the disease. Take 1,000 people, of whom 240 truly have the disease: a prevalence of 24 percent, as in the IDx-DR trial. Suppose the reference standard misses 5 percent of true disease, labeling 12 of the 240 as healthy, and wrongly labels 1 percent of healthy people as diseased, about 8 of the 760. Now score a model that is always right. The reference calls 236 people diseased, 228 truly diseased and 8 not; the model flags the 228, so its apparent sensitivity is 228 out of 236, 96.6 percent. The reference calls 764 people healthy, including the 12 it missed; the model clears 752 of them, so its apparent specificity is 98.4 percent. The reference's own errors lower a flawless model's apparent sensitivity by more than three percentage points.
The direction of the bias depends on how the reference and the device err. If both struggle with the same hard cases, such as small lesions or poor images, they may agree on the wrong answer and inflate the score. Shared errors of this kind are one reason the IDx-DR trial used a reference built from more and different information than the device received: wider photographs and stereo views, with a scan of the macula for a secondary analysis, read by people who never saw the device's answer. Disagreements between device and reference then reflect the device's limits rather than errors the two share.
A reference fit for the claim
Regulators have changed their wording on this point. In 2021 the FDA, Health Canada and the UK's regulator, the MHRA, published principles of good machine-learning practice, which stated that selected reference datasets should be "based upon best available methods". The International Medical Device Regulators Forum finalized an international version in January 2025, and that version states instead that selected reference standards are "fit-for-purpose". One reading of the change is that the best method may be impossible or unethical to apply to every patient, and that the right reference depends on the claim.
A stroke-triage program authorized in 2018 was scored against neuroradiologists' readings, with an additional neuroradiologist settling disagreements. Researchers at Stanford assembled a dermatology dataset for testing skin-lesion models and took every label from a biopsy result, so a pathologist had confirmed each one. The IDx-DR trial used a reading-center protocol because the device's claim was a screening decision about retinopathy severity, which is exactly what that protocol grades.
Whatever the choice, the reference must be defined before the test, applied the same way to every case, independent of the device's output, and described in the labeling, so that a reader of the results knows what the device was compared with. Chapter 17 shows how to find that description in a public summary and what it implies about the limits of the numbers.
Write down the reference standard in the study protocol: who grades, with what information, using which scale, how disagreements are resolved, and how graders are kept from seeing the device's output. Measure and report the agreement among graders, plan adjudication for disagreements, and prefer a reference that draws on more or different information than the device uses.
- IDx-DR's sensitivity and specificity were scored against a reading-center grade of wider stereo photographs, read by three graders who did not see the program's output; macular scans were added to the reference only in a secondary analysis.
- Single experts disagree with an adjudicated consensus; most retina specialists working alone missed about a quarter of moderate or worse cases, and the best about one in six.
- An imperfect reference lowers a perfect model's apparent accuracy and can hide errors that model and reference share.
- A reference standard must be fit for the claim, fixed before testing, independent of the device and described in the labeling.
A threshold turns a score into a decision.
+ The questionA model outputs a score; who chooses where yes begins, and what does moving that point trade?
Four counts
Of the 819 participants whose results could be analyzed in the IDx-DR trial, 198 had more than mild retinopathy according to the reference standard and 621 did not. The program flagged 173 of the 198 and missed 25; it cleared 556 of the 621 and flagged 65 who did not have the disease. Every performance figure for the trial comes from those four counts. Sensitivity, the share of diseased participants flagged, is 173 out of 198, 87.4 percent, and specificity, the share of disease-free participants cleared, is 556 out of 621, 89.5 percent. The positive predictive value, the share of flagged participants who had the disease, is 173 out of 238, 72.7 percent; the negative predictive value, the share of cleared participants who did not have it, is 556 out of 581, 95.7 percent.
Each figure is an estimate from a sample and carries uncertainty. The exact 95 percent confidence interval for the observed sensitivity runs from about 81.9 to 91.7 percent; for specificity, it runs from about 86.9 to 91.8 percent. An exact interval is computed directly from the counts rather than by an approximation. The paper's own estimates are slightly different, 87.2 and 90.7 percent, because the trial deliberately recruited extra participants with poorly controlled diabetes, and the authors corrected for that enrichment with a statistical model.
The trial's statistical test asked whether the lower end of each interval cleared a floor set in advance: 75 percent for sensitivity and 77.5 percent for specificity. Two higher values, 85 and 82.5 percent, were the alternatives used to size the trial: the performance the program was assumed to have when the trial's designers calculated how many participants were needed. The FDA's summary calls these higher values the prespecified thresholds, and the estimates exceeded both the floors and the thresholds. Because the test applies to the lower end of each interval, a study has to clear its floor with room to spare. A small study cannot, because its intervals are wide.
A threshold turns a score into a decision
Most models do not output a decision. They output a score, and a threshold turns the score into a decision: above the threshold, refer; below it, do not. Moving the threshold trades one error for the other. A lower threshold flags more patients, catching more disease and raising more false alarms; a higher one does the reverse. The receiver operating characteristic curve, or ROC curve, plots that trade across every possible threshold: sensitivity against the false-positive rate, which is the share of disease-free patients wrongly flagged. The area under the curve of Chapter 11 is the area beneath this curve. It summarizes how well the score separates the two groups whatever the threshold, from 0.5 for a coin toss to 1 for perfect separation.
The Google network of Chapter 10 shows the choice in practice. Its authors scored it on 8,788 gradable images from EyePACS-1, a set they kept apart from all the data used to train and tune the network, and its area under the curve was 0.991. The authors call such a held-out set a validation set; in the terms of Chapter 11 it is a test set. Each threshold gives one operating point on the curve, with its own sensitivity and specificity, and the authors reported two: a high-specificity point with 90.3 percent sensitivity and 98.1 percent specificity, and a high-sensitivity point with 97.5 percent sensitivity and 93.4 percent specificity. The model was the same at both points. Which point to ship depends on what a false alarm and a missed case each cost in the setting of use. That is a clinical judgment rather than a property of the model, and the point must be fixed before the test, because a threshold tuned on the test set inflates the result, as Chapter 11 warned.
An area under the curve describes the whole curve, including thresholds nobody will use. Two models with the same area can have very different sensitivity at the specificity a clinic needs. A claim can be interpreted only when it comes with the threshold, the sensitivity and specificity at that threshold, their confidence intervals, the prevalence in the test population and the share of cases the device declined.
Prevalence changes what a positive means
Sensitivity and specificity describe the device; predictive values describe what its answers mean in a particular population, and they change with how common the disease is. The PCR and immunoassay guides in this library work through the same arithmetic for laboratory tests, and the arithmetic applies unchanged to software.
Take a population in which 24 percent have the disease, as among the trial's analyzed participants, and apply IDx-DR's observed sensitivity of 87.4 percent and specificity of 89.5 percent. Of every 1,000 people, 240 are diseased and the program flags 210 of them; 760 are healthy and it flags 80 of them by mistake. A positive answer is right 210 times out of 290, about 72 percent, close to the trial's 72.7 percent. Now take a population in which 7 percent have the disease, closer to an unselected screening population than the trial's enriched sample: 70 diseased, of whom 61 are flagged, and 930 healthy, of whom 98 are flagged. A positive answer is now right only 61 times out of 159, about 38 percent, while a negative answer is right about 99 percent of the time. The program is unchanged, but the meaning of its answers has changed.
Low prevalence and a weak score together produce a heavy burden of false alarms. In 2021 researchers at Michigan Medicine tested a sepsis prediction model on 38,455 hospitalizations, 7 percent of which involved sepsis, the body's life-threatening response to an infection. The model was built into a widely used electronic health record. Its area under the curve was 0.63, against 0.76 to 0.83 reported in the developer's own documentation. At the alert threshold the hospital used, the model flagged 18 percent of hospitalizations, caught 33 percent of sepsis cases and missed 67 percent. Only 12 percent of the hospitalizations it flagged involved sepsis: if the model alerted once per patient, clinicians would evaluate about eight patients to find one with sepsis. Each rate follows from the published counts; for instance, 843 of the 6,971 flagged hospitalizations involved sepsis.
Answers the program declines to give
A diagnostic program can also decline to answer, and how those cases are counted changes the figures. IDx-DR returned "insufficient quality" for some participants even after their pupils had been dilated, and its labeling states that such patients should be referred. The trial's headline sensitivity and specificity exclude them: the program gave a usable answer for 96.1 percent of participants whose reference images could be graded. A complete account therefore reports how often the device answers, as well as its sensitivity and specificity when it does. Chapter 17 recomputes IDx-DR's sensitivity with the declined cases counted. The immunoassay guide takes up the same question for a blood test that has a zone of results giving no answer.
Decide with clinicians what a missed case and a false alarm each cost in the intended setting, choose the threshold on the tuning data accordingly, and freeze it before the test set is opened. Report sensitivity, specificity and their intervals at that threshold, the prevalence of the test population, predictive values for the populations where the device will be used, and the rate of declined answers.
- Sensitivity, specificity and predictive values all come from four counts; IDx-DR's trial gave 87.4 percent sensitivity and 89.5 percent specificity.
- A threshold turns a model's score into a decision, trading missed cases against false alarms along the ROC curve; it should be chosen with clinicians and fixed before the test.
- Predictive values change with prevalence: the same program's positive answers are right about 72 percent of the time at 24 percent prevalence and about 38 percent at 7 percent.
- A full account also reports how often the device declines to answer, and the operating point behind any summary number.
+ Part IV · Patients the model never met
Performance where it is used.
IDx-DR is not labeled for children and young adults. In two studies of children and young adults with diabetes, its specificity was about 79 percent, against about 90 percent in the adults of the 2017 trial. Among patients without the disease, that difference means roughly twice as many false referrals. The two results are not strictly comparable. The youth studies used their own reference readings rather than the adult trial's reading-center protocol, and the company's founder was an author of the later study. But the patients were different, and so was the result. A model's accuracy is a measurement made on particular people, images, references and conditions, and it changes when any of them changes. The four chapters of this part follow those changes: what happens when a model meets a different setting, how it can be accurate on average and wrong for a group, what changes when a clinician and a model decide together, and how to read a published performance claim for what it does and does not cover.
Tested on yesterday's patients.
+ The questionWhy does a device whose model passed its test do worse in another clinic, on another camera, or a year later?
Eleven clinics in Thailand
By one press account, Thailand had about 4.5 million people with diabetes, and around 200 retinal specialists to examine their eyes, when Google and the Thai Ministry of Public Health began deploying a screening system in primary care. The system's model was a deep-learning network: a neural network with many layers. Over eight months, researchers made regular visits to 11 clinics in the provinces of Pathum Thani and Chiang Mai, where they watched nurses photograph patients' eyes and interviewed the nurses about the system. The researchers presented the study at a conference on human-computer interaction in 2020. The study did not measure grading accuracy. Instead, it found that the system often would not grade an image at all.
The system had been built to refuse images below a quality threshold, a sensible safety choice in the laboratory. In the clinics, lighting differed from room to room and images came out blurred or with dark areas, so the system marked those images ungradable. A press account of the study reported that more than a fifth of images were rejected. Under the first protocol, every patient whose image was rejected was referred to a specialist. The researchers changed that protocol so that eye specialists reviewed the ungradable images together with each patient's records, instead of referring every such patient automatically. Most of the study's authors were Google staff. They described the core finding as a tension between the image quality the model required and the images an under-resourced setting could produce.
Google's system in a prospective cohort
A separate study in Thailand measured accuracy directly. It was a prospective cohort study: it enrolled a group of patients and collected data as their screening happened, rather than looking back at stored records. Between December 2018 and March 2020, nine primary-care sites in the national screening program used the system in real time, while regional retina specialists also graded every image as a safety measure. Of 7,940 people screened, 7,651 were analyzed, and 31.5 percent were referred. The authors published the results in 2022. For vision-threatening retinopathy, they reported that the system's sensitivity was 91.4 percent against 84.8 percent for the specialists, and its specificity 95.4 percent against their 95.5 percent. The width of the reported confidence intervals suggests that these figures come from a subset of the patients rather than from all 7,651. The paper's authors include Google staff, and its funding came from Google and a Thai hospital.
Taken together, the two studies show one common pattern in how performance changes between a test and real use. In the cohort study, the model's accuracy matched or beat that of the specialists. What went wrong in the clinics arose outside the grading model: in the quality check that rejected images, and in the workflow around that check, from the lighting and the quality of the images to the referrals that followed a rejection. A test on curated images measures only one step of a longer process, which runs from taking the photograph, through the quality check and the grading, to the referral. In the Thai clinics, the steps that such a test does not measure decided much of what patients experienced.
Kinds of shift
A model is fitted and tested on one distribution of cases: patients, images and labels in particular proportions. Dataset shift is any difference between that distribution and the one the model meets in use, and it comes in recognizable kinds. Input shift, also called covariate shift, changes what the model sees: a different camera, different lighting, a different scanner protocol, an older or younger population. Prevalence shift changes how common the disease is, and Chapter 13 showed that prevalence changes what a positive answer means. Concept shift changes what counts as the right answer: a new grading guideline, a different reference. Acquisition shift arises inside the chain of processing that produces the image: a software update on a scanner, a new image format or a change in compression, each invisible to a person looking at the image and possibly visible to a model.
The study behind the pneumonia network of Chapter 11 also measured a shift between hospital systems. A network trained only on Mount Sinai's images had an area under the curve of 0.802 on Mount Sinai's test data, and its area fell to 0.717 on the NIH's. Time produces shifts too. Populations change, treatments change the appearance of disease, and clinical practice changes which patients are tested at all. A model tested once gives evidence only about the cases it was tested on, at the time it was tested.
In the adult trial, IDx-DR cleared 89.5 percent of participants without the disease, so among every 1,000 such adults it would refer about 105 unnecessarily. In the youth studies its specificity was about 79 percent, so among every 1,000 young people without the disease it would refer about 210 unnecessarily, twice as many. Sensitivity in one of those studies was reported as about 86 percent, close to the adult figure. A screening program that used the device outside its labeled age range would send about twice as many people without the disease to its specialist clinics, and no study in adults could have shown that cost.
External and prospective tests
Two kinds of test reach further than a test set drawn from the same data as the training set. Chapter 11 introduced the external test, which uses data from sites, devices or periods that contributed nothing to development; it shows how well the model performs beyond the data it was built on. A prospective test runs the device forward in time, in its intended setting, on patients enrolled for the purpose, with the workflow and the operators it will really have. The IDx-DR trial was a prospective test in primary care, and for that reason its figures carry more weight than those of a retrospective test on archived images. The principles of good machine-learning practice state that testing should demonstrate performance under clinically relevant conditions, and both kinds of test address that principle.
A model that holds its accuracy at a second hospital has passed one more test, not the test of every hospital. Sites differ in equipment, protocols, populations and workflow at once, and a single external result cannot separate those effects. Several external sites, chosen to span the intended use, give far more evidence than one large site.
Neither kind of test removes the problem of shift, because the next site is always new. External and prospective tests can show that performance holds across the kinds of difference the intended use contains, and a report of such tests can state the conditions they ran under, so that a user can see when a site falls outside them. Chapter 19 describes the check a hospital can run before it relies on a device, and Chapter 20 describes the monitoring that continues afterwards.
For every test set, record the sites, the devices and their software versions, the image formats, acquisition protocols, operators, population characteristics and dates. State them in the labeling as the conditions under which performance was shown, and give users a way to compare their own setting with those conditions before use.
- In Thai clinics a retinal screening system rejected many images taken in real lighting, and the workflow around those rejections decided much of what patients experienced.
- In a prospective Thai cohort, the same kind of system matched or beat retina specialists' accuracy, so much of what went wrong in the clinics lay outside the grading model.
- Dataset shift changes inputs, prevalence, the right answer or the acquisition chain; outside its labeled age range, IDx-DR's specificity was about 79 percent in youth studies with their own reference readings, against about 90 percent in the adult trial.
- External and prospective tests show whether performance holds beyond the data a model was built on, and the conditions they cover belong in the claim.
Accurate on average, wrong for some.
+ The questionCan a model be accurate overall and still fail one group of patients, and how would anyone find out?
A test set built to compare skin tones
Three published models for telling malignant skin lesions from benign ones had been reported to score areas under the curve between 0.88 and 0.94 on their own test sets. In 2022 a group at Stanford tested the three models on a new collection, the Diverse Dermatology Images set: 656 images from 570 patients seen at Stanford clinics between 2010 and 2020. A biopsy report confirmed every diagnosis, and the images of the lightest and darkest skin tones were matched by diagnosis, age, sex and date.
The scores on the new set come from the group's preprint, the version of its paper shared before peer review. On the whole set, the three models scored between 0.56 and 0.67. Split by skin tone, the drop was steeper for darker skin: the best of the three models scored 0.72 on the lightest skin types and 0.57 on the darkest, and another scored 0.61 and 0.50, no better than chance.
The published paper reports two further findings. Dermatologists read the same images, and when their readings were scored against the biopsy results, they too were less accurate on darker skin. The authors note that dermatologists often label the images in AI datasets, so the same weakness can affect the labels that models are trained and tested on. The group also fine-tuned two of the models: it continued their training on images from the new set. Fine-tuning closed the gap between light and dark skin, which suggests that the original failure came largely from what the models had been trained on.
Averages hide groups
A single accuracy figure for a whole test set is an average over its members, weighted by how many of each kind it contains. If one group is small, the average can be high while that group's performance is poor, and the headline figure does not reveal the poor result. A 2021 study examined chest-radiograph classifiers trained on three large public datasets, which together held more than 700,000 images, and measured how often each classifier called a sick patient's radiograph normal. The rate was higher for female patients, for patients under 20, for Black and Hispanic patients, and for patients insured through Medicaid, the US public health insurance program for people with low incomes. The rate was higher still where these characteristics combined, for example in Hispanic women compared with white women.
The remedy is to look at each group. The IDx-DR trial reported the make-up of its participants: 28.6 percent were African American, 63.4 percent White and 1.6 percent Asian; 16.1 percent were Hispanic, an ethnicity recorded separately from race; and their median age was 59. The investigators analyzed sensitivity by age, sex, race, ethnicity, blood-sugar control, site and lens status, which records whether an eye still has its natural lens. None of these factors had a significant effect on sensitivity, and specificity was somewhat higher in participants over 65. An analysis of this kind cannot prove that no group does worse. It can show that no large difference was detected among the groups present in meaningful numbers, and it tells a reader which groups were too small for any conclusion.
A subgroup result of 90 percent from a few dozen patients is compatible with performance far below the overall figure. Reporting the single figure alone, without its interval, invites the conclusion that the group is served well. A reader needs the interval, and the number of patients behind it, to judge whether the evidence is strong enough to conclude anything.
How many patients a subgroup needs
The precision of a subgroup estimate depends on the number of patients in that subgroup, which is usually a fraction of the whole test set. A study sized for a precise overall sensitivity may give a vague one for every group within it.
A model detects disease in 45 of 50 diseased patients in a subgroup: 90 percent. The Wilson method is a standard formula for the interval of a proportion, and the FDA uses it in the examples of its guidance on diagnostic statistics. By that method the 95 percent confidence interval runs from 78.6 to 95.7 percent; the exact method of Chapter 13 gives a slightly wider 78.2 to 96.7 percent for the same 45 of 50. With 450 detected out of 500, the estimate is still 90 percent, but the interval narrows to 87.1 to 92.3 percent. With 9 out of 10, the interval spans 59.6 to 98.2 percent and leaves the subgroup's sensitivity almost unknown. To show that a subgroup's sensitivity is within a few points of the overall figure, the subgroup needs hundreds of diseased patients of its own. That need is why subgroups are planned when the study is designed, not discovered afterwards.
Looking at many subgroups creates a problem of its own. The usual test calls a difference significant when chance alone would produce one that large less than 1 time in 20, so a study that compares twenty subgroups will usually find one or two that differ by chance alone. Investigators therefore name the subgroups in the protocol before they see the data, and they treat a difference found afterwards as a question for the next study rather than a conclusion. When a planned subgroup does fall short, a maker has three honest responses. It can add data from that group and retrain the model, which is what fine-tuning did for the dermatology models. It can narrow the claim: IDx-DR's labeling covers only the adults its trial enrolled. Or it can tell users plainly where performance is unproven. Each response changes the device or its labeling, and needs the evidence that such a change requires.
Representative data, by rule
Regulators have written representativeness into their expectations. The good machine-learning practice principles of 2021 stated that clinical study participants and data sets should be representative of the intended patient population, and the international version of 2025 states that clinical evaluation should use datasets representative of that population. When the FDA granted the stroke-triage program of Chapter 12 its De Novo in 2018, it created a regulation for radiological triage and notification software in general. Its special controls require results showing effective triage across relevant subgroups, which the decision summary defines by clinical factors and by how the images were acquired.
India's 2026 guidance on medical device software is more demanding for AI. It states that makers should disclose the demographic and geographic composition of their training, validation and test data, justify models trained or validated outside India, and assess bias, generalizability and robustness across relevant Indian sub-populations. The AI Act is the European Union's law on artificial intelligence, and it sets duties according to the risk that a system poses. For high-risk systems, it adds duties of data governance: rules on how the data used to train, validate and test a system are collected, prepared and checked. Those data must be relevant and sufficiently representative, with appropriate statistical properties, where applicable, for the groups of people the system is intended for, and the instructions for use must give the system's performance for those groups where appropriate. For AI in medical devices these duties apply from 2 August 2028, on the timeline that Chapter 29 sets out.
Representative does not mean proportional. A subgroup that is rare in the population but at high risk may need to be over-sampled so that its performance can be measured at all: the study includes more of its members than their share of the population would give. The study then corrects for the over-sampling when it reports overall figures, as the IDx-DR trial did for its enrichment. What a regulator asks to see is that the groups the intended use contains were present, in numbers that support a conclusion, and that their results are reported.
List the subgroups the intended population contains, including those defined by skin tone, sex, age, ethnicity, disease severity, device model and site. Size the study so that each group whose performance matters has enough diseased and healthy patients for a useful interval. Report every planned subgroup with its interval, and say plainly which groups were too small to assess.
- Dermatology models with reported areas under the curve of 0.88 to 0.94 on their own test sets scored 0.50 to 0.57, at or near chance, on the darkest skin in a balanced, biopsy-confirmed set.
- An overall figure can hide poor performance in a small group; chest-radiograph classifiers missed disease more often in several under-served groups.
- A subgroup estimate is only as precise as the number of patients in it: a sensitivity of 90 percent from 50 diseased patients has an interval of about 79 to 96 percent.
- The US and India already expect data representative of the intended population and results across subgroups, and the EU's AI Act will require representative data for high-risk AI in medical devices, with group results where appropriate, from 2 August 2028.
The reader and the machine.
+ The questionWhen AI assists a clinician rather than deciding alone, what has to be measured: the model, the person, or the pair?
When the suggestion is wrong
An assistant that is right most of the time can still make an expert worse. In a study published in 2023, 27 radiologists read 40 test mammograms. Each mammogram came with a suggested category on the standard breast-imaging scale, which grades how likely a finding is to be cancer, and the readers were told that the suggestions came from an AI system. In fact the investigators had set the suggestions, and 12 of the 40 were deliberately wrong. When the suggestion was right, readers of every level of experience rated about 80 percent of mammograms correctly. When it was wrong, inexperienced readers rated 19.8 percent correctly, moderately experienced readers 24.8 percent and very experienced readers 45.5 percent.
The pattern has a name: automation bias, the tendency to accept a machine's output in place of one's own judgment. A 2012 systematic review found that it varies with a person's experience of the task and with trust and confidence, and that heavy workload, complex tasks and time pressure increase it. Automation bias turns the model's errors into the clinician's errors, and it is strongest among the people an assistant is often meant to help most, those with less experience.
In the study's design, the suggestion was right in 28 of the 40 cases, 70 percent. An inexperienced reader was right 79.7 percent of the time with a correct suggestion and 19.8 percent of the time with a wrong one, so that reader's overall score is 0.7 times 79.7 plus 0.3 times 19.8, about 62 percent. For a very experienced reader, right 82.3 percent and 45.5 percent of the time, the same sum gives about 71 percent. Of the gap of about 9.5 points between the two readers, about 7.7 points come from the cases with a wrong suggestion: 0.3 times the 25.7-point difference in how often the two readers rated those cases correctly. The investigators chose the 30 percent rate of wrong suggestions to measure the effect, not to mimic any product. A better model would make wrong suggestions rarer, and each one may then be harder to spot.
Three roles for a model
Software that analyzes clinical data plays one of three roles, and each role changes what has to be measured. An autonomous device gives the answer itself: IDx-DR's output is the screening decision, and no clinician interprets the images first, so the device's own accuracy is what reaches the patient. An assistive device gives a reader extra information, such as marks on suspected lesions, and the reader decides; the performance that matters is the reader's with the device compared with the reader's without it. A triage or notification device flags cases that may need urgent attention and works alongside the usual workflow. The stroke-triage program of Chapter 12, which the FDA authorized in February 2018, is one. The FDA's decision summary describes it as a "notification-only, parallel workflow tool". The program alerts a specialist to a suspected blocked artery, while "trained radiologists read all images per standard of care, regardless of the performance" of the device.
The role also decides where automation bias can do harm. An autonomous device has no human to bias, but no human to catch its errors either. A parallel triage tool is designed so that its misses do not remove the usual reading by a radiologist, although its alerts may still change which images are read first. An assistive device places its output directly in front of the reader at the moment of decision, which is where the mammography study showed the risk. Clinical decision support is software that gives clinicians information or recommendations. The FDA last revised its guidance on such software in January 2026. Its criteria address the same risk: software that a clinician cannot independently review, or that supports a time-critical decision where the clinician is likely to rely on it without review, is a device software function.
Measuring the pair
For an assistive device, the evidence comes from a reader study. The FDA describes the usual design in its guidance on computer-assisted detection in radiology, which it last revised in September 2022. Many readers each read many cases, with and without the device. In a "fully-crossed" design, every reader reads every case in both conditions, and the guidance states that such a design "offers the greatest statistical power for a given number of cases."
Under the guidance, reading sessions are separated "by at least four weeks to avoid memory bias," so that readers do not remember cases from an earlier session. The primary measure is usually the area under the ROC curve for readers with the device against readers without it, and the number of readers must be justified rather than fixed by rule. The guidance states that case sets may be enriched with diseased cases, but that enrichment can change how readers behave and biases predictive values, the problem of Chapter 13 in another form.
The good machine-learning practice principles name the target directly. The 2021 version stated that the focus should be on the performance of the human-AI team, and the international version of 2025 states that a device should be assessed with a focus on human-AI interaction in the intended use environment. A model's standalone figures remain necessary, because they show what the model does on its own, without a reader. For an assistive device, however, they do not show what patients receive.
A strong model and a capable reader can work badly together if the reader follows the model's errors and overrides its correct calls. Only a reader study can show whether the pair beats the reader alone, and its cases must include the model's failures at a realistic rate. Standalone accuracy cannot answer that question.
Telling users what they need to know
Clinicians can resist a wrong suggestion only if they know when the model is likely to be wrong. In June 2024 the FDA, Health Canada and the MHRA published guiding principles on transparency for machine-learning devices. The principles state that makers should share the device's intended use, performance and limitations with the right users, at the right time and in a form designed for those users, and should also share the logic behind its outputs where that can be done. IDx-DR's labeling states the patients, the camera and the conditions under which performance was shown. It also states that an "insufficient quality" result after dilation may itself be a sign of disease that needs referral. Chapter 17 reads that kind of information as an outside reader would, and asks what it leaves out.
Decide whether the device is autonomous, assistive or parallel, and design the display for that role. Show the model's confidence or the evidence behind a suggestion where that helps review, and avoid showing a suggestion before the reader has looked at the case independently, because a suggestion seen first can anchor the reader's judgment. Measure the reader with and without the device in a study whose cases include the model's errors at a realistic rate, and train users on the failure modes the study finds.
- In a 2023 mammography study, wrong suggestions labeled as AI cut readers' accuracy from about 80 percent to between 20 and 46 percent.
- Autonomous, assistive and parallel triage devices place the model differently, and each role changes what has to be measured.
- Reader studies measure the clinician with and without the device, ideally with every reader reading every case in both conditions and with reading sessions at least four weeks apart.
- Users can resist a wrong output only if they know when the model is likely to be wrong, and the 2024 transparency principles ask makers to share the device's intended use, performance and limitations with them.
What a decision summary says, and leaves out.
+ The questionWhat should a reader look for in a software device's published performance, and what will it not tell them?
A public document with the evidence in it
The FDA publishes a decision summary for every device it grants through De Novo, and a shorter summary for most devices it clears through a 510(k). IDx-DR's decision summary is free to download. It gives the indications for use and the limitations, and it describes the software and its architecture. It states the special controls that every later device of the same type must meet, summarizes the software documentation and the clinical study, lists the risks and their mitigations, and ends with the agency's judgment that the probable benefits outweigh the probable risks. For an outside reader, such as a hospital evaluating a product or an engineer studying a competitor, the decision summary is the most complete regulatory account of the evidence.
The summary also has to be read carefully. Its heading gives "DATE OF DE NOVO: January 12, 2018", which is the date the FDA received the request; the FDA's database lists 11 April 2018 as the date the request was granted. The summary's table of results gives an observed sensitivity of 87.4 percent with a 95 percent confidence interval of 81.9 to 92.9 percent, but an exact interval runs from 81.9 to 91.7 percent when it is computed from the trial's counts, 173 of 198, and that is the interval the device's 2022 510(k) summary prints. The date is not a slip: other decision summaries, such as the one for the cardiac ultrasound program Caption Guidance, also put the date of receipt under that heading, so it misleads only a reader who takes it for the grant date. The interval appears to be a misprint, since its upper end matches neither an exact nor a Wilson interval, and although it does not change the conclusion, it shows that a reader should check a summary's figures with a calculator at hand.
Rebuilding a table's percentages and intervals from its counts takes minutes and catches transcription errors, mismatched denominators and intervals of the wrong kind. A missing set of counts is itself worth noting, because without the counts a reader cannot check how the study treated the cases the device declined.
Questions to put to any claim
A performance claim can be read against a short list. Who was tested: the intended population, how participants were recruited, and whether the study was enriched. Where and by whom: the setting, the operators and their training. Against what: the reference standard, who applied it, and whether they were masked, that is, kept from seeing the device's output. Which endpoints: the outcome measures the study was designed to test, and whether they were fixed before the study. How many: the counts behind each percentage, and the intervals. For which groups: subgroup results and the groups too small to assess. On what inputs: the camera or scanner, its software, the image format, the version of the device's own software. What happened to the cases the device declined. And who ran the study: its funding, and whether the authors work for the maker.
For IDx-DR, a reader can answer most of these questions well from the published record. The population was adults with diabetes not previously diagnosed with retinopathy, partly enriched with participants whose diabetes was poorly controlled. The setting was ten primary-care practices with operators who had never imaged an eye, and the reference was a masked reading-center protocol. The endpoints were set in advance with input from the FDA, the counts are published, results by subgroup are reported, and the camera and software version are named.
The trial's paper also discloses that its first author was a shareholder, director and employee of IDx, which funded the study, and that another author held shares and the statistician was paid by the company. By the company's account, the first author founded IDx. The record cannot describe how the device performs outside the conditions of the trial. That limit is why Chapter 14's youth studies and Chapter 20's monitoring matter.
Counting the cases it declined
The trial's headline sensitivity and specificity cover the 819 participants for whom both the reference standard and the program gave a usable result. Another 33 participants had gradable reference images but received an "insufficient quality" result from the program, and 10 of those 33 had the disease. How a study counts these 33 participants changes the result.
With the 33 left out, as in the headline figures, the program flagged 173 of 198 diseased participants: 87.4 percent. With the 10 diseased participants among the 33 counted as misses, it flagged 173 of 208: 83.2 percent. The exact 95 percent interval for that figure runs from 77.4 to 88.0 percent, so its lower end is still above the trial's floor of 75 percent. The labeling directs that a persistent insufficient-quality result lead to referral, and with the 10 counted as referrals on that basis, the program sent 183 of 208 diseased participants onward: 88.0 percent. The same choice applies to the 23 declined participants without the disease: counting them as false referrals moves specificity from 89.5 percent to 86.3 percent. A full claim states which convention it uses and gives the others. The immunoassay guide in this library recommends the same for a blood test with an indeterminate zone, a range of results read as neither positive nor negative.
What a summary leaves out
Some information is absent by design. A decision summary rarely describes the training data, the design of the model inside the software or how its threshold was chosen. It reports the clinical evidence for the locked device, not how the device was built. It describes performance at authorization, not in use: IDx-DR's summary cannot report how the program has done in the clinics that bought it. It also describes the device only as authorized, with the inputs the summary names, and says nothing about cameras or populations outside the label.
Other gaps come from the difference between a regulator's words and a seller's. In a press release in 2018, the FDA called IDx-DR the first device authorized to provide "a screening decision without the need for a clinician to also interpret the image or results." In its own materials, the maker speaks of a diagnosis and calls the device "De Novo-cleared", although the FDA grants a De Novo rather than clearing it. A developer's own performance figure can differ widely from an independent one: the developer of Chapter 13's sepsis model reported an area under the curve of 0.76 to 0.83, and an outside team measured 0.63. The rule that follows is to read a claim in the regulator's record and the peer-reviewed paper, with the maker's own statements labeled as such. An independent evaluation is worth more than any of them.
Before buying, deploying or competing with a software device, fill in a single page: intended population, setting, operators, reference standard, prespecified endpoints, counts and intervals, subgroups, inputs and versions, the handling of declined cases, the study's funding and authors, and any independent evaluation. Mark each answer with its source, and treat any blank line as a question for the maker or a reason for a local test.
- A decision summary is the most complete regulatory account of a device's evidence, but IDx-DR's gives the receipt date under the heading of the De Novo date and prints a sensitivity interval that the trial's counts do not reproduce, so a summary should be read with a calculator.
- A claim is read against who was tested, where, against what reference, with which endpoints, counts, subgroups, inputs and sponsors.
- How declined cases are counted changes the result: IDx-DR's sensitivity is 87.4, 83.2 or 88.0 percent depending on the convention.
- Summaries leave out training data and real-world performance, and a maker's wording and figures need labeling and independent checks.
+ Part V · Into the hospital, and after
Deploying, watching and changing a model.
On 10 June 2021 the FDA cleared a new version of IDx-DR. The new version could receive images in DICOM, the format hospitals use for medical images, and it put its instructions on screen step by step. It let clinics configure file names and keep images locally, and it told operators during the exam whether an image was good enough to send. None of these changes altered what the program claimed to detect, and the FDA cleared the version without new clinical data. The changes were made to fit the device into the way clinics work. A year later the FDA cleared another version, in which the maker had replaced one of the models inside the analysis. The four chapters of this part follow a model from authorization into use: connecting it to a hospital's systems, checking it on the hospital's own patients, watching it once it runs, and changing it without losing the evidence. By the end of the part, the central question has its answer.
Fitting the software into a hospital.
+ The questionAn authorized AI device arrives at a hospital. What has to be connected, configured and agreed before it is used on the first patient?
The label on every image
Every image that a hospital's scanners and cameras produce in the DICOM format carries a header: a list of fields describing where the image came from. Each field is identified by a tag made of two four-digit codes. The field tagged 0008,0060 records the modality, such as CT or ophthalmic photography, and the field tagged 0008,0070 records the manufacturer of the equipment. Field 0008,1090 holds the model name, 0018,1020 the versions of the software running on the equipment, 0008,1030 a description of the study and 0010,0020 the patient's identifier. In DICOM a study is the set of images from one examination. The DICOM standard defines thousands of such fields. It is revised several times a year, and a 2026 release was current in October 2026.
The header is an AI device's first safeguard: the device can check it for each image before it analyzes the image. A device authorized for images from one camera can read the manufacturer and model fields and refuse anything else, and it can record the software version of the equipment that produced each image, so that a later change in performance can be traced to an update on the scanner. Many fields, however, are set by local protocols or typed by staff, and two hospitals can describe the same examination differently. A device that decides which images to analyze from their free-text descriptions will meet descriptions its makers never saw.
Where the software sits
An AI device rarely works alone. A scanner sends each study to the picture archiving and communication system, or PACS, which stores the hospital's images and displays them to radiologists. An imaging model receives the study from the PACS and returns its results to the PACS, to the radiologist's worklist of studies waiting to be read, to the electronic health record or to a specialist's phone. The model may run on a server in the hospital or in the maker's cloud. Each of those connections is an interface with its own standard.
DICOM carries images and the results derived from them. HL7 version 2 carries orders and reports as messages; HL7, the standards body that publishes it, says that 95 percent of US healthcare organizations use it. FHIR is HL7's newer standard, built from web resources that can each be addressed individually. It has been at release 5 since March 2023. In July 2026 release 6 was in its second normative ballot, a formal vote on making its text a stable standard.
Standards for the AI step itself are younger. The IHE initiative writes profiles, which describe how vendors should combine existing standards for a task. It has published one profile for AI results, which defines how analysis results are encoded in DICOM objects and displayed, and one for AI workflow, which defines how a request for analysis is sent, managed and performed. In October 2026 both were still at trial implementation, the stage before final text. A hospital connecting several AI products therefore often builds part of each integration itself, or buys a platform that does that work.
Inputs inside and outside the specification
A device's authorization covers specified inputs. IDx-DR is indicated for use with the Topcon NW400, and the public summary of its 2021 clearance states a minimum image resolution of 22 pixels per degree of the camera's field of view. Inside those limits the trial's evidence applies; outside them it does not, whatever the images look like. When an input falls outside its specification, a device has two choices: it can refuse the input, with a message the user can act on, or it can analyze the input anyway and risk an answer the evidence never covered. IDx-DR refuses: it returns its "insufficient quality" output, and since 2021 it has also given feedback during the exam so that the operator can retake the image.
Inputs also change after installation. A hospital may change a scanner protocol to reduce dose, or a vendor may update the scanner's reconstruction software, which computes the image from the scanner's raw measurements. The PACS may begin compressing images to save storage, or a new interface may convert images to a different format on the way to the model. Such a change can be invisible to a person reading the images and still be visible to the model. That kind of change is the acquisition shift of Chapter 14. In a hospital study that Chapter 19 describes, a change of image format, more than any change in the patients, cut a model's performance from excellent to no better than chance.
A model's inputs can move outside the conditions it was tested under without any change to the device itself. A new scanner protocol, a reconstruction update or compression in the archive can each do so, and so can a format conversion in an interface engine, the software that translates and routes data between hospital systems. Agreeing in advance who notifies whom of such changes is part of installing the device.
A retrospective comparison with usual care tested the stroke-triage program authorized in 2018 on 44 cases with a confirmed blocked artery that the program had flagged correctly. With the program, the specialist was notified a median of 5.6 minutes after the CT angiogram; when notification came through the radiology report, the median was 51.5 minutes. A 2023 randomized trial tested the same company's software at four Houston stroke centers, in 243 treated patients. The time from arrival to the start of clot removal fell by an estimated 11.2 minutes from a median of 100, about 11 percent. The two studies measured different intervals in different hospitals, so their numbers cannot be subtracted. Together, though, they show that the model shortens one step and that the other steps decide how much of the gain reaches the patient: transferring the images, getting the alert to the right person, assembling the team and having the procedure room ready. The trial found no significant difference in patients' functional outcome, their level of disability after the stroke. That was an exploratory measure, one the trial was not designed to settle.
Who answers for what
The maker answers for the device within its specified environment, and under the EU's medical device regulation it must state that environment: manufacturers "shall set out minimum requirements concerning hardware, IT networks characteristics and IT security measures" needed to run the software as intended. The hospital answers for the network, the interfaces, the other systems and the way the device is used. The integration lies between them, and neither party controls it alone.
The standard for that integration is IEC 80001-1. Its 2021 edition sets general requirements for how an organization applies risk management before, during and after it connects a medical device or health software to its IT infrastructure, covering safety, effectiveness and security. ISO has already marked the standard for revision and intends to replace it with a new standard in the 81001 series. An earlier technical report in the same family describes responsibility agreements: written statements of what each party does across the life of the connection. The parties include the hospital, the IT supplier and the device maker. The questions such an agreement must settle are concrete: who tests the interfaces, who approves changes on each side, who is told when a scanner or the archive is updated, and what clinicians do when the AI is unavailable.
Before the first patient, list every interface the device depends on and test each with real local data, and confirm that the local images and their headers fall inside the device's specification. Measure the time from acquisition to result against the clinical need. Define what clinicians do when the device is down or declines an input, and sign a responsibility agreement that names who notifies whom of changes on either side.
- A device can read the DICOM header of every image to identify the equipment and software behind it, refuse inputs outside its specification and trace changes.
- AI devices depend on interfaces to PACS, health records and cloud services; the IHE profiles for AI were still trial implementations in October 2026.
- A model's evidence covers only inputs inside its specification, and changes on the hospital's side can move inputs outside it without any change to the device.
- The maker states the required environment, the hospital manages the connection under IEC 80001-1, and a responsibility agreement divides the rest.
Site acceptance, local performance and computerized system validation.
+ The questionThe maker validated the software; why does the site have to validate it again, and how much is enough?
Two validations, two questions
A device that arrives at a hospital has already been validated by its maker, so a second validation can look redundant. It is not, because the two validations answer different questions. The maker's validation shows that the device meets its intended use in its specified environment, on the patients and inputs of its studies. The hospital's question is whether that evidence carries over to this hospital: its scanners and their settings, its interfaces, its patients and its way of working. Only the hospital has the data to answer that question.
Professional societies now say so plainly. In a joint statement published in January 2024, the radiology societies of the United States, Canada, Europe, Australia and New Zealand called testing on local data, with local systems and workflows, "essential", and recommended that each site "perform a statistically rigorous evaluation of performance on their own local data." Where such an evaluation is not feasible, the societies advised comparing the local data with the vendor's test data and, where the two differ, proceeding "with great caution, if at all."
A trial with its outputs hidden
In August 2020 a team at the Hospital for Sick Children in Toronto began running a deep-learning model on live clinical data with its outputs hidden from clinicians, an approach the team calls a silent trial. The model predicted, from ultrasound images of the kidneys, which children with swelling of the kidney would need surgery. Its development data covered 1,643 kidneys from 294 patients, and on a random 20 percent of those data, held out from training, the model's area under the curve had been 0.90. The silent trial covered 523 kidneys from 150 patients seen between August and December 2020, and there the area under the curve was 0.50, no better than chance.
The team traced the fall to three differences. The live patients were younger, the share of obstructed right kidneys was higher, and the images reached the model in a different form. The development data had been processed JPEG files, while the live data arrived as unprocessed PNG files that looked different to the model even after the same preprocessing. Adjusting for age and side barely helped, raising the area under the curve to 0.51. Reprocessing the live images to match the original processing raised it to 0.84 to 0.85. After the team retrained the model on the original and silent-trial data together, a second silent trial on 711 kidneys from 202 patients gave 0.91 to 0.92. The decisive fault was not clinical at all. It lay in the integration chain of Chapter 18, and it came to light only because the team ran the model on the hospital's own data before anyone acted on the model's outputs.
Acceptance and a local check
A site's own validation has two layers. The first is acceptance: confirming that the device is installed as specified, that each interface passes the right data, that results reach the right screen for the right patient, and that the fallback procedure works when the device is unavailable. The second is a local performance check: running the device on a sample of the site's own cases, silently or retrospectively, and comparing its outputs with a local reference. The societies suggest focusing that review on the predictive values that clinicians will see in practice, identifying the cases where the device would change care, and categorizing its false positives and negatives before deciding whether to deploy. They recommend repeating the review whenever the AI software or the equipment used with it changes.
A hospital wants evidence that a device's sensitivity on its patients is above 75 percent, and it expects the sensitivity to be about 87 percent, as in the device's studies. The hospital plans the check at one-sided 5 percent significance and 80 percent power. The check then has a 5 percent chance of wrongly concluding that sensitivity is above 75 percent when it is really 75 percent, and an 80 percent chance of showing that it is above 75 percent if it really is 87 percent. The standard calculation for comparing a proportion with a fixed value gives about 69 diseased cases. If 10 percent of the hospital's screened patients have the disease, the check needs about 690 consecutive cases. With 60 detections out of 69, 87 percent, the exact one-sided lower confidence bound is about 78 percent, above the target: at the same 5 percent level, the result rules out any sensitivity below that bound. A check run on a few dozen convenient cases cannot reach that conclusion, which is why the societies call for statistical rigor rather than a demonstration.
The results of the check become the baseline for the monitoring of Chapter 20: the distribution of inputs, the rate of declined cases, and the local sensitivity, specificity and predictive values against which later drift is measured.
Computerized system validation
Some sites must go further, because the rules they work under require it. Laboratories, blood services, organizations running clinical trials and pharmaceutical manufacturers work under good-practice rules, known collectively as GxP, and under these rules the user of any computerized system that affects product quality, patient safety or data integrity must validate that system. Many of these organizations follow ISPE's GAMP 5, the guide of Chapter 8, whose second edition appeared in July 2022. GAMP 5 is risk-based: the validation effort scales with the system's impact, complexity and novelty. Its authors describe it as placing patient safety and product quality ahead of compliance for its own sake, and as supporting iterative and agile methods.
GAMP 5 sorts software into categories: infrastructure software, such as operating systems and databases, and then a continuum from products used without configuration, through configured products, to custom software. The more a system is configured or customized for the site, the more verification the site must do itself; for a standard product, the user relies on the supplier's evidence after assessing the supplier. The work then follows a sequence of steps. The site writes user requirements that state what it needs, a risk assessment shows where a failure would matter, and the system is specified, configured and verified.
Verification has traditionally taken the form of installation, operational and performance qualification, which confirm in turn that the system is installed correctly, operates as specified and performs as needed in routine use. Increasingly it takes the form of risk-based testing, some of it scripted in advance and some unscripted (Chapter 9). A traceability matrix links each requirement to its tests, release into use is controlled, and the system is reviewed periodically for as long as it is used. Data integrity applies at every step: records must be attributable to the person who made them, complete and protected, with audit trails that log who changed what and when. The FDA has set requirements for such electronic records since 1997, in its rule on electronic records and signatures, 21 CFR Part 11.
With the supplier's documentation, a site can avoid repeating tests that do not depend on its environment. The documentation cannot cover the site's interfaces, configuration, data or patients. A site that accepts a package without assessing the supplier, or without local performance data, leaves its own question unanswered.
ISPE has extended the method to AI. The 2022 edition of GAMP 5 included an appendix on machine learning, which set out a life cycle for a machine-learning subsystem, and in July 2025 ISPE published a separate GAMP guide to AI, covering data governance, model testing, monitoring and change management for AI-enabled systems used in GxP work. In July 2025 the European Commission also put out for consultation a new annex on AI to the EU's good manufacturing practice rules. As drafted, the annex would admit only static, deterministic models in critical manufacturing uses: models that do not change during use and always give the same output for the same input.
The draft annex would not apply to generative AI, which produces text, images or speech rather than a score, or to large language models, which are trained on very large amounts of text to generate text. It would also require acceptance criteria at least as high as the performance of the process the model replaces. As of October 2026 the annex was still a draft. Its scope is medicines manufacturing, not medical devices, but its rules on test data read like a summary of Chapters 11 to 13.
The builder's standard and the user's
IEC 62304 and GAMP 5 describe the same software from opposite sides. IEC 62304 is written for the organization that builds device software, and under it the maker records how the device was made. GAMP 5 is written for the organization that uses a computerized system, and under it the user records evidence that the system does what that user needs, in that user's configuration. A laboratory that works under good-practice rules and installs an AI-enabled analyzer needs both: the maker's evidence, and its own validation built on it. So does a contract research organization, a company that runs clinical trials for sponsors, when it uses an AI reading tool in a trial. Makers can shorten the site's validation by supplying a validation package with their product: requirements and traceability, test scripts the site can re-run, and installation and operational qualification documents.
Before purchase, write the site's requirements and acceptance criteria, including the local performance check and its sample size, and ask the supplier for its validation package and its specified environment. Where GxP rules apply, scope the work with GAMP 5's categories and risk assessment; in every case, run the device silently or retrospectively on local cases before clinicians rely on it, and keep the results as the monitoring baseline.
- The maker's validation covers the intended use in a specified environment; only the site can show that the evidence carries over to its systems and patients.
- In a silent trial in Toronto, a model's area under the curve fell from 0.90 to 0.50, recovered to 0.84 to 0.85 once an image-format mismatch in the integration was fixed, and reached 0.91 to 0.92 after retraining.
- A local check needs enough diseased cases to support a conclusion, about 69 in the solved example to show a sensitivity above 75 percent, and its results become the monitoring baseline.
- Where GxP rules apply, sites validate AI systems with GAMP 5's risk-based method, which ISPE has extended to AI, and build on the supplier's evidence.
When the true answer arrives late.
+ The questionHow can a maker, or a hospital, tell a model is still working when the right answer arrives months later, or never?
Truth arrives late
A model can be wrong for months before anyone can know. When a screening program tells a patient to come back in a year, nobody checks whether that answer was right until the patient returns, if the patient returns at all. When the program refers a patient, the specialist confirms or overturns the answer weeks later, and the result is kept in another clinic's records. For most outputs of most diagnostic models, the reference standard of Chapter 12 is never applied at all in routine use. The direct measure of performance is agreement with the right answer, and that agreement is available late, partially and only for some patients.
The early months of the COVID-19 pandemic showed how quickly the conditions of use can change in the meantime. A study of 24 hospitals in four US health systems found that daily alerts from the sepsis prediction model of Chapter 13 rose by 43 percent while the number of patients in the hospitals fell by 35 percent, ahead of the surge in COVID admissions. The patients had changed, and the model's inputs had changed with them. Michigan Medicine was one of the four systems, and the university's account of the study states that the rise in alerts there led to the alerts being paused. Nobody needed to wait for confirmed diagnoses to see that something had shifted, because the change was visible in the number of alerts the model itself produced.
What can be watched at once
Monitoring therefore starts with what is available immediately. The inputs can be checked as they arrive: the equipment and software versions in the image headers of Chapter 18, measures of image quality, patient age and the other characteristics the device records. The outputs can be counted: the share of positive results, the share of declined cases and the distribution of scores. None of these measures shows directly whether the model is right. Each can show within days that the conditions the model was validated under no longer hold, and that finding is the trigger for a closer investigation.
Suppose a screening device with 87.4 percent sensitivity and 89.5 percent specificity runs at a site where 10 percent of patients have the disease. The device should return a positive result for about 18.2 percent of patients: 8.7 percentage points come from the diseased patients it detects, and 9.5 from false alarms among the healthy. At 200 screens a week, that is about 36 positives a week, and chance alone makes the weekly count vary around that figure with a typical spread of about 5. If a camera fault lowers specificity to 80 percent, the expected positive rate rises to about 26.7 percent, or about 53 positives a week, which is more than three typical spreads above the baseline. A simple chart of weekly positives, with its alarm line set three typical spreads above the expected count, would flag the change within a few weeks, months before referral outcomes could confirm that the extra positives were false.
Some drift does not show in counts of inputs and outputs. A 2017 study at US Veterans Affairs hospitals followed seven models that predict acute kidney injury, a sudden loss of kidney function, over the nine years after the period their development data came from. The seven included machine-learning models and regression models, the traditional statistical kind that adds up weighted risk factors; the word has nothing to do with the regression tests of Chapter 7. All kept their discrimination, the ability to rank patients by risk. What drifted was their calibration, the agreement between the risks a model predicts and the rates that actually occur.
Every model came to overstate the risk as the rate of kidney injury changed, and the overprediction grew steadily for the regression models while staying small for the random forest, a model that combines many decision trees, and for the neural network. If a hospital had used one of the regression models with a fixed threshold on predicted risk, the model would have alerted more and more often for the same patients, while its area under the curve stayed stable throughout.
A model's discrimination can hold while its calibration drifts. Monitoring that tracks only the area under the curve can miss a drift that moves every decision made at a fixed threshold, and that kind of drift changes alert rates and referrals.
Closing the loop with outcomes
The direct measures come later. A site can collect the specialist findings for referred patients and match them to the device's outputs, which gives a running estimate of the positive predictive value. It can also send a random sample of negative results for a reference reading, which is the only way to estimate missed cases before patients return with them. The maker uses complaints, adverse event reports and service records in its postmarket surveillance, the system through which it reviews information about a device once the device is on the market. In their joint statement of 2024, the radiology societies recommend re-evaluating each AI tool on updated local data "at specified intervals, but at least annually," and after any new version. They note that "naturally occurring data drift will cause AI model performance to degrade over time," and they ask institutions to define in advance how a problem will be escalated and resolved.
Who watches
Responsibility for monitoring is shared, and its division is not yet settled. The maker has postmarket obligations everywhere a device is sold, and the last of the good machine-learning practice principles states the expectation that deployed models are monitored for performance and that the risks of retraining are managed. The hospital sees the inputs, the workflow and the outcomes, and only the hospital can link the device's outputs to its own patients' later records. In September 2025 the FDA asked for public comment on how the real-world performance of AI-enabled devices should be measured and evaluated, with questions about metrics, monitoring methods, data sources and the triggers for a response. The FDA described its request as proposing no policy, and no follow-up had been published by October 2026. Under the EU's AI Act, hospitals that deploy high-risk AI will take on monitoring duties, on the timetable set out in Chapter 29.
What a monitoring plan needs is clear even where the division of labor is not. It needs a baseline from the site validation of Chapter 19, measures of inputs and outputs watched continuously, a reference sample and outcome matching on a schedule, thresholds that trigger investigation, and named people on each side who act when a threshold is crossed. When monitoring shows that performance has moved, the response is a change to the device, its labeling or its use, and Chapter 21 sets out how a model can be changed without losing the evidence behind it.
Define the input checks, the output and declined-case rates to chart, their baseline from the local validation and the limits that trigger investigation. Schedule a reference reading of a random sample of negatives and the matching of positives to specialist findings, assign who on the maker's and the site's side reviews each, and state what happens when a limit is crossed: investigation, restricted use, suspension or retraining.
- For most outputs of a diagnostic model the right answer arrives late or never, so accuracy cannot be monitored directly in real time.
- Checks on inputs, output rates and declined-case rates can show within days that conditions have shifted, as the rise in sepsis alerts did early in the pandemic.
- Calibration can drift while discrimination holds, so monitoring must watch decisions at the threshold, not only the area under the curve.
- Reference readings of sampled negatives, outcome matching and agreed triggers close the loop, and the maker and the site share the work.
Agreeing in advance what may change.
+ The questionHow can a model be retrained after authorization without a new review every time, and what keeps the retrained model as good as the one that was authorized?
From a discussion paper to a final guidance
On 2 April 2019 the FDA published a discussion paper on a new way to regulate changes to machine-learning software. It asked for feedback and was explicitly not a draft guidance. The paper started from the way such software improves: by retraining. It argued that if every retrained version needed a new premarket review, makers would either freeze their models at the first version or file a submission for each improvement. The FDA proposed instead that a maker describe in advance a "region of potential changes", which the paper called the pre-specifications, together with an algorithm change protocol: the methods the maker would use to make, test and control those changes. If the FDA authorized the plan with the device, the maker could make changes inside it without a new submission.
The idea took five years to become final guidance. In October 2023 the FDA, Health Canada and the UK's MHRA published five guiding principles: a plan should be focused and bounded, risk-based, evidence-based and transparent, and it should take a view of the whole product lifecycle. The FDA issued its final guidance on predetermined change control plans for AI-enabled device software on 4 December 2024 and reissued it on 18 August 2025. It had also issued a broader draft guidance on such plans for all devices in August 2024, and as of October 2026 that guidance was still a draft.
What a plan contains
Under the final guidance, a plan has three parts. A description of modifications lists the planned changes and the characteristics and performance the modified device will have: retraining on new data from the same kinds of sites, for example, or adjusting a threshold within stated limits. A modification protocol describes how the maker will develop, validate and implement each change. It says how new data will be collected and kept separate, how the model will be retrained and its performance evaluated, which acceptance criteria the model must meet, and how users will be told. An impact assessment sets out the benefits and risks of carrying out the plan and how the maker will control those risks.
The plan has firm limits. Every modification must stay within the device's intended use, and generally within its indications; a change of what the device is for needs a new submission. Changing the plan itself generally needs a new submission too. The device's labeling should say that it contains machine learning and has an authorized plan. The public summary of the authorization should describe the plan, so that users and other makers can see what may change without further review. A plan can be authorized through a 510(k), a De Novo or a premarket approval.
A plan that lists permitted changes but sets loose or self-adjusting acceptance criteria gives the regulator nothing to rely on. In this guide's view, the criteria that matter are fixed before any modification is made. The modified model must perform at least as well as the authorized version, overall and in each named subgroup, on a sequestered test set and against the same reference standard. A modified model that misses these criteria does not ship. A sequestered set is the sealed test set of Chapter 11: held apart from all training and tuning data, and not used by the developers before the final test.
IDx-DR's 2022 change
IDx-DR went through such a change before the guidance existed, and its record with the FDA shows the evidence its maker used. When the FDA granted IDx-DR's De Novo in 2018, one of the special controls it set already required "a protocol... that describes the level of change in device technical specifications that could significantly affect the safety or effectiveness of the device." That was an early, device-specific form of the same idea. In June 2022 the FDA cleared version 2.3, whose changes included a new classifier that judges image quality inside the analysis. To show that the change did no harm, the maker re-ran the new version on the images from the 2017 trial, 850 of which could be evaluated, and compared its answers with those of version 2.0.
Against the reading-center reference, version 2.3 detected 171 of the 195 diseased participants it analyzed, 87.7 percent; version 2.0 detected 173 of 198, 87.4 percent. Version 2.3 cleared 553 of the 614 participants without disease that it analyzed, 90.1 percent, against version 2.0's 556 of 621, 89.5 percent. On the participants each version analyzed, both figures rose slightly. Counted over the same 850 participants, however, version 2.3 detected two fewer diseased participants and cleared three fewer without disease. Its percentages rose because it declined more people. The new quality classifier changed which participants received an answer. After operators had retaken and resent images as the protocol allowed, version 2.3 gave a usable result for 809 of the 850 participants, 95.2 percent, and version 2.0 for 819, 96.4 percent: version 2.3 declined ten more, 1.2 percent of the 850. The 96.1 percent of the opening scene is a different count, over the 852 participants who had a reference grade. At the trial's case mix, a program screening 10,000 people a year would send about 120 more of them for re-imaging or referral because of image quality alone. The re-run was a regression test on the original patients. It showed that the new version performed close to version 2.0 on them while declining more of them, and it could say nothing about patients the trial never included.
Plans in use
The plans authorized so far point to a pattern: they permit bounded changes to locked models, and their public summaries say little about monitoring. One of the earliest belongs to Caption Guidance, a program that guides ultrasound users to acquire images of the heart; the plan came with its De Novo authorization in February 2020. When the maker modified the plan through a 510(k) that September, the summary stated that all algorithm modifications would be "trained, tuned, and locked prior to release", and the plan excluded continuously learning algorithms. A peer-reviewed study published in July 2026 counted, in FDA data to April 2026, 170 devices of all kinds, not only AI devices, cleared with a change control plan across all review panels. Of these, 34 were radiology AI devices, 22 of them cleared in 2025. The study found that the public summaries described neither ongoing performance monitoring nor predefined triggers for retraining. The insulin-dosing app UpDoc, cleared in December 2025, has a plan that lists five categories of permitted change, from new insulin formulations to alternative ways of entering data; none of them alters the cleared dosing logic.
A plan changes what the maker may do without a new review. It does not change what a hospital needs to know. Each new version alters the device that the site validated in Chapter 19. A site that relies on its baseline needs to be told what changed, to re-run its local check when the change could move performance, and to reset the monitoring limits it set under Chapter 20.
The answer to the central question
Parts II to V together answer the central question. For conventional software, the evidence that it will be right for the next patient is a lifecycle: requirements derived from risk and traced to tests, an architecture that confines the dangerous code, an account of every borrowed component, verification at each level and regression testing after every change. For an AI model, the evidence adds a test kept apart from training and drawn from the setting of use, against a reference fit for the claim, reported with its uncertainty, by subgroup and with an account of the cases the model declined. Where a clinician acts on the output, it adds a study of the clinician with and without the device. That evidence is then checked at each site, on inputs inside the device's specification, and watched in use. What keeps that evidence true after an update is change control: every change is judged against the claims it could move, and a model's changes are bounded by a plan agreed in advance that fixes the tests each new version must pass. Part VI turns to attacks, a threat that can arrive at any time, whether or not the device has been updated.
Decide before the first submission which changes the model is likely to need: retraining on new data, new input devices, threshold adjustments. Describe them, the data they will use, the sequestered test sets and the acceptance criteria overall and by subgroup, and how users and sites will be told of each new version. Keep the plan narrow enough that each permitted change can be tested against criteria fixed today.
- The FDA's 2019 discussion paper led to final guidance in December 2024: under a predetermined change control plan, changes authorized in advance proceed without a new submission.
- A plan holds a description of modifications, a modification protocol with fixed acceptance criteria, and an impact assessment, all within the intended use.
- When IDx-DR's maker re-ran the 2017 images through version 2.3, cleared in 2022, accuracy on the analyzed participants rose slightly because the new version declined ten more of the 850.
- Each new version changes the device a site validated, so sites must be told, re-check performance where needed and reset their monitoring.
+ Part VI · Defending the device
Cybersecurity.
On 28 August 2017 Abbott wrote to physicians that a firmware update was ready for several of its pacemakers, formerly made by St. Jude Medical. The update was meant to close vulnerabilities in the devices' radio link, weaknesses that an attacker could use. It could not be sent remotely. Each patient had to visit a clinic, where the update was sent through the wand of the programmer, the clinic's instrument for reading and setting an implanted pacemaker. The update took about three minutes, and the pacemaker ran in a backup mode while it was installed. Press reports of the FDA's communication put the number of affected devices in the US at about 465,000. Nothing in those pacemakers had failed. According to a cardiology society's summary of the FDA's communication, the flaws could allow intrusions, and some of those intrusions could affect how the device operated. The three chapters of this part deal with weaknesses of that kind: what threats a device faces, which defenses must be designed in before release, and how weaknesses found after release are judged and fixed.
An attacker is an input.
+ The questionHow can a device that meets every requirement still be made to harm a patient?
A pump that could not be patched
On 27 June 2019 the FDA warned patients and clinicians that certain Medtronic MiniMed insulin pumps were being recalled because an unauthorized person "could potentially connect wirelessly to a nearby MiniMed insulin pump and change the pump's settings." The recalled pumps were the MiniMed 508 and several models of the Paradigm series. They communicated by radio with remote controllers and glucose meters, and the protocol they used lacked proper authentication, the check that confirms who sent a message, so a pump could not tell its own patient's devices from an attacker's transmitter within range. Medtronic had identified about 4,000 patients in the US who were potentially using vulnerable pumps.
The remedy was unusual. The FDA said that Medtronic was "unable to adequately update" the pumps "with any software or patch," and the company offered patients replacement pumps with stronger built-in protection. The FDA said it was not aware of any confirmed patient harm. On the same day CISA, the US cybersecurity agency, published an advisory that described the weakness as improper access control. According to the advisory, only an attacker nearby could exploit it, doing so required high skill, and there was no known public exploit, meaning no published code or method built to use the weakness. The FDA posted its recall record for the MiniMed 508 in March 2020, and the recall was still open in October 2026.
Safety risk and security risk
The pumps met their requirements. They delivered insulin as programmed, alarmed as specified and passed their verification. What they lacked was a requirement that nobody had written in the form that mattered: refuse commands from anyone but the patient's own devices. Like every software fault in this guide, the fault was present on the day the pumps shipped; what differed was the input that reached it, which was not an unlucky combination of events but a deliberate one.
That difference changes how risk is judged. Safety risk management estimates how likely a sequence of events is, following the ISO 14971 chain of Chapter 4. An attacker is not a sequence of events with a probability; an attacker chooses the inputs, searches for the rare combination and repeats it at will. For that reason the FDA's 2026 guidance on cybersecurity in premarket submissions describes security risk assessment as non-probabilistic. Such an assessment judges a vulnerability by its exploitability, how easily an adversary could use it, and by the harm that would follow. The FDA also treats known vulnerabilities as reasonably foreseeable. Where a security weakness could lead to patient harm, the two assessments overlap, and the security control that blocks the attack is also a safety control.
In its 2019 advisory, the US cybersecurity agency scored the pump weakness 7.1, High, on version 3 of the Common Vulnerability Scoring System. The score comes with a vector, a string that records each factor behind it: AV:A/AC:H/PR:N/UI:N/S:U/C:L/I:H/A:H. From left to right, its factors are these: the attack vector is adjacent, meaning within radio range rather than over the internet, and the attack complexity is high. The attacker needs no privileges, meaning no access rights on the device, and no action by the user. The scope is unchanged, meaning the effects of an attack stay within the component that has the weakness. The impact is low on confidentiality, the secrecy of data. It is high on integrity, whether data stay correct, and on availability, whether the device stays usable, because settings and delivery could be changed or stopped. In a 2018 advisory on the same pumps' remote controllers, two weaknesses had been scored 4.8 and 5.3, Medium. They scored lower because each affected only one of confidentiality or integrity, and one also needed user interaction. The score summarizes technical severity but says nothing about how many patients use the device or what a wrong dose of insulin does, so a maker adds its own assessment of the clinical harm alongside the score.
The attack surface
Every way data or commands can enter a device is part of its attack surface: radio and Bluetooth links, Wi-Fi and Ethernet, USB and serial ports, the connection to a cloud service, a service interface used only by technicians. Since 29 March 2023 a US law has set requirements for every premarket submission for what it calls a "cyber device": among other conditions, a device that includes software and can connect to the internet. Chapter 29 sets out those requirements. The FDA reads the condition of connecting to the internet broadly, "whether intentionally or unintentionally, through any means": Wi-Fi, cellular, Bluetooth, magnetic inductive links, USB, Ethernet and serial ports all count, and so does a USB connection used briefly for service.
A threat model lists those entry points, the assets behind them and the adversaries who might want those assets. The assets include the dose settings, the patient's data and the device's software itself. For each path, the threat model records what an attacker could do and what would stop the attack. In 2021 MITRE, a US non-profit research organization, and the Medical Device Innovation Consortium published a playbook for threat modeling medical devices, with funding from the FDA. The FDA's premarket cybersecurity guidance expects the threat model to be part of the design and to be updated as the architecture of Chapter 5 changes. The special controls set for IDx-DR in 2018 already required a cybersecurity vulnerability and management process, because the device's images and answers cross the internet between the client and the server.
Estimates that an attack is unlikely rest on assumptions about who would try, with what skill and for what gain, and those assumptions go out of date. Research, published exploits and automated tools lower the skill an attack needs. Judging a weakness by how easily it could be used, and by the harm that would follow, stays valid longer than judging it by how likely an attack seems today.
When security becomes safety
Some attacks threaten patients directly, as a changed insulin dose would. Others threaten data, and an attack on data can undermine trust and care. In January 2025 the FDA warned that the software of the Contec CMS8000 patient monitor could send patient data outside the care setting and contained a backdoor, a hidden way into the software that bypasses its normal access controls. The monitor was also sold under another name. The FDA advised hospitals that relied on the monitors' remote monitoring to unplug them and stop using them, and advised other users to use them for local monitoring only. The fix released in July 2025 removed the monitors' networking entirely: the remedy for the unsafe connection was to have no connection at all.
The pump and the monitor show two limits of device security. A defense that was not designed in may be impossible to add later, and for some weaknesses the safest response is to give up a function. Chapter 23 describes the defenses that have to be in the design from the start.
Start from the architecture diagram of Chapter 5 and mark every interface, including service and debug ports and cloud connections. For each, list what could be read, changed or stopped through it, by whom and with what effort, and the control that prevents it. Judge each path by exploitability and clinical harm rather than by an estimate of how likely an attack is, and update the model with every design change.
- MiniMed pumps recalled in 2019 accepted radio commands without proper authentication and could not be patched, so patients were offered replacements.
- Security risk is judged by exploitability and harm, not probability, because an attacker chooses the inputs; known vulnerabilities count as foreseeable.
- Every interface is part of the attack surface; US law and the FDA's broad reading make almost any connectable device a cyber device.
- A threat model built from the architecture names each path, its harm and its control, and some weaknesses can be removed only with the function.
Built to be defended.
+ The questionWhich defenses have to be designed in before release, because they cannot be added later?
A signature checked at every start
A firmware image is one file: the complete program a device will run, compiled, tested and released under a single version number. Before the file leaves the maker, a release system computes a fingerprint of it, the cryptographic hash of Chapter 6, which changes completely if a single bit of the file changes. The release system then signs the fingerprint with the maker's private key, a secret value kept on a guarded machine. The signature is shipped with the file. The device holds the matching public key, which can check a signature but cannot make one, and a small piece of start-up code, both in memory that cannot be rewritten. Every time the device powers on, the start-up code computes the fingerprint again, checks it against the signature, and runs the firmware only if the two agree.
This sequence of checks is called secure boot, and the protected code and key at its start are the root of trust. Each stage checks the next before handing over control, so a program that someone has altered, however slightly, is refused before it runs. In May 2018 NIST published guidelines on firmware resiliency for computing platforms in general. They describe three duties that such a design serves: protecting the firmware against unauthorized changes, detecting changes that occur anyway, and "recovering from attacks rapidly and securely," usually by returning to a copy known to be good.
The design depends on something present in the hardware from the first unit shipped: protected memory for the key and the start-up code. A device built without such memory cannot gain it through an update. Memory that cannot be rewritten is part of the hardware, and without it the device has no trusted code to tell a genuine update from a forged one.
Why a checksum is not a lock
Embedded software has long carried checksums, most often a cyclic redundancy check, or CRC, which detects a file corrupted by a faulty memory chip or a noisy cable. In its premarket cybersecurity guidance, revised on 3 February 2026, the FDA states that such checks "do not provide integrity or authentication protections in a security environment," and it tells makers not to rely on them as security controls. A CRC is computed by a published method from the file alone, so anyone who changes the file can compute a new CRC that matches. A signature can be produced only with the private key, so a matching signature proves both that the file is unchanged and that it came from the holder of the key.
The same distinction applies to every defense in this chapter. A safety mechanism protects against chance: a corrupted bit, a failed sensor, an unlikely input. A security control must hold against someone who knows how the mechanism works and is trying to defeat it, the adversary of Chapter 22. Many safety mechanisms and security controls look alike from outside; their designs differ in what they assume about the person who supplies the input.
A lifecycle with security in it
The standard for building security into health software is IEC 81001-5-1, published in December 2021. It was written to extend a device maker's existing process rather than to run beside it. Its scope is based on IEC 62304, its clauses follow IEC 62304's order, and it takes its security activities from IEC 62443-4-1, the equivalent standard for industrial control products. Where IEC 62304 has a software risk management process, IEC 81001-5-1 has a security risk management process. It adds security tasks such as threat modeling, security requirements, security testing and the handling of vulnerabilities after release. A team already following IEC 62304 therefore extends its procedures rather than writing a second set. The standard has been flagged for revision since September 2025.
The FDA calls such a process a secure product development framework, and in its 2026 guidance it gives two standards as examples of one: IEC 81001-5-1 and ANSI/ISA 62443-4-1, the US adoption of IEC 62443-4-1. For the security risk assessment itself, the FDA points to two documents from AAMI. The first is a technical report, TIR57, published in 2016. The second is the standard ANSI/AAMI SW96, published in 2023 and aligned with the ISO 14971 process of Chapter 4.
Eight kinds of control
In the first appendix of its guidance, the FDA sorts the defenses a device may need into eight categories: authentication; authorization, which decides what an authenticated sender may do; cryptography; the integrity of code, data and execution; confidentiality; event detection and logging; resiliency and recovery; and firmware and software updates. The list could serve as a summary of what the MiniMed pumps lacked. Their radio protocol did not authenticate the sender, so a pump could not tell its patient's remote from an attacker's transmitter. A 2018 advisory on the pumps' remote controllers had also described a replay attack, in which recorded radio traffic could be sent again to trigger a bolus when the remote options were enabled.
In the same appendix, the FDA recommends a control for each of those faults. It asks makers to "use cryptographically strong authentication, where the authentication functionality resides on the device," and not to use passwords "that are hardcoded, default, easily guessed, or easily compromised." For commands that could do harm, it asks for anti-replay measures, such as a number used only once in each exchange. It also asks that updates be authenticated before they are installed, and that hardware-based security be used where possible.
The FDA expects the architecture that a maker submits to show the device in four views: the whole system with its connections, the paths by which one compromise could harm many patients, how updates travel from the maker to every unit, and the security of each use case. The tests it lists include penetration testing, in which testers attack the device as an adversary would, and fuzz testing, which feeds the software large numbers of automatically generated, randomly varied inputs to find those that make it fail. The list also includes tests with malformed and unexpected inputs, analysis of the attack surface and of how weaknesses can be chained, and software composition analysis of the compiled binaries, which identifies the third-party components they contain and the known vulnerabilities of those components.
Several features have to be decided early in design: protected memory for keys, a processor able to check signatures quickly, room for a second firmware image and a radio protocol with authentication. A device shipped without them can sometimes be protected by its surroundings, but it can never be fully repaired, and the MiniMed recall shows that the remedy may be a new device.
Designing for the patch
A device that can be updated has a security control the MiniMed pumps lacked: a way to change its own code after release. The Abbott pacemakers of 2017 had such a way, though only through the programmer's wand in a clinic, and Abbott's letter to physicians shows that the update path itself carries risk. Abbott recommended against replacing the devices to avoid the vulnerability. It asked physicians to weigh the update's risks for each patient, with particular care for patients who depend on their pacemaker for every heartbeat.
In its letter, Abbott gave three failure rates from its own earlier firmware updates: the device reloading its previous firmware after an incomplete update, 0.161 percent; loss of the programmed settings, 0.023 percent; complete loss of device function, 0.003 percent. Per 100,000 updates, those rates mean about 161 devices left working on the old version, 23 needing to be reprogrammed, and 3 that stop working. For a patient whose heart does not beat without pacing, the last outcome is the one that matters. For that reason Abbott advised physicians to consider the update for such patients where temporary pacing was available. The most common failure was also the safest, because the device kept its previous firmware and returned to it when the new one did not load completely. Keeping the old firmware as a fallback was a design decision, and it left most devices whose update failed still working on their old version.
The solved example shows the decision a maker faces after release. A patch closes a vulnerability and opens a small, measurable risk of its own. For these pacemakers, Abbott stated in its letter that it had received no reports of a compromised device. Good design makes that trade easier. A well-designed update path authenticates the update, writes it to a spare slot in memory, checks it, and keeps the old version until the new one has started correctly. Chapter 24 follows what happens when a vulnerability comes to light after release, in code the maker did not write.
Decide at the start which hardware the device needs for security: protected key storage, signature checking at start-up, room for a fallback image. Add authentication, anti-replay and authorization to every interface in the threat model, plan how updates will reach every unit and how a device recovers from a failed update, and list the security tests, from fuzzing to penetration testing, in the verification plan.
- Secure boot checks a signed firmware image against a key held in protected hardware at every start, so altered code never runs.
- A CRC detects accidents but not attackers; only a signature made with a private key proves a file is genuine.
- IEC 81001-5-1 adds security risk management and security tasks to an IEC 62304 process, and the FDA gives it as an example of a secure product development framework.
- Controls such as authentication and an update path with a fallback must be designed in; a device built without them may have to be replaced.
The threat changes while the code does not.
+ The questionOnce a device ships, how are new weaknesses in its code and its components found, judged and fixed?
Eleven flaws in someone else's code
In July 2019 eleven vulnerabilities were disclosed in IPnet, a network stack: the part of an operating system that sends and receives data over a network. Six of them were serious enough to let an attacker on the network run code of the attacker's own choosing. Armis, the security company that found them, named them URGENT/11. It reported that the affected code had been part of the VxWorks real-time operating system since version 6.5, and of several other operating systems that had used the same stack. On 1 October 2019 the FDA warned patients, clinicians and makers that medical devices and hospital networks using those operating systems could be affected. The FDA named six operating systems from six suppliers. It said that device makers' notices so far included an imaging system, an infusion pump and an anesthesia machine, and that the agency knew of no adverse events.
No device maker had written IPnet. The stack reached their products inside operating systems that the makers had chosen and treated as SOUP under the rules of Chapter 6. The code in those devices had not changed since release. What changed was public knowledge about the code, and with that knowledge the risk. URGENT/11 shows the general pattern of security after release: the threat changes while the code does not, and most of the code at issue came from someone else.
Matching the list to the news
When a vulnerability is published, the first task is finding out which products contain the affected component, and in which version. The software bill of materials of Chapter 6 exists for that search. A maker that keeps a machine-readable SBOM for every released version can search all of them for the component within hours; a maker that must ask each development team, or inspect old build files, may take weeks. In its 2026 premarket guidance, the FDA expects makers to monitor sources such as NIST's National Vulnerability Database and to assess the known vulnerabilities in each component. It also expects them to give particular weight to those listed in CISA's catalog of vulnerabilities already being exploited in real attacks.
Whether the component is present is only the first question. A device may contain the vulnerable component without using the vulnerable function, or without exposing that function to any interface an attacker can reach. The maker records its answer for each vulnerability and each product as a VEX statement, short for Vulnerability Exploitability eXchange. In 2021 NTIA, the US Commerce Department's telecommunications agency, described a VEX as an assertion about the status of a vulnerability in a specific product, with four values: not affected, affected, fixed, or under investigation.
Vulnerabilities were also disclosed in another widely used network stack, made by Treck. According to a trade report, by June 2020 the infusion pump maker B. Braun had received 24 patches from Treck for its Outlook 400ES pump and had determined that 20 of them did not apply. The 4 that applied were 17 percent of the 24. Each of the 20 needed a reasoned statement of why the pump was not affected; each of the 4 needed an assessment of exploitability and harm, and a fix. The amount of that work grows with the number of components in the bill of materials. A code-audit vendor reported a mean of 1,180 open-source components per application (Chapter 6). If a product carries that many components, and 1 percent of them receive a new advisory in a year, about 12 advisories arrive for that product each year. At B. Braun's ratio, about 2 of the 12 would need a fix, and every one would need a written conclusion. Without an SBOM, the first step for each advisory is the slowest: finding out whether the component is present at all.
Judging what a weakness means for patients
A vulnerability that applies is judged on two scales. The first is technical severity, usually expressed with the Common Vulnerability Scoring System, the system behind the MiniMed score read in Chapter 22. The system's fourth version was published on 1 November 2023 and added a supplemental Safety metric. That metric records whether exploitation could cause injury, using the consequence categories of IEC 61508, the standard for functional safety: the part of safety that depends on electrical, electronic and programmable electronic systems working correctly. The specification states explicitly that supplemental metrics do not change the score. Under an FDA contract, MITRE wrote a rubric for applying the scoring system to medical devices. In October 2020 the FDA qualified the rubric as a medical device development tool, which means the agency has evaluated the tool and accepts that it works as intended within a stated context of use.
The second scale is clinical. In its postmarket cybersecurity guidance of December 2016, the FDA divides risks into controlled and uncontrolled by their residual risk, the risk that remains after controls are in place. A risk is controlled where the residual risk of patient harm from a vulnerability is acceptable, and uncontrolled where it is not.
For an uncontrolled risk, the agency states that it does not intend to enforce its usual reporting rules for corrections if four conditions hold. The first is that no serious adverse events or deaths are known. The other three concern the maker: it must tell its customers within 30 days of learning of the risk, with interim measures and a plan; fix the risk, validate the fix and distribute it within 60 days; and take part in an information-sharing organization for the sector. Since 29 March 2023 US law has required a maker that submits a cyber device to the FDA to make updates and patches available, with timing that depends on the vulnerability. Known unacceptable vulnerabilities must be fixed on a "reasonably justified regular cycle," and critical ones that could cause uncontrolled risks "as soon as possible out of cycle."
Coordinated disclosure
Many device vulnerabilities are found by people outside the company: academic researchers, security firms, hospital engineers. Outside researchers reported the MiniMed flaws, and a security firm found URGENT/11. Coordinated vulnerability disclosure is the practice by which they report a weakness privately to the maker, the maker acknowledges and investigates it, and the two agree when the details are published, ideally with a fix or mitigation available. A government coordinator often works between the two. The US government's coordinator, now part of CISA, published the MiniMed advisories of 2018 and 2019 after researchers reported the flaws and Medtronic analyzed which models shared them.
Two international standards describe the maker's side. ISO/IEC 29147 covers receiving and publishing reports, and ISO/IEC 30111 covers handling them inside the company; both were flagged for revision in 2025. Since March 2023 US law has required a maker that submits a cyber device to the FDA to include a plan for monitoring and addressing vulnerabilities after release, "including coordinated vulnerability disclosure and related procedures."
The hospital's share
A fix protects no patient until it is installed, and for many devices installation happens in the hospital. Each update has to be scheduled around clinical use and tested against the hospital's own configuration where the integration of Chapter 18 could be affected. It is then applied unit by unit, sometimes by the maker's service engineers and sometimes by the hospital's clinical engineering staff. In December 2020 Armis, which had found URGENT/11, estimated that 97 percent of the affected devices in operational technology had not been patched. Operational technology means the computers that control machines and physical equipment rather than handle office data. The estimate covers far more than medical devices, but it shows how slowly fixes reach equipment that is not a personal computer.
When a maker publishes a patch, the open risk moves from the maker's work queue to the hospital's. Until a unit is updated, its protection depends on what surrounds it: restricted access, disabled services, monitoring and network segmentation, which keeps the device on a separate part of the network. Hospitals need the maker's SBOM and VEX statements to know which units are exposed, and makers need to know which versions hospitals have installed in order to see which patients remain at risk.
The EU's medical device guidance on cybersecurity was issued in 2019 and revised in 2020. It describes security as a joint responsibility of makers, integrators, operators and users, while noting that only makers carry the legal duties. It also states that a device should not depend on security controls in its environment. In practice, the defense of a device in the field is shared: the maker designs the device to be patched, watches for new weaknesses and issues fixes; the hospital keeps an inventory of what it runs, installs fixes and protects what it cannot yet fix.
Keep the SBOM of every released version searchable, monitor vulnerability databases and the catalog of exploited vulnerabilities daily, and record a VEX statement for each match. Judge each applicable vulnerability by exploitability and clinical harm, set a target date in the regular cycle or out of it, publish a disclosure policy with a contact address, and tell hospitals which versions are affected and what to do until they can patch.
- URGENT/11 put eleven vulnerabilities into devices through an operating system's network stack that no device maker had written.
- With a searchable SBOM a maker finds affected products within hours, and a VEX statement records whether each one is actually exposed.
- Severity scores measure the weakness; FDA guidance ties its relief from reporting to fix deadlines for uncontrolled risks, and US law requires makers of cyber devices to make patches available on a regular cycle for known unacceptable vulnerabilities, or out of cycle for critical ones.
- Researchers report through coordinated disclosure, and a fix protects patients only once it is installed, often in the hospital; maker and hospital share the defense.
+ Part VII · The field in 2026
Who builds software devices.
The FDA last updated its public list of AI-enabled medical devices on 4 September 2026. The list is a table of authorizations, not a count of devices: each row is one marketing authorization, and the agency says the list "is not a comprehensive resource." The FDA's page gives no total. Independent analysts counted 1,451 entries decided by the end of 2025, and 1,524 including decisions up to 30 March 2026. In November 2025 the director of the FDA's device center told an advisory committee that the agency had authorized more than 1,200 AI-enabled devices. The single chapter of this part uses the list and the makers' own records to describe the field as it stood in October 2026: which devices reach patients, who makes them, and where generative AI stands. It is the chapter of the guide most tied to its date, and it will be the first to go out of date.
Software and AI devices in 2026.
+ The questionWhich software and AI devices reach patients in 2026, who makes them, and where does generative AI stand?
Mostly images
About three in four entries on the list are radiology devices. In a count reported in July 2026, one analyst put radiology at 1,104 of the 1,451 entries decided by the end of 2025, or 76 percent, and at about the same share of the devices authorized in 2025 alone. The rest come from other specialties, among them ophthalmology, the specialty of IDx-DR.
The reasons follow from the earlier chapters. Images are digital from the moment they are made, and they carry the standard headers of Chapter 18. Hospitals have kept them for decades in archives alongside the reports that describe them, which gives makers training data with something close to a label attached. A reading task also has a natural reference against which a model's answers can be scored: the expert reader of Chapter 12.
Regulation has reinforced the pattern. When the FDA granted a De Novo to the stroke-triage program ContaCT on 13 February 2018, it created a new classification for radiological software that flags suspected findings and notifies a specialist. Later devices of the same type could then be cleared through the faster 510(k) route, by showing that they were substantially equivalent to one already on the market, and later devices were cleared that way. A first authorization in a category makes that faster route available to the products that follow.
Who makes them
The same analyst also counted radiology authorizations by company through the end of 2025. Five makers of scanners and imaging systems led: GE HealthCare with 120, Siemens Healthineers with 89, Philips with 50, Canon with 45 and United Imaging with 38. Aidoc, a company that makes only AI software, came next with 31, and then DeepHealth with 28. Much of the equipment makers' AI runs inside or beside their own scanners and reaches hospitals with the hardware. The software specialists sell triage and detection programs that run beside the hospital's image archive, often through a platform that hosts several makers' tools, so that the hospital connects one platform rather than each tool.
A third group rarely appears on the list but accounts for much of this guide: makers of pumps, pacemakers, ventilators and monitors, whose software drives therapy rather than reading images. Most of their software is conventional code, judged by the lifecycle of Part II and defended as in Part VI. The MiniMed and Abbott cases of Chapters 22 and 23 came from this group, not from AI.
GE HealthCare said in July 2025 that it had 100 listed authorizations, and the analyst counted 120 radiology authorizations for the company by the end of 2025. Both figures can be right, because they were counted at different dates and by different rules. What is being counted matters more. An indication is one use for which a device is authorized, such as one finding that a program detects. Qure.ai, a company based in Mumbai, said in February 2026 that it held 26 FDA-cleared indications across 9 products, about 2.9 indications per product; its own regulatory page listed 7 cleared products. In January 2026 Aidoc announced a clearance covering 14 acute findings on CT from a single model, 11 of them newly cleared. If that clearance came from one submission, it is one entry counted by submission and 14 counted by indication. Two analysts counted radiology's share of the 2025 authorizations and got 75 percent and 71.5 percent, the second from 211 of 295. The gap of 3.5 points comes from the counting method and the cutoff date. A count of "AI devices" means little until it states whether it counts submissions, products or indications, and at what date.
Company figures for clearances, indications and "firsts" are written for investors and customers. Each figure can be checked against the FDA's databases, where every authorization has a number and a date and most have a public summary, and against the entries on the list itself. Where a company's figure and the regulator's record disagree, the regulator's record is the evidence.
Makers in India
Indian makers reach patients first through other regulators. The FDA cleared Qure.ai's head-CT triage program, qER, on 17 June 2020, under the classification created for ContaCT. The company's regulatory page lists CE certificates under the EU's device regulation for several products; such a certificate comes from an independent assessor and allows a product to carry the CE mark, which a device needs before it can be sold in the EU. The page also reports an approval from India's regulator but gives no number for it.
India's regulator first gathered its requirements for software devices in a single document in July 2026, its final guidance on medical device software (Chapter 26). In August 2026 an Indian manufacturer announced what it called India's first manufacturing license for diagnostic software. That claim concerns in vitro diagnostic software, used with tests on samples taken from the body, and is not specific to AI. This guide found no reliable source naming the first AI software device licensed by India's regulator.
Generative AI at the door
Generative AI reached the FDA's advisory process before it reached the list. The agency's Digital Health Advisory Committee met for the first time in November 2024 to discuss the whole product lifecycle for generative AI devices, and again in November 2025 on generative AI for mental health. At the second meeting the device center's director said that none of the digital mental-health devices authorized so far involved generative AI.
On the list page, as updated in September 2026, the FDA states that it will explore ways to identify and tag devices that incorporate foundation models, "from large language models (LLMs) to multimodal architectures." A foundation model is a large model trained on broad data so that it can be adapted to many tasks, and a multimodal model handles more than one kind of input, such as text and images. No entry on the list was tagged.
One device whose maker says it uses large language models is UpDoc, an insulin-management program for adults with type 2 diabetes. The FDA cleared it through a 510(k) on 23 December 2025 as a drug-dose calculator, class II, six months before the maker announced the clearance. The public summary of the clearance describes conversational data collection by voice or chat and insulin instructions computed from parameters set by the patient's clinician, and it reports non-clinical testing only. The summary does not use the words "large language model" or "generative AI."
On 25 June 2026 the company announced that this was the first FDA clearance of software using patient-facing large language models. The FDA has not confirmed the claim, and when a reporter asked whether generative AI made treatment decisions, the company's chief executive would not say. The clearance that Aidoc announced in January 2026 differs in kind. The cleared software is built on what the company calls a foundation model: a large image model, not a language model, doing a triage task of a kind already regulated.
For generative AI, the evidence question is the central question of this guide, asked of a new kind of software. A language model that collects a patient's answers and passes them to fixed, cleared rules is an interface, and the evidence can concentrate on whether the model captures what the patient meant. A model that writes the recommendation itself is the decision, and such a model would need the test sets, references and subgroup results of Parts III and IV for an output that can vary with every phrasing. The public record should say which of the two a device is. For UpDoc, the summary describes doses computed from parameters that the clinician sets, which points to the first.
For any software device, find its FDA submission number, decision date and summary, and note the device it was compared with and the product code, the FDA's three-letter code for the device type. Check which indications the authorization covers, whether it includes a change control plan, and what testing it reports. Treat a maker's counts and "firsts" as claims until the record supports them, and look for where any language model sits relative to the clinical decision.
- About three in four entries on the FDA's list of AI devices are radiology devices, helped by digital archives, natural references and early classifications.
- Scanner makers lead the counts and AI software specialists follow, while the makers of therapy devices, whose software is mostly conventional code, rarely appear on the list.
- Counts of AI devices depend on whether they count submissions, products or indications; the regulator's record settles disagreements.
- UpDoc, cleared in December 2025, is claimed by its maker as the first device with a patient-facing language model; the FDA record does not say so.
+ Part VIII · Quality and regulation, as of October 2026
What a software device must show.
On 27 July 2026 an EU regulation called the Digital Omnibus on AI entered into force, three days after its publication. It made several changes to the AI Act of 2024. Among them, it moved the date from which AI in products such as medical devices must meet the act's high-risk requirements from 2 August 2027 to 2 August 2028. The European Parliament had proposed a further change, moving medical devices to a section of the act's list of product laws to which lighter obligations apply, but the final compromise left medical devices where they were. Rules for software devices change in this way, through amendments, guidance and revised dates, while the mechanism beneath them stays the same. The four chapters of this part therefore put the lasting mechanism first and the clauses last: what makes software a device and sets its class, what evidence its maker must file, which updates need a regulator, and what the rules on security and AI add, including what they ask of the hospital that deploys the device. Every status in these chapters is as of October 2026.
Intended use draws the line.
+ The questionWhich software is regulated as a medical device in the US, the EU and India, and in what class?
Two guidances on one day
On 6 January 2026 the FDA reissued two guidances that together mark where its oversight of software stops. In the first, on general wellness products, the FDA stated that certain non-invasive products that estimate blood pressure, oxygen saturation, blood glucose or heart rate variability can count as wellness products, and so fall outside device regulation, when their outputs are "intended solely for wellness uses." The conditions are strict. Such a product may suggest that a user see a health professional when a reading falls outside a wellness range. It may not name a disease, call a value abnormal, use clinical thresholds or give alerts for managing a disease, and it may not show values that look clinical unless they have been validated.
In the second, on clinical decision support software, the FDA revised its criteria for software that advises clinicians without being a device. It added enforcement discretion for software that gives a single recommendation where only one is clinically appropriate. Enforcement discretion means that the FDA does not intend to enforce device requirements on such software, even where it may meet the definition of a device. The FDA also moved its treatment of time-critical decisions to the fourth of the criteria listed in the next section, the one on whether a clinician can independently review the basis of a recommendation. Chapter 16 described the effect of that criterion on software that supports a time-critical decision. It corrected the decision-support guidance on 29 January 2026. Neither document changed the law. Both moved the boundary in practice by explaining how the agency reads the law, and the boundary of software regulation usually moves in this way.
The US line
In US law a device is an instrument or related article intended for the diagnosis of disease or other conditions, or for the cure, mitigation, treatment or prevention of disease, and software can meet that definition. Under the 21st Century Cures Act of December 2016, five kinds of software function fall outside that definition. The first four are software for the administrative support of a health care facility, software for maintaining a healthy lifestyle unrelated to any disease, software that serves as an electronic patient record, and software that transfers, stores, converts or displays laboratory or device data without interpreting it. The fifth is clinical decision support, which must meet four criteria at once.
The software must not acquire, process or analyze a medical image, a signal from an in vitro diagnostic test, or a pattern or signal from a signal acquisition system. The software must display, analyze or print medical information. It must support or provide recommendations to a health care professional. And it must let that professional independently review the basis for its recommendations, so that the professional does not rely primarily on them. IDx-DR fails the first criterion because it analyzes images, and it fails the fourth because no professional reviews the images before the result is given. It is a device on two counts.
A function that is a device is then placed in a risk class, I, II or III, by a classification regulation that names the device type. When the FDA granted IDx-DR's De Novo in 2018, it created such a regulation for retinal diagnostic software, in class II. The class sets the route to market: most class II software reaches the market through a 510(k) or a De Novo, and class III software through premarket approval.
The EU line
Under the EU's medical device regulation, software is classified by Rule 11. Software that provides information used to take decisions for diagnosis or therapy is class IIa. It is class IIb if those decisions could cause a serious deterioration of a person's health or a surgical intervention, and class III if they could cause death or an irreversible deterioration. Software that monitors physiological processes is IIa, or IIb where it monitors vital parameters whose variations could result in immediate danger. All other software is class I.
In practice, almost any software that informs a clinical decision is at least class IIa under Rule 11, and so it needs a notified body, an independent organization designated to assess conformity, before it can carry the CE mark. The Medical Device Coordination Group, a body of experts appointed by the EU countries, issues guidance on how the device regulations apply. It first endorsed its guidance on qualifying and classifying software in October 2019 and revised it in June 2025. A regulatory consultancy reviewed the revision and reported that it made no substantive change. By the consultancy's account, the revision did use the term "medical device artificial intelligence" for the first time, with a reference to the AI Act, and added examples, including software that informs a prognosis or prediction.
Every scheme in this chapter weighs what could happen to a patient if the software's information or action were wrong. A maker that classifies by how the software usually performs will usually reach a lower class than the regulator would. A clinician who is expected to catch the software's errors counts only where the scheme gives credit for that review: in the US that review can take decision support out of the device definition, and on India's grid, software that only informs a clinician sits below software that diagnoses.
India's grid
India's regulator published a draft guidance on medical device software in October 2025 and its final guidance on 21 July 2026. The final guidance no longer uses the terms software as a medical device and software in a medical device. Instead, it distinguishes standalone software from software that is part of a hardware device or drives one. Software that drives or influences hardware takes the hardware's class. Standalone software is classed on a grid that follows the 2014 international framework of Chapter 1, with India's classes A to D.
Take autonomous screening for diabetic retinopathy from retinal photographs, used by primary-care staff, as in the opening scene. In the US it is a device, because it analyzes images and no clinician reviews the basis of its answer, and its device type is class II. Under the EU's Rule 11 it provides information used for a diagnostic decision, so it is at least class IIa. Whether it is IIb depends on whether a missed referral could cause a serious deterioration of health. Rule 11 would make it class III if such a decision could cause an irreversible deterioration; because a patient whose disease is missed is told to return for screening in a year, the maker can argue against that. The maker must make each argument, and the notified body must accept it. On India's grid, software that diagnoses in a serious situation is class C. A critical situation would make it class D, and software that only informed a clinician in the same serious situation would be class A. India's guidance also states that software used by non-clinical users in a serious situation without specialist support may be treated as critical. That provision would move this function to class D if lay operators ran it at home. The one function is therefore class II in the US, IIa or IIb in the EU, and C or D in India, and in each scheme the claim decides the class, which in India the user can also change.
The rest of India's scheme follows the same logic. Where several classes could apply, the highest applies, and the regulator keeps a list of software it has classified. State authorities issue manufacturing licenses for classes A and B, and the central authority issues them for classes C and D. The guidance states that it reflects current practice under India's Medical Devices Rules of 2017 rather than creating a new control. Internationally, the medical device regulators' forum published a further document on characterizing software risk in January 2025. In it, the forum suggests assuming, where possible, that the software will fail, so that the risk assessment rests on the consequences rather than on an estimate of probability.
State in one paragraph what the software is for, who uses it, in which clinical situation, and what its output decides. Classify it under each market's scheme from that paragraph, record the reasoning, and check it against the regulators' examples. Keep the intended use under change control, because a new claim or a new user can change the class without a line of code changing.
- In January 2026 the FDA moved its boundary by guidance, for wellness estimates and for decision support that gives a single recommendation, while the law stayed unchanged.
- US law excludes five kinds of software function, and decision support is excluded only if it does not analyze images or signals and a clinician can review the basis of its recommendations.
- The EU's Rule 11 makes almost any software that informs a clinical decision class IIa or higher, and India classes standalone software from A to D on a grid.
- The same screening function is class II in the US, IIa or IIb in the EU and C or D in India, and in each the claim decides the class, which in India the user can also change.
Documents, validation and clinical evidence.
+ The questionWhat does each regulator ask to see before a software or AI device is sold, and what do the validation rules ask of the sites and makers that use software under good manufacturing practice?
Two levels of documentation
The FDA's guidance on premarket submissions for device software has been final since 14 June 2023. It replaced a guidance from 2005 and applies to every device with software, AI or not. In it, the FDA defines a documentation level: the amount of software evidence a submission must contain, which depends on what a failure could do. The level is Enhanced where a failure or latent flaw in the software "could present a hazardous situation with a probable risk of death or serious injury," judged before any risk control is counted. Otherwise the level is Basic.
Most of the file is the same at both levels: the evaluation of the level itself, a description of the software, the risk management file, the software requirements, an architecture diagram, the version history and the list of unresolved anomalies. At the Basic level, the testing evidence is a summary of testing and the protocol and report of system-level testing. The Enhanced level adds the detailed design specification, the configuration management and maintenance plans, and the protocols and reports of unit and integration testing. A maker can describe its development practices by declaring conformity to the clauses of IEC 62304 that the guidance names for its level, fewer at Basic than at Enhanced. The FDA has recognized the standard in full since 2019, but the guidance does not ask for a declaration to the whole standard, because the two schemes classify software differently (Chapter 4). Chapters 4 and 7 traced where each of these records comes from, and the guidance lists which of them the maker sends to the FDA.
A submission draws its records from the maker's quality system. The FDA's quality rule for devices is now called the Quality Management System Regulation. Since 2 February 2026 it has incorporated the international standard ISO 13485:2016 by reference, which makes the standard's requirements part of the rule, and the agency's inspectors have stopped using their old inspection technique. The surgical energy guide in this library treats IEC 62304 and the documentation levels as items on a generator maker's compliance list. This chapter has followed them from the side of the software.
What the FDA asks of AI
For AI, the FDA's main document is a draft guidance on lifecycle management and marketing submissions for AI-enabled device software functions. The FDA published it in January 2025, and as of October 2026 it was still a draft. In it, the FDA proposes the sections a submission should contain beyond the general software file: a description of the device and its user interface, labeling, risk assessment, data management, the model's description and development, validation, plans for monitoring performance in the field, cybersecurity, and a public summary.
The draft's expectations read like Parts III and IV of this guide. Test data should be independent of training data, for example drawn from different clinical sites, and sequestered from the developers. Validation data should represent the intended population, and relying on a single collection site is generally not appropriate. As an example of what may be appropriate, the draft gives at least three geographically diverse US sites. Performance should be reported across subgroups such as sex, age, race, ethnicity, disease severity, site and acquisition equipment, at each operating point. The draft suggests, but does not require, a model card: a short structured summary of the model, its data and its performance, placed in the labeling. The final guidance on change control plans, the subject of Chapter 21, completes the FDA's set of documents for AI devices.
The EU and India
In the EU, the maker puts a software device's evidence into its technical documentation, which a notified body reviews. For clinical evidence, the coordination group's guidance of March 2020 sets out three components. The first is a valid clinical association: the software's output is associated with the clinical condition it targets. The second is technical performance: the software generates its intended output accurately and reliably from its input data. The third is clinical performance: the output is clinically relevant in accordance with the intended purpose. The maker keeps all three current through post-market evaluation. The guidance predates the recent growth in AI devices and does not mention AI.
Neither IEC 62304 nor the security standard of Chapter 23 is a harmonized standard under the EU's device regulation. A harmonized standard is one whose reference the EU has published for a regulation, and a device that meets it is presumed to meet the requirements the standard covers. Neither standard appears in the EU's list of harmonized standards for the device regulation as consolidated on 17 June 2026. A maker can still use them as the state of the art, the generally accepted current practice that the regulation requires makers to take into account, but meeting them gives no presumption of conformity.
In its final guidance of July 2026, India's regulator uses the same three components. For AI, it adds requirements of its own on datasets: the composition of training, validation and test data, including demographic distribution, geographic origin and clinical diversity; whether the data are real-world, from public databases or synthetic; and a justification where models were trained or validated outside India. A maker that shows performance on non-Indian data should supplement it with validation in representative Indian populations. The guidance lists IEC 62304, IEC 81001-5-1, ISO 14971 and ISO 13485 among the standards that may apply.
Three drafts describe where regulators are heading: the FDA's AI-enabled device guidance, the FDA's draft on change control plans for all devices, and the EU's draft manufacturing annex on AI. Makers sensibly design toward them. A regulator still judges a submission against the final documents in force on the day it is filed, so the file needs to show which version of each document the maker followed.
Standards in motion
Standards change more slowly than guidance, and their status is easy to misreport. IEC 62304's first edition dates from 2006, with an amendment in 2015. A second edition, widened to health software in general, has been in preparation for years. Its status can be read directly from the international project record.
ISO and IEC track every project with a stage code. Stage 30 is the committee stage, where drafts circulate among experts; 40 is the enquiry, when a draft goes to national bodies for a vote; 50 is the approval of a final draft; and 60.60 is publication. A national standards body mirrors the international project record. In that copy, the second edition of IEC 62304 was at stage 30.20 from 7 August 2026, with the next milestone due on 27 November 2026; stage 30.20 means that a ballot on a committee draft has begun. A consultancy that follows the committee reported almost 1,500 comments on the first committee draft and expected the final draft no earlier than 2028. A vendor's web page, by contrast, stated that the final draft stage had begun on 22 May 2026 and that publication was scheduled for 12 August 2026. A standard cannot be at stage 30 and stage 50 at once, and the project record is the authority on its stage. In October 2026 the 2015 consolidated first edition is the one in force. Anything said about the second edition describes a draft, including the proposed replacement of the three safety classes by two levels of process rigor that the German association VDE reported.
Validation rules for those who use software
Sites and makers that use software, rather than sell it, follow a parallel set of rules. In the US, the FDA's rule on electronic records and signatures, 21 CFR Part 11, has applied since 1997. The FDA finalized its guidance on computer software assurance in September 2025 and revised it on 3 February 2026 under a new title that refers to quality management system software. The guidance sets out the risk-based testing of Chapter 9 for software used in production and in the quality system, and it does not cover the device software itself.
In the EU, a revised Annex 11 on computerized systems and a new Annex 22 on AI went out for consultation from July to October 2025. Both are annexes to the EU's guidelines on good manufacturing practice, or GMP, which is the set of rules on how regulated products such as medicines are made and controlled. Annex 22 is the draft described in Chapter 19. On 8 October 2026 the EU's list of GMP guidelines still contained the Annex 11 of 2011 and no Annex 22.
Determine the documentation level before design starts, and keep each required record, from requirements to unresolved anomalies, current with every release, so that preparing a submission means exporting records rather than starting a project. For AI, keep the data management records, the sequestered test sets and the subgroup results the FDA, EU and Indian documents ask for, and record which version of each guidance and standard the file follows.
- The FDA asks for Basic or Enhanced software documentation, depending on whether a software failure could present a probable risk of death or serious injury before risk controls are counted.
- The FDA's AI guidance is still a draft; it asks for test data that are independent, sequestered and representative, with results by subgroup, and it suggests a model card.
- The EU and India ask for evidence of valid clinical association, technical performance and clinical performance, and India adds validation in Indian populations for AI.
- IEC 62304's second edition is a committee draft; Part 11, the FDA's computer software assurance guidance and the EU's Annex 11 of 2011 govern software used under GMP, while a revised Annex 11 and a new Annex 22 are still drafts.
Changes, jurisdiction by jurisdiction.
+ The questionWhich software changes can a maker release on its own, and which need a regulator first?
One release, three answers
The same two-week release can be routine in one jurisdiction and need a new submission in another. Suppose a team finishes version 2.4 of an AI screening program. It retrains the model on more data from the same kinds of clinics, moves the operating threshold within the range already validated, applies a security patch to the operating system and upgrades an image-processing library to a new major version. The code was built, verified and released through the pipeline of Chapter 9, and every record is in place. Whether version 2.4 may now reach patients depends on where they are.
The three regulators ask the same underlying question, the one this guide has asked of every change: could the change move a claim the evidence supports, or create a risk the evidence did not cover? Each regulator turns that question into a different procedure. The FDA asks a series of questions that lead to a likely answer, the EU sorts changes into significant and not significant, and India divides them into major and minor. The FDA and India's regulator also offer a way to agree some changes in advance, and the EU's AI Act will offer one for medical device AI from 2 August 2028.
The FDA's questions
The FDA's guidance on deciding when to submit a 510(k) for a software change has been final since 25 October 2017. It applies to cleared devices and to those authorized through De Novo, and it works as a flowchart. A change made solely to strengthen cybersecurity, with no other effect on the device, is likely not to need a new submission, and neither is one made solely to return the device to the specification of its most recently cleared version.
A change likely does need a submission if it introduces a new risk or modifies an existing one, and that risk could result in significant harm and is not effectively mitigated. So does a change that adds or modifies a risk control for a hazard that could cause significant harm. The last question is whether the change could significantly affect clinical functionality or performance specifications; if so, a submission is likely needed. For changes to what the device is indicated for, the maker follows the FDA's general guidance on device changes.
Under the flowchart, the decision is the maker's. The maker must document its reasoning and is accountable for it in an inspection. For a model, the question on performance usually decides the outcome: retraining is meant to change performance, so a retrained model is likely to need a new submission unless an authorized change control plan already covers it. That is why the change control plans of Chapter 21 matter. UpDoc's plan, for example, lists five categories of change the maker may make without a new submission, none of which alters the cleared dosing logic.
The EU: significant change and substantial modification
Under the EU's medical device regulation, a maker informs the notified body of planned changes to a device that body certified, and changes that could affect conformity need that body's approval. The most detailed published list of what counts as a significant change for software is in the coordination group's guidance on significant changes. The group wrote that guidance for devices still certified under the older directives that the regulation replaced, but its chart for software sets out the reasoning plainly.
On that chart, a change is significant if it brings a new or modified architecture or database structure, or a change of an algorithm, or if a closed-loop algorithm replaces a user's input. So is a new operating system or any new component, or a major change to either. A new diagnostic or therapeutic feature, a new channel of interoperability, a new user interface or a new presentation of medical data are significant too. Safety-neutral bug fixes, security updates, changes to appearance and gains in efficiency are not significant, provided they do not affect diagnosis or treatment. If any one answer on the chart is significant, the whole change is significant.
The AI Act adds its own test, which turns on conformity assessment: the check that a product meets a law's requirements before it is placed on the market. Under the act, a substantial modification is a change made to an AI system after it has been placed on the market, when the change was not foreseen in the system's initial conformity assessment and affects its compliance or modifies its intended purpose. A high-risk system that undergoes such a change needs a new conformity assessment.
The act calls the maker of an AI system its provider. Changes that the provider predetermined at the initial assessment, and documented, are not substantial modifications, and these predetermined changes are the EU's form of a change control plan. For medical device AI, those provisions apply from 2 August 2028. In June 2025 the EU published joint guidance on the device rules and the AI Act. It notes that retraining may trigger reassessment under both, and that a plan agreed in advance can reduce how often that happens.
A model retrained on new data has new parameters and, by design, new performance. Every regulator in this chapter treats retraining as a change to evaluate against the device's claims, and none treats it as maintenance. A plan agreed before authorization is the main way to make retraining routine.
India: major and minor
In its final guidance on medical device software, published in July 2026, India's regulator divides software changes into two lists. Major changes need the licensing authority's approval before they are made. They include design changes that affect specifications, indications or performance; changes to software design or system requirements, new clinical claims or new types of input data; and changes of intended use. Label changes other than font, color or layout are major, and so are version changes that affect intended use, safety, effectiveness or risk controls. Minor changes are notified to the authority: bug fixes and security patches that do not affect intended use, safety or performance, version changes with no such effect, and re-tuning of performance within validated ranges. Software is identified by its version, revision level and build or release date. Under India's rules a minor change must be notified within 30 days.
For in vitro diagnostic devices, the regulator added an addendum to its questions and answers on 13 March 2026. The addendum states that each new version of approved software needs a post-approval change application, the route that the immunoassay guide in this library describes for analyzers. For AI, the July guidance adds an optional algorithm change protocol with four parts: a data management plan, a plan for evaluating and monitoring performance, a retraining plan where retraining is intended, and a plan for software updates and rollback.
Take the four changes of the opening section. In the US, the security patch to the operating system falls under the flowchart's first question and is likely not to need a submission. The threshold move within the validated range and the library upgrade need a documented assessment of whether they could significantly affect performance. The retraining is likely to need a 510(k) unless an authorized plan covers it. In the EU, the security patch is not significant, but the retraining is a change of an algorithm and the library's new major version is a major change of a component. By the reasoning of the coordination group's chart, at least two of the four changes are therefore significant, and the threshold move may be a third if it changes the algorithm's output. Under the regulation, the notified body must approve any of them that could affect conformity. In India, the patch and the threshold move are minor and are notified within 30 days. The library upgrade is minor if it affects nothing the guidance lists, and the retraining is major if it changes performance specifications. The one release thus yields one likely submission in the US, at least two significant changes in the EU and one major change in India, all before the AI Act applies to any of them.
Patches and borrowed code
The three schemes come closest on security patches and updates to third-party components. All three regulators let a maker release a patch that only closes a vulnerability without prior approval. Under the FDA's flowchart such a patch likely needs no submission, the coordination group's chart counts security updates as not significant, and India's guidance lists the patch as a minor change, which the maker notifies within 30 days. In each case the maker documents its assessment. Because no prior approval is needed, the regular and out-of-cycle fixes of Chapter 24 remain possible. The regulators differ on larger component changes: under the EU's chart a new major version of an operating system or library is significant, while the FDA and India ask what the new version could do to safety and performance. A maker that keeps its SBOM, its regression evidence and its assessment of each component change in one record can answer all three regulators from the same file.
For each change in a release, record which claims and risks it could affect, the verification run, and the regulatory decision for each market: documented assessment or submission in the US, significant or not in the EU, major or minor in India. Group changes that need a regulator into planned releases, keep security patches separate so they can ship fast, and write a change control plan for the changes a model will need.
- Under the FDA's 2017 software-change flowchart, changes made only for security or to return to specification likely need no new submission, but changes that affect risk or performance usually need a 510(k).
- The coordination group's chart for legacy devices treats changes of algorithm, architecture and major components as significant, and the AI Act adds the test of substantial modification, from which predetermined changes are excluded.
- India's regulator lists major software changes, which need approval, and minor ones, which are notified, and it offers an optional algorithm change protocol.
- One release can need one submission in the US, notified-body review in the EU and approval in India, and a plan agreed in advance is the main way to make retraining routine.
Section 524B, the AI Act and the deployer's duties.
+ The questionWhat do the cybersecurity and AI-specific rules add, what do they ask of the hospital that deploys the device, and how does one device meet two EU regulations at once?
Section 524B
On 29 December 2022 the US enacted an appropriations law, one of the laws that fund the federal government. Ninety days later, on 29 March 2023, a new section, 524B, took effect in the Food, Drug, and Cosmetic Act, the US law under which the FDA regulates devices. Section 524B applies to premarket submissions made on or after 29 March 2023 for any cyber device: one that includes software validated, installed or authorized by its maker, can connect to the internet, and has characteristics that could be vulnerable to cybersecurity threats.
A maker submitting such a device must do four things. It must provide a plan to monitor, identify and address vulnerabilities after release, including coordinated disclosure. It must design, develop and maintain processes giving reasonable assurance that the device and related systems are secure, and make updates and patches available on a regular cycle and, for critical vulnerabilities, out of cycle. It must provide a software bill of materials covering commercial, open-source and off-the-shelf components. And it must meet any further requirements the FDA sets by regulation.
The FDA gave makers a transition: until 1 October 2023 it worked with companies whose submissions lacked the new information, and after that date it could refuse to accept such submissions. The FDA revised its premarket cybersecurity guidance, the source of Chapters 22 to 24, on 3 February 2026. Two law and certification firms reviewed the revision and reported that it aligned the guidance with the new quality management system rule without adding technical expectations. In March 2026 a standards body announced that the FDA had recognized its consensus report on cybersecurity considerations unique to machine-learning devices.
One device, two EU regulations
In the EU, cybersecurity is part of the medical device regulation itself. The regulation's general safety and performance requirements state that software must be developed according to the state of the art, taking account of the development lifecycle, risk management including information security, verification and validation. They also require makers to set out the minimum requirements for hardware, IT networks and IT security measures, including protection against unauthorized access, needed to run the software as intended. Chapter 24 described the coordination group's cybersecurity guidance of 2019, revised in 2020, which states that a device should not rely on security controls in its operating environment. The EU's Cyber Resilience Act applies in full to most products with digital elements from 11 December 2027, and the duty of the makers of those products to report actively exploited vulnerabilities has applied since 11 September 2026, but the act excludes devices covered by the medical device and diagnostic regulations. For a hospital, that act therefore matters for the other software it buys, not for its medical devices.
For a device that uses AI, a second regulation applies as well. Under the AI Act, an AI system is high-risk if two conditions hold: it is a product covered by one of the EU laws listed in the act's first annex, or a safety component of such a product; and that law requires the product to be assessed by a third party. A safety component is a part that performs a safety function for a product, or whose failure endangers the health and safety of people or property. Medical devices are on that list, and almost all software that informs a clinical decision needs a notified body under Rule 11, so most medical device AI is high-risk.
The high-risk requirements cover risk management, data governance, technical documentation, automatic logging, information for those who deploy the system, human oversight, and accuracy, robustness and cybersecurity. The Digital Omnibus of July 2026 kept medical devices in that annex and moved the date from which these requirements apply to medical device AI to 2 August 2028. According to the consolidated text, the Omnibus also clarified that AI used solely for purposes other than safety does not qualify as a safety component.
The two regulations are applied in one assessment. A device that is also high-risk AI follows the conformity assessment of the medical device regulation, and within that assessment the notified body checks the AI Act's requirements. In June 2025 the device coordination group and the EU's AI board published joint guidance that describes one technical documentation, an integrated quality system and integrated risk management, rather than two parallel files.
Segmentation, firewalls and access control at the site are valuable layers of protection, and the EU's minimum IT requirements belong in the instructions for use. A device designed to be safe only behind those measures, though, depends for its security on an organization that did not design it and may change its network tomorrow.
The deployer's duties
The AI Act places duties on the organization that uses a high-risk system under its own authority, which the act calls the deployer. For an AI device used in a hospital, the deployer is the hospital. The deployer must use the system in accordance with its instructions and assign human oversight to people with the necessary competence, training and authority. Input data under its control must be relevant and sufficiently representative for the intended purpose. It must monitor the system's operation and suspend use where it sees a risk. It must inform the provider of risks without undue delay, and of serious incidents immediately.
Where the logs that the system generates automatically are under the deployer's control, it must keep them for at least six months. It must inform workers before the system is used in their workplace. The Omnibus also rewrote the act's article on AI literacy, the skills and understanding people need to use AI systems in an informed way and to be aware of their risks. Under the rewritten article, providers and deployers must take measures to support AI literacy, without having to guarantee any level of literacy in an individual.
The deployer's duties are the site practices of Chapters 18 to 20 made into law, almost line for line: integration within the specification, a local check that the inputs match the conditions of the evidence, monitoring with named people, and a channel back to the maker. For hospitals using AI that is a medical device, the deployer's duties set out above apply from 2 August 2028. The duty on AI literacy applies earlier: it sits in the act's first chapter, which has applied since 2 February 2025, and the Omnibus rewrote it without changing that date.
India
In its guidance of July 2026, India's regulator brings security into the same file as safety. Trade reports of the text state that the guidance asks makers to keep a bill of materials that includes an SBOM, covering third-party, open-source and commercial components, and to link it to the monitoring and mitigation of known vulnerabilities. By the same reports, the guidance treats cybersecurity as a risk area, asks that vulnerabilities be monitored after deployment, and asks that devices keep operating safely during incidents, including partial loss of connectivity, denial of service and loss of integrity. The guidance lists IEC 81001-5-1 among the standards that may apply, and Chapter 27 set out its rules on AI datasets.
Patient data are governed by the Digital Personal Data Protection Act of 2023 and its rules. The government formally published the rules in November 2025, and they take effect in phases. The rules on the data protection board applied at once. The registration of consent managers applies from 13 November 2026; consent managers are registered intermediaries through which people give, manage and withdraw their consent to the use of their data. Most obligations on those who process data, including security safeguards and breach notification, apply from 13 May 2027. The penalty for failing to keep reasonable security safeguards can reach 250 crore rupees, which is 2.5 billion rupees.
Counted from 8 October 2026, the next dates that apply to a software device or to the site that runs it are spread over two years. India's consent-manager rules apply in about one month, on 13 November 2026, and India's main data protection duties, including breach notification, apply in about seven months, on 13 May 2027. The Cyber Resilience Act applies in full on 11 December 2027, about 14 months away, to the hospital's software that is not a device. On 2 August 2028, about 22 months away, the AI Act's high-risk requirements for medical device AI apply, including the hospital's duties as deployer. A device authorized today and updated every quarter will go through about seven releases before the last of these dates, and each release will be assessed against the rules in force on its own release day.
For each device, list the rules that apply in each market and the date each takes effect: section 524B for cyber devices, the device regulation and the AI Act together in the EU, India's software guidance and data protection rules. Assign each duty to the maker or the site, write the site's duties into the responsibility agreement of Chapter 18, and review the list whenever a regulation changes its dates.
The four-hour operators, again
The operators of the opening scene had four hours of training and a locked program. What their trial showed in 2017 was narrow and solid: that program, on that camera, run by people like them, gave the right screening decision for adults like those enrolled often enough to clear performance thresholds agreed in advance. The trial could not show what the program would do in a clinic whose systems sent images differently, after a camera or model changed, or against someone trying to make it fail.
The same device filed in October 2026 would carry evidence for each of those gaps. Because its images and answers cross the internet, it is a cyber device, so its file would hold a threat model, an SBOM, security test results and a disclosure plan. Its software documentation would be at the level set by its potential for harm. Following the FDA's draft for AI, the maker would show test data drawn from several sites and kept from the developers, with results by subgroup and by camera, and would very likely include a change control plan for the retraining the model will need. In India the maker would be expected to add validation in Indian populations. In the EU, from August 2028, the notified body would assess the device's logging and human-oversight measures. The hospital that installed the device would test its interfaces, check its inputs against the specification, run it silently on local patients before relying on it, watch its positive and declined rates, and tell the maker when something changed.
None of that changes the fact that has shaped every chapter. Every way the software can fail was built in on the day it shipped or arrived with an update, and for the model the training data is part of what was built. The evidence that the software will be right for the next patient is a measurement of what was built, taken on patients chosen to resemble the next one. What keeps that evidence true after an update is the discipline of measuring again. Four hours were enough to take the pictures. Everything else in this guide is what it takes to keep trusting the answer.
- Since March 2023, a maker that submits a cyber device to the FDA must file a vulnerability plan, show secure processes with patching, and provide an SBOM.
- In the EU, a medical device that is high-risk AI goes through one assessment under both the device regulation and the AI Act, and the AI duties apply from 2 August 2028.
- Hospitals that deploy such AI become deployers under the AI Act, with duties for oversight, input data, monitoring and logs that mirror good site practice.
- Under India's 2026 guidance, SBOMs and cyber resilience belong in the device file, and India's data protection duties phase in through May 2027.
Lessons.
Fourteen working rules follow from the chapters, for anyone who builds, buys, installs, regulates or writes about software and AI in medical devices. Each names the chapters that explain why it holds.
- Write the intended use before anything else. The claim made for the software decides whether it is a device, sets its class in each market and fixes what the evidence must show. A new claim or a new user can change all three without a line of code changing (Chapters 1 and 26).
- Look for the fault and the input that reaches it, not a wear-out rate. Software does not age. Every fault was either built in on the day the software shipped or arrived with an update, and a fault causes a failure only when some input reaches it (Chapter 2).
- Build confidence from the process, because testing alone cannot supply it. A program has more paths than any test campaign can exercise, and very low failure rates cannot be demonstrated by running the software. The evidence is therefore requirements derived from risk and traced to tests, an architecture that keeps the dangerous part small, and rigor scaled to harm (Chapters 3 to 5 and 7).
- Account for every component someone else wrote. Record each one's exact version, published anomalies, support status and replacement plan, generate the SBOM from the build, and treat pretrained models and public datasets as components too (Chapter 6).
- Let every sprint and every pipeline run leave its records. IEC 62304 prescribes activities and records, not phases, so agile increments and automated builds can produce its evidence, provided the team's definition of done includes verification and a person still decides each release (Chapters 8 and 9).
- Keep the data that judges a model away from the people and data that built it. Separate test sets by patient and, where possible, by site, sequester them from developers, and watch for the leaks through which a model learns cues of the hospital instead of signs of the disease (Chapters 10 and 11).
- Ask who decided the right answer and where a positive result begins. The reference standard limits what measured accuracy can mean. The operating point and the endpoints must be fixed before the test and the results reported with their intervals; a claim must also state how it counted declined cases and give the results under the other ways of counting them (Chapters 12, 13 and 17).
- Test where the device will be used, and report by subgroup. Accuracy is a measurement on particular patients, equipment and settings. It changes when any of them changes, and an average can hide a group the model fails (Chapters 14 and 15).
- Measure the reader and the model together when a clinician uses the output. An assistant that is usually right can still make a reader worse. The evidence for an assistive device is therefore the performance of the reader and the model together, measured with attention to automation bias (Chapter 16).
- Treat installation as part of the evidence. Test every interface with local data, confirm that local inputs fall inside the specification, run the device silently on local patients before anyone relies on it, and write down who notifies whom of changes on either side (Chapters 18 and 19).
- Watch inputs and output rates from the first day, and check accuracy on a schedule. A shift in what the device receives and returns can show within days. A change in accuracy is confirmed only by reference readings of sampled negatives and by matched outcomes, and named people must act on agreed limits (Chapter 20).
- Agree in advance what a model may change, and fix the tests each version must pass. A change control plan agreed in advance is the main way to make retraining routine. The FDA authorizes such plans, India offers an optional algorithm change protocol, and from August 2028 the EU's AI Act excludes predetermined changes from substantial modification. Outside such a plan, every regulator treats retraining as a design change, not as maintenance (Chapters 21 and 28).
- Design the device on the assumption that it will be attacked and will need patches. Authenticate every interface, sign the code and keep a fallback image, build the threat model from the architecture, search the SBOM whenever a vulnerability is published, and publish a way to report vulnerabilities (Chapters 22 to 24).
- Read the record, and date every status. Makers' counts and claims to be first remain claims until a regulator's database supports them. Every rule, guidance and standard in this field should be quoted with the date of its status, because the clauses change faster than the mechanism beneath them (Chapters 25 to 29).
Glossary.
Terms are defined as they are used in this guide.
- 510(k)
- A US premarket submission by which the FDA clears a device shown to be substantially equivalent to one already on the market. Once a De Novo has created a device type, later devices of that type can follow this route; IDx-DR's new versions were cleared this way in June 2021 and June 2022.
- AAMI TIR45
- A technical information report from AAMI, the Association for the Advancement of Medical Instrumentation, that maps agile development, in short cycles each ending with tested software, onto IEC 62304 in four layers: project, release, increment (a sprint of one to four weeks) and story (a few days of user-visible function). The FDA recognizes its 2012 and 2023 editions in full.
- AI Act
- The EU's 2024 law on artificial intelligence, under which most medical device AI is high-risk, because medical devices are on its product list and almost all decision-informing software needs a notified body. Its high-risk requirements, from risk management and data governance to logging and human oversight, apply to medical device AI from 2 August 2028, as moved by the Digital Omnibus of July 2026.
- AI model
- In this guide, an algorithm whose decision rules were fitted to example data rather than written line by line. Its accuracy is a measurement on a test set and holds only for patients like those in it, against the reference chosen, for the exact model that was locked.
- Algorithm
- The procedure the software follows. When its decision rules are fitted to example data rather than written line by line, the guide calls it an AI model.
- Algorithm change protocol
- In the FDA's 2019 discussion paper, the methods a maker would use to make, test and control changes to a machine-learning model within an agreed region of potential changes. India's July 2026 guidance offers an optional AI protocol of the same name in four parts: data management, performance evaluation and monitoring, retraining, and updates with rollback.
- All-pairs testing
- Testing every pair of parameter settings rather than every combination, which grows slowly with the number of settings. In 98 percent of 109 FDA recall reports that NIST researchers could trace, it would have revealed the failure; the HAMILTON-C6 ventilator fault needed four conditions at once, so a pairwise campaign would not have been sure to find it.
- Attack surface
- Every way data or commands can enter a device: radio and Bluetooth links, Wi-Fi and Ethernet, USB and serial ports, a cloud connection, a service interface used only by technicians. The MiniMed pumps' unauthenticated radio link was the part of theirs that mattered.
- AUC
- Area under the curve: the area under a receiver operating characteristic curve, from 0.5 for a coin toss to 1 for perfect separation, which summarizes how well a score separates diseased from healthy cases regardless of where the threshold is set. It hides the operating point and can stay stable while calibration drifts; the Google retinal network of 2016 scored 0.991.
- Automation bias
- The tendency to accept a machine's output in place of one's own judgment, which studies suggest is strongest when the person is uncertain. In a 2023 study, wrong suggestions labeled as AI cut radiologists' accuracy on mammograms from about 80 percent to between 20 and 46 percent, with the least experienced readers affected most.
- Autonomous, assistive and triage devices
- The three roles of software that analyzes clinical data. An autonomous device gives the answer itself, as IDx-DR does; an assistive device gives a reader extra information and the reader decides, so the reader's performance with and without it is what counts; a triage device works in parallel with the usual workflow, alerting a specialist while radiologists still read every image.
- Baseline
- The results of a site's local performance check, kept as the reference for later monitoring: the distribution of inputs, the rate of declined cases, and the local sensitivity, specificity and predictive values. In agile work the word also names the controlled set of requirements approved at each release.
- Calibration
- Whether a model's scores mean what they say, such as predicted risks matching the true rate; discrimination is whether it ranks sick patients above healthy ones. Seven kidney-injury models at US Veterans Affairs hospitals kept their discrimination for nine years while their calibration drifted, so a fixed threshold would have alerted ever more often.
- CDSCO
- Central Drugs Standard Control Organisation: India's device regulator. Its final guidance on medical device software, of 21 July 2026, classes standalone software A to D, divides software changes into major and minor, and asks makers to justify AI models trained or validated outside India; licenses for classes A and B come from state authorities, for C and D from the central authority.
- CE mark
- The marking under which a device is sold in the EU. Under Rule 11 almost any software that informs a clinical decision is at least class IIa, so it needs a notified body's assessment before it can carry the mark.
- Change control
- The discipline that judges every change to a device against the claims it could move and the risks it could create before it reaches patients; for an AI model it adds a plan agreed with the regulator in advance that bounds what may change and fixes the tests each new version must pass. It is the guide's answer to what keeps the evidence true after an update.
- Clinical decision support
- Software that advises clinicians, which US law excludes from the device definition only if it meets four criteria at once: it does not analyze a medical image or a diagnostic signal or pattern; it displays, analyzes or prints medical information; it supports recommendations to a health care professional; and that professional can independently review their basis. IDx-DR fails the first and the fourth.
- Computerized system validation (CSV)
- Validation by its user of a computerized system that affects product quality, patient safety or data integrity, which GxP rules require of laboratories, blood services, clinical-trial organizations and drug manufacturers. Most follow GAMP 5, scaling the effort to the system's impact, complexity and novelty and building on the supplier's evidence.
- Confidence interval
- The range that expresses the uncertainty of an estimate made from a sample, at a stated confidence, usually 95 percent; it narrows as the number of cases behind the estimate grows. IDx-DR's observed sensitivity of 87.4 percent has an exact interval of 81.9 to 91.7 percent, and a sensitivity of 90 percent spans 78.6 to 95.7 from 50 patients but 87.1 to 92.3 from 500.
- Configuration management
- Identifying and controlling every item that goes into a build, from source code and SOUP versions to build scripts and the compiler with its settings, so that the same inputs produce the same software and the version tested is the version shipped. IEC 62304 requires it at every class.
- Coordinated vulnerability disclosure
- The practice by which someone who finds a weakness reports it privately to the maker, the maker investigates, and the two agree when the details are published, ideally with a fix; a government coordinator such as the US cybersecurity agency CISA often stands between them. Since 29 March 2023 US law has required it in the vulnerability plan of every premarket submission for a cyber device.
- Coverage
- A measure of how much of the code a test campaign executed. The FDA's 2002 guidance describes a ladder from statement coverage, every line run once, which it calls insufficient on its own, through decision, condition and multiple-condition coverage, to full path coverage, which it calls generally not achievable; coverage shows what ran, not whether it was right.
- CRC
- Cyclic redundancy check: a checksum computed from a file by a published method, which detects corruption by a faulty memory chip or a noisy cable. Anyone who changes the file can compute a matching one, so the FDA tells makers not to rely on it as a security control; only a signature made with the maker's private key proves a file is genuine.
- CSA
- Computer software assurance: the FDA's risk-based approach to validating software used in production and in the quality system, finalized in September 2025 and reissued in February 2026. For functions whose failure could foreseeably compromise safety it suggests more rigorous assurance, such as documented, scripted testing, and for the rest lighter, unscripted methods; it does not cover the device software itself.
- CVSS
- Common Vulnerability Scoring System: the usual measure of a vulnerability's technical severity, built from factors such as the attack vector, its complexity and the impact on confidentiality, integrity and availability. The 2019 MiniMed weakness scored 7.1, High, on version 3; the score says nothing about how many patients use a device or what harm follows.
- Cyber device
- Under section 524B of the US Food, Drug, and Cosmetic Act, a device that includes software validated, installed or authorized by its maker, can connect to the internet, and has characteristics that could be vulnerable to cybersecurity threats. The FDA reads connection broadly, down to a USB port used briefly for service; IDx-DR, whose images and answers cross the internet, would be one.
- Cyber Resilience Act
- The EU law, CRA for short, on the cybersecurity of products with digital elements, which applies in full from 11 December 2027; its makers' duty to report actively exploited vulnerabilities has applied since 11 September 2026. It excludes devices covered by the medical device and diagnostic regulations, so for a hospital it governs other software it buys, not its medical devices.
- Dataset shift
- Any difference between the distribution of cases a model was fitted and tested on, meaning patients, images and labels in particular proportions, and the one it meets in use. It can change the inputs (a new camera or population), the prevalence, the right answer itself, called concept shift (a new grading guideline), or the acquisition chain (a scanner update, compression, a new format), and time produces it too.
- De Novo
- The US route by which the FDA grants authorization to a new kind of device and creates a classification regulation for its type, so that later devices of the type can be cleared through a 510(k). A De Novo is granted, not cleared: IDx-DR's, on 11 April 2018, created the class II type of retinal diagnostic software device.
- Decision summary
- The FDA's published account of the review behind a De Novo, with a shorter summary for most 510(k) clearances, covering the indications and limitations, the software, the special controls, the clinical study and the risks with their mitigations. IDx-DR's is the most complete public record of its evidence, though its date and one interval need checking against the counts.
- Declined case
- A case for which a device gives no usable answer, such as IDx-DR's result of insufficient image quality, which its labeling directs be referred. Headline figures usually leave such cases out: counting IDx-DR's 10 declined diseased participants as misses lowers its sensitivity from 87.4 to 83.2 percent, and counting them as referrals raises it to 88.0.
- Definition of done
- The conditions a story must meet before it counts as finished in agile work: requirement written and reviewed, design and code reviewed, unit and integration tests passed, trace links in place, risk analysis updated and any SOUP change recorded. Applied strictly, it lets the evidence for a release accumulate in small pieces as the work is done.
- Deployer
- Under the EU's AI Act, the organization that uses a high-risk AI system under its own authority; for a hospital AI device, the hospital. From 2 August 2028 it must follow the instructions, assign competent human oversight, keep its input data relevant and representative, monitor the system, report risks and serious incidents to the provider, and keep the logs for at least six months.
- Device software function
- Software that meets the legal definition of a medical device, whether it runs inside a pump or on a server by itself. What makes it one is its intended use, not anything in its code.
- DICOM
- Digital Imaging and Communications in Medicine: the standard format hospitals use for medical images and image-based results. Each image carries a header of tagged fields, such as the modality, the manufacturer, the model name and the equipment's software versions, which a device can read to refuse inputs outside its specification; IDx-DR accepted DICOM images from its 2021 version.
- Documentation level
- The FDA's measure, since June 2023, of how much software evidence a submission must contain: Enhanced where a failure could present a hazardous situation with a probable risk of death or serious injury, judged before any risk control is counted, and Basic otherwise. Enhanced adds the detailed design and the unit and integration test protocols and reports.
- End of support
- The date after which a software supplier no longer publishes security fixes. Extended support for Windows XP ended on 8 April 2014; devices that outlive their operating system's support stay exposed, as WannaCry showed in 2017 on unpatched or unsupported Windows systems; the FDA asks makers to state each component's support level and end-of-support date.
- Enrichment
- Recruiting more of one kind of participant or case than the population holds, such as the extra participants with poorly controlled diabetes in the IDx-DR trial, whose paper also gave figures corrected for it. In reader studies, enrichment with diseased cases can change how readers behave and biases predictive values.
- Evidence
- In this guide, records that someone outside the team could check: requirements, test results, study data, monitoring reports.
- Exploitability
- How easily an adversary could use a vulnerability. Because an attacker chooses the inputs, the FDA's cybersecurity guidance judges security risk by exploitability and by the harm that would follow, not by probability, and treats known vulnerabilities as reasonably foreseeable.
- External and prospective tests
- An external test uses data from sites, devices or periods that contributed nothing to a model's development and shows how it travels; a prospective test runs the device forward in its intended setting, on patients enrolled for the purpose, with its real workflow and operators. The IDx-DR trial was prospective, in primary care.
- Fault
- A flaw in a program's logic or data, present from the moment it was written, that causes a failure when an input reaches it, in every copy of that version. The HAMILTON-C6 ventilator fault waited inside three software versions for four ordinary events to coincide.
- FHIR
- Fast Healthcare Interoperability Resources: HL7's newer standard, built from web resources that can be addressed individually. Release 5 has been current since March 2023, and release 6 was in its second normative ballot in July 2026.
- Foundation model
- A class of large model, ranging in the FDA's words from large language models to multimodal architectures; in September 2026 the agency said it would explore tagging devices that incorporate one, and none was yet tagged. Aidoc's January 2026 clearance of 14 acute CT findings from one model rests on what the company calls a foundation model, a large image model.
- GAMP 5
- Good Automated Manufacturing Practice: the guide from the industry body ISPE that regulated companies follow to validate the computerized systems they buy and build, in its second edition of July 2022. It is risk-based, sorts software into categories from infrastructure to custom software, supports iterative and agile methods, and was joined in July 2025 by a separate GAMP guide to AI.
- Generative AI
- AI models that produce text, images or speech rather than a score. Where such a model sits relative to the clinical decision decides the evidence a device needs: one that only collects a patient's answers for fixed, cleared rules is an interface, while one that writes the recommendation is the decision.
- Good machine-learning practice
- Principles for developing AI devices, published by the FDA, Health Canada and the UK's MHRA in 2021 and finalized by the International Medical Device Regulators Forum in January 2025. Among them: test sets independent of training data, reference standards fit for purpose, data representative of the intended population, assessment of the human-AI team, and monitoring of deployed models.
- GxP
- Good practice: the collective name for the good-practice rules, such as good manufacturing practice, under which laboratories, blood services, clinical-trial organizations and drug manufacturers work. They require any computerized system that affects product quality, patient safety or data integrity to be validated by its user.
- Hazard, hazardous situation and harm
- The chain that ISO 14971 traces: a hazard is a potential source of harm, a hazardous situation is one in which a sequence of events exposes someone to it, and harm is what may follow. Each chain is judged by the severity of the harm and the probability of the sequence; for software, regulators suggest assuming the failure will happen.
- HL7
- Health Level Seven: a standards body, and its version 2 standard, which carries orders and reports between hospital systems as messages; the body says version 2 is used by 95 percent of US healthcare organizations.
- IEC 62304
- The international standard for the life cycle of medical device software, published on 9 May 2006, amended in June 2015 and recognized in full by the FDA since January 2019. It prescribes activities and their records, not their order, and sets three software safety classes; a second edition widened to health software was a committee draft in October 2026, not expected before 2028.
- IEC 80001-1
- The standard, in its 2021 edition, for organizations applying risk management before, during and after connecting a medical device or health software to their IT infrastructure, covering safety, effectiveness and security. It is marked for revision and is to be replaced by a standard in the 81001 series.
- IEC 81001-5-1
- The standard, published in December 2021, for building security into health software. It follows IEC 62304's scope and clause order and takes its security activities from IEC 62443-4-1, adding security risk management, threat modeling, security testing and vulnerability handling to an existing process; the FDA names it as one way to meet its secure product development framework.
- IHE
- Integrating the Healthcare Enterprise: an initiative that writes profiles telling vendors how to combine standards for a task. Its profiles for AI results and AI workflow, which define how results are encoded in DICOM and how analysis requests are sent and managed, were still at trial implementation in October 2026.
- Intended use
- What the maker says software is for, in its labeling and claims. It, not anything in the code, makes software a medical device, sets its class in each market and fixes what the evidence must show; IDx-DR's covers adults with diabetes not diagnosed with retinopathy, imaged on the Topcon NW400.
- ISO 13485
- The international standard for the quality management systems of device makers, in its 2016 edition. The FDA's device quality rule has incorporated it by reference since 2 February 2026, and it requires makers to validate software used in the quality system and in production, in proportion to risk.
- ISO 14971
- The international standard for the risk management of medical devices, whose 2019 edition was confirmed in 2025. It traces hazards through hazardous situations to harms and gives each unacceptable risk a control.
- Large language model
- A generative AI model that produces text, abbreviated LLM by the FDA. UpDoc's maker announced in June 2026 that its insulin program, cleared in December 2025, was the first FDA clearance of software using patient-facing large language models, a claim the FDA has not confirmed; its public summary points to an interface rather than the decision.
- Leakage
- Any route by which the data that judges a model influences how it is built, inflating the reported score: the same patient on both sides of a split, preprocessing statistics computed with the test images included, duplicates in both sets, or repeated looks at the test set until it becomes a second tuning set.
- Legacy software
- Software legally on the market but without enough evidence that it was built to IEC 62304. The standard's 2015 amendment lets it stay in use after a risk analysis using field experience, a gap analysis against the standard, a plan to close the gaps that matter and a documented rationale.
- Local performance check
- Running a device on a sample of a site's own cases, silently or retrospectively, and comparing its outputs with a local reference, after acceptance has confirmed the installation, interfaces and fallback. Showing a sensitivity above 75 percent when about 87 percent is expected takes about 69 diseased cases, or 690 consecutive cases where 10 percent have the disease.
- Locked model
- A model whose parameters are fixed after training, so that the same input always gives the same output and it can be verified like any other software; a continuously learning model, by contrast, keeps adjusting them from new data in use. IDx-DR's analysis was locked before the 2017 trial, and in 2019 the FDA wrote that the AI devices it had authorized were typically locked.
- Machine-learning model
- A fixed structure chosen by engineers whose parameters are set by training on data rather than by hand. The structure is code that can be reviewed; the parameters cannot be read as rules, so evidence about the model comes from its behavior on carefully chosen data.
- MDR
- Medical Device Regulation: the EU's law for medical devices, which classifies software by Rule 11, so that almost any software informing a clinical decision needs a notified body, and makes makers state the minimum hardware, network and IT security their software needs. Its counterpart for diagnostic tests is the in vitro diagnostic regulation, IVDR.
- Medical Device Coordination Group (MDCG)
- The EU body whose guidance interprets the device regulations: on qualifying and classifying software (2019, revised June 2025), clinical evidence for software (March 2020), significant changes, cybersecurity (2019, revised 2020) and, jointly with the EU's AI board in June 2025, meeting the device rules and the AI Act in one assessment.
- Model card
- A short structured summary of a model, its data and its performance, placed in the labeling; the FDA's draft guidance on AI-enabled devices suggests one without requiring it.
- Modification protocol
- The part of a predetermined change control plan that says how each planned change will be developed, validated and implemented: how new data are collected and kept separate, how the model is retrained and evaluated, the acceptance criteria it must meet and how users are told. Its acceptance criteria, fixed before any change is made, carry the plan's weight.
- More than mild diabetic retinopathy
- The condition IDx-DR screens for: in the 2017 trial, a worse eye at level 35 or higher on the severity scale of the Early Treatment Diabetic Retinopathy Study (ETDRS), or macular edema, as graded at the Wisconsin reading center. It was present in 24 percent of the analyzable participants.
- Notified body
- An independent organization designated in the EU to assess a device's conformity before it can carry the CE mark. Under Rule 11 almost all software that informs a clinical decision needs one, and for a device that is also high-risk AI, the notified body checks the AI Act's requirements within the same assessment.
- OTS
- Off-the-shelf software: in the FDA's term, a generally available software component used by a device maker that cannot claim complete control of its life cycle. A commercial operating system is OTS, and so is a compiler or test tool that never ships in the device but must be validated for its use.
- Overfitting
- Fitting a model so closely to its training examples that it follows their noise and predicts new data worse. In the guide's example, a ten-parameter polynomial passes through all ten training points yet misses new points by 1.67 on average, where a straight line misses them by 0.78.
- PACS
- Picture archiving and communication system: the hospital system that stores images after the scanner sends them, from which an imaging model receives studies and to which it often returns its results.
- Parameter
- One of the adjustable numbers inside a model, a weight that multiplies one value before it is passed on; ResNet-50, a widely used image network of some fifty layers, has 25,557,032. Parameters are not written but found by training.
- Part 11
- 21 CFR Part 11, the FDA's rule on electronic records and signatures, in force since 1997, which requires records that are attributable, complete and protected, with audit trails.
- Pipeline
- An automated sequence that takes every submitted change, builds the software from source, runs the tests and analysis tools and packages the result, with no step done by hand. For a device maker it makes every internal build a complete record; a person still judges the unresolved anomalies and approves the release.
- Predetermined change control plan (PCCP)
- A plan, authorized with a device, that lets specified changes be made without a new submission. The FDA's final guidance for AI-enabled software, of December 2024, asks for a description of the planned modifications, a modification protocol and an impact assessment of their benefits and risks, all within the intended use.
- Predictive value
- What a device's answer means in a given population: the positive predictive value (PPV) is the share of positive answers that are right, the negative predictive value (NPV) the share of negative answers that are right. Both change with prevalence: IDx-DR's positive answers are right about 72 percent of the time at 24 percent prevalence and about 38 percent at 7 percent.
- Premarket approval (PMA)
- The US route to market for class III devices. A predetermined change control plan can be authorized through it, as through a 510(k) or a De Novo.
- Prevalence
- How common a condition is in the population tested: about 24 percent for more than mild retinopathy among the IDx-DR trial's analyzed participants, 7 percent for sepsis in the Michigan hospitalizations. It leaves sensitivity and specificity unchanged but changes predictive values, and a change in it is one kind of dataset shift.
- Quality Management System Regulation (QMSR)
- The FDA's quality rule for devices under its name since 2 February 2026, when it began to incorporate ISO 13485:2016 by reference and the agency's inspectors stopped using their old inspection technique.
- Random and systematic failure
- A random failure is one of a hardware part, which wears and fails at its own moment, so that a rate per hour describes a population; a systematic failure occurs every time the same conditions recur, in every copy of the same version, which is how software fails. Two copies of a program fed the same input fail together, so redundancy does not protect against a software fault.
- Reader study
- A study of clinicians reading cases with and without an assistive device. In the fully crossed design of the FDA's guidance, every reader reads every case in both conditions, sessions are separated by at least four weeks to avoid memory bias, and the primary measure is usually the area under the ROC curve with the aid against without it.
- Reference standard
- The best available judgment of each case's true state, against which a device is scored, often called ground truth. For IDx-DR it was the majority grade of three masked readers at the Wisconsin reading center, from four stereo photograph pairs of each retina, with macular scans added only in a secondary analysis; an imperfect reference lowers a perfect model's apparent accuracy and can hide errors both share.
- Regression testing
- Re-running the tests that cover what a change touched, and for anything beyond a trivial change the whole system-level suite, so that behavior verified before the change is verified again after it. Because 79 percent of the software recalls the FDA counted in 1992 to 1998 came from defects introduced by changes after release, it is the most important habit in maintenance.
- Release
- In agile work, a period of one to several months ending with software that could be delivered. A release to patients is a regulated event: IEC 62304 requires it to be verified as complete, its known anomalies evaluated, the released version archived and reproducible, and the release approved.
- Reproducible build
- A build in which rebuilding a past release from its recorded inputs gives a bit-for-bit identical result, so that the version tested and the version shipped can be shown to be the same.
- Responsibility agreement
- A written statement of which party, the hospital, the IT supplier or the device maker, does what across the life of a device's connection to a hospital's systems: who tests the interfaces, who approves changes on each side, who is told when a scanner or the archive is updated, and what clinicians do when the AI is unavailable.
- ROC curve
- Receiver operating characteristic curve: a plot of sensitivity against the false-positive rate at every possible threshold of a model's score, showing the trade between missed cases and false alarms. The Google retinal network's curve carried a high-specificity point, 90.3 percent sensitivity at 98.1 percent specificity, and a high-sensitivity point, 97.5 percent at 93.4 percent.
- Root of trust
- The protected start-up code and key, held in memory that cannot be rewritten, at the start of a secure boot chain. It must be in the hardware from the first unit shipped, because a device built without it cannot gain one by update.
- Rule 11
- The EU device regulation's rule for classifying software: IIa if it informs diagnostic or therapeutic decisions, IIb if those decisions could cause a serious deterioration of health or a surgical intervention, III if they could cause death or an irreversible deterioration. Monitoring software is IIa, or IIb for vital parameters whose variations could mean immediate danger; all other software is class I.
- Rule of three
- When no failures are seen in a number of independent, realistic trials, the upper bound of the 95 percent confidence interval for the failure probability is about three divided by that number. Showing a rate of 1 failure in a billion hours this way would take about three billion failure-free hours, some 342,000 years.
- SBOM
- Software bill of materials: a machine-readable list of every component in a piece of software, with supplier, name, version and dependencies, now extended to models and datasets. Required for US cyber devices since March 2023, it tells a maker or a hospital which devices contain a newly vulnerable component within hours rather than weeks.
- Section 524B
- The section of the US Food, Drug, and Cosmetic Act, in effect since 29 March 2023, that requires the maker of a cyber device to file a plan for monitoring and addressing vulnerabilities after release, including coordinated disclosure; to keep processes that make the device secure, with patches on a regular cycle and out of cycle for critical vulnerabilities; and to provide an SBOM.
- Secure boot
- A start-up chain in which each stage checks the next before handing over control. At every power-on, protected code recomputes the firmware's hash, a fingerprint that changes completely if a single bit changes, checks it against the maker's signature made with a private key, and runs the firmware only if the two agree, so altered code never runs.
- Secure product development framework
- The FDA's name for a development process with security built in, from threat modeling and security requirements to security testing and the handling of vulnerabilities after release; its 2026 guidance names IEC 81001-5-1 and IEC 62443-4-1 as examples that can satisfy it.
- Segregation
- In IEC 62304, any mechanism that prevents one software item from negatively affecting another: a separate processor, operating-system memory protection, or checks on data crossing a boundary. Demonstrated segregation lets the strictest class apply only to the dangerous code, in the guide's pump example 8,000 of 200,000 lines.
- Sensitivity
- The share of people with the condition whom a device flags: 173 of 198, 87.4 percent, in the IDx-DR trial. Its precision depends on the number of diseased patients tested, not on the total.
- Significant change
- Under the EU's device rules, a change that brings in the notified body. The coordination group's software chart counts a new or modified architecture, database structure or algorithm, a new operating system or component or a major change to one, and a new feature, interoperability channel or user interface as significant, but not safety-neutral bug fixes, security updates or cosmetic changes.
- Silent trial
- Running a model on live clinical data with its outputs hidden from clinicians, so that its local performance is measured before anyone acts on it. At the Hospital for Sick Children in Toronto one took a kidney-ultrasound model from an area under the curve of 0.90 to 0.50; reprocessing the live images, which arrived as unprocessed PNG files rather than processed JPEGs, restored most of the loss.
- Software as a medical device
- Software that is a medical device on its own, running on ordinary computers, phones or cloud servers rather than inside hardware; IDx-DR's analysis program is an example. India's 2026 guidance drops the term in favor of standalone software.
- Software requirement
- A statement of what the software must do, or must never do, written so that a test, an inspection or an analysis can show whether it holds. A display that is easy to read is not one; a dose shown with its unit in characters at least 5 mm high until the user confirms it is.
- Software safety class
- IEC 62304's class A, B or C, set by the worst harm software could contribute to once risk controls outside it are counted: A where no unacceptable risk remains, B for non-serious injury, C for death or serious injury. Until classified, software is treated as class C; the class sets how much process the standard demands, and only controls outside the software can lower it.
- Software system, item and unit
- IEC 62304's division of a program: the software system is divided into software items, any identifiable parts, and these in turn until they reach software units, items not divided further. IDx-DR has three separately versioned items: the client, the service that passes images and results, and the analysis.
- SOUP
- Software of unknown provenance: in IEC 62304, a software item already developed and generally available that was not developed for the device, or one developed earlier without adequate records. For each, a maker states the requirements the device places on it and its exact version, evaluates its published anomalies and keeps it under configuration control.
- Special controls
- The requirements the FDA sets for a device type when it creates one, which every later device of that type must meet. IDx-DR's included a protocol on which changes could significantly affect safety or effectiveness and a cybersecurity vulnerability and management process.
- Specificity
- The share of people without the condition whom a device clears: 556 of 621, 89.5 percent, in the IDx-DR trial, and about 79 percent in studies of young people outside its label.
- Subgroup analysis
- Reporting a device's performance separately for groups within the intended population, such as by sex, age, ethnicity, skin tone, site or equipment, each with its confidence interval, with the groups named in the protocol before the data are seen. Dermatology models that had scored 0.88 to 0.94 on their own test sets fell as low as chance on the darkest skin.
- Substantial modification
- Under the EU's AI Act, a change to an AI system after it is placed on the market that was not foreseen in its initial conformity assessment and affects its compliance or modifies its intended purpose; a high-risk system that undergoes one needs a new conformity assessment. Changes predetermined and documented at the initial assessment are not substantial modifications.
- Threat model
- A list of a device's entry points, the assets behind them, such as dose settings, patient data and the software itself, and the adversaries who might want them, with what an attacker could do on each path and what would stop it. The FDA expects it to be part of the design, updated as the architecture changes.
- Threshold
- The value that turns a model's score into a decision, refer above it and not below. Moving it trades missed cases against false alarms; the point chosen, the operating point, is fixed from the clinical costs before the test set is opened.
- Traceability
- The links from each hazard to the requirements that control it, from each requirement to the design that implements it and the tests that verify it, and from each test to its result, so that a reviewer can follow any risk to its control and its evidence.
- Training
- Finding a model's parameters from examples whose right answers are known: the model scores a batch, a measure of its error called the loss is computed, and every parameter is nudged in the direction that would have made the loss smaller, over millions of batches.
- Training, tuning and test sets
- The three separate bodies of data behind a model: the training set its parameters are fitted to, the tuning set used for choices training does not make (which engineers usually call the validation set), and the test set, held back and used once to estimate performance on the next patient. Only the test set, kept independent and split by patient, judges the model fairly.
- Unit, integration and system testing
- The three levels of verification by test: unit testing checks each smallest piece in isolation against its detailed design, integration testing checks that units and items work together across their boundaries, and system testing checks the complete software, usually on the real hardware, against the software requirements.
- Unresolved anomaly
- A known defect that software ships with. Submissions list each one with an assessment of its effect on safety and effectiveness, and a release may carry it only if a person has judged that it leaves no unacceptable risk.
- Valid clinical association
- The first of three components of clinical evidence for software in the EU and India: the software's output is associated with the clinical condition it targets. The others are technical performance, producing the intended output accurately and reliably from the input, and clinical performance, an output clinically relevant to the intended purpose.
- Validation
- Confirmation, by examination and objective evidence, that software specifications conform to user needs and intended uses; it asks whether the specification was right and reaches to usability and clinical evidence. In machine learning the same word names the tuning set, and a submission that uses both senses must say which it means.
- Verification
- Objective evidence that the outputs of a stage of development meet the requirements set for that stage, that is, that the software was built as specified. It runs at unit, integration and system level, with reviews, code inspection and static analysis finding faults without running the code.
- VEX
- Vulnerability Exploitability eXchange: a statement of whether a published vulnerability affects a specific product, with one of four values, not affected, affected, fixed or under investigation. An SBOM says which components a device contains; a VEX says whether a weakness in one of them can be exploited there.
- Watchdog
- A timer, often a separate circuit, that the software must reset at regular intervals; if the program hangs, the timer runs out and forces the device into a safe state. A program that runs on time while computing a wrong dose keeps resetting it, so a watchdog cannot catch a wrong output.
Sources.
Each source is listed under the chapter whose text first relies on it, with the opening scene and the front matter first. Company documents, papers by a company's founders or staff, and secondary reports are marked as such. Regulatory status, standards editions and product facts are as of October 2026.
- Opening — Abràmoff M.D., Lavin P.T., Birch M., Shah N., Folk J.C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. npj Digit Med 1, 39 (2018). doi:10.1038/s41746-018-0040-6 (first author founded the company; trial funded by IDx)
- Opening — US Food and Drug Administration. De Novo DEN180001, IDx-DR (IDx, LLC): database record and decision summary; received 12 Jan 2018, granted 11 Apr 2018. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN180001; accessdata.fda.gov/cdrh_docs/reviews/DEN180001.pdf
- Opening — Digital Diagnostics. IDx rebrands to lead AI health care revolution (5 Oct 2020). digitaldiagnostics.com/idx-rebrands-to-lead-ai-health-care-revolution/ (manufacturer)
- Opening — US Food and Drug Administration. FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. Press release, 11 Apr 2018. drugdiscoverytrends.com/fda-permits-marketing-of-ai-based-device-to-detect-certain-diabetes-related-eye-problems/ (read via the drugdiscoverytrends.com reprint; the fda.gov page returned 404)
- Opening — US Food and Drug Administration. General Principles of Software Validation; Final Guidance for Industry and FDA Staff. 11 Jan 2002 (Section 6 since superseded by the computer software assurance guidance). fda.gov/media/73141/download
- Ch. 1 — US Code of Federal Regulations (eCFR). 21 CFR 886.1100, Retinal diagnostic software device (codified at 87 FR 3205, 21 Jan 2022). ecfr.gov/current/title-21/chapter-I/subchapter-H/part-886/subpart-B/section-886.1100
- Ch. 1 — Central Drugs Standard Control Organisation, India. Guidance document on Medical Device Software under MDR-2017. Doc No. CDSCO/MD/GD/MDSW/01/2026. 21 Jul 2026. cdsco.gov.in/opencms/export/sites/CDSCO_WEB/Pdf-documents/Guidance-document-on-Medical-Device-Software-under-MDR-2017.pdf (text read to section 12.5.2)
- Ch. 1 — US Food and Drug Administration. Changes to Existing Medical Software Policies Resulting from Section 3060 of the 21st Century Cures Act. Final guidance, docket FDA-2017-D-6294. 27 Sep 2019. fda.gov/media/109622/download
- Ch. 1 — US Food and Drug Administration. General Wellness: Policy for Low Risk Devices. Final guidance, docket FDA-2014-N-1039. 6 Jan 2026 (supersedes the version of 27 Sep 2019). fda.gov/media/90652/download
- Ch. 1 — US Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance, docket FDA-2022-D-2628. 4 Dec 2024, reissued 18 Aug 2025. fda.gov/media/166704/download
- Ch. 1 — Akin Gump Strauss Hauer & Feld. IMDRF releases international framework for regulating device software (Oct 2014), on IMDRF/SaMD WG/N12 FINAL:2014, Software as a Medical Device: Possible Framework for Risk Categorization and Corresponding Considerations (18 Sep 2014). akingump.com/en/insights/alerts/imdrf-releases-international-framework-for-regulating-device (secondary; IMDRF page not opened)
- Ch. 2 — US Food and Drug Administration. Ventilator Software Correction: Hamilton Medical Issues Correction for HAMILTON-C6 Medical Ventilators to Address Risk of Failed Ventilation Restart. Recall notice, posted 11 Jul 2024. fda.gov/medical-devices/medical-device-recalls-and-early-alerts/ventilator-software-correction-hamilton-medical-issues-correction-hamilton-c6-medical-ventilators
- Ch. 2 — US Food and Drug Administration. Medical Device Recalls: Class 1 recall Z-2020-2024, ventilator HAMILTON-C6 (Hamilton Medical), software versions 1.1.4–1.1.6; initiated 15 May 2024, posted 18 Jun 2024. accessdata.fda.gov/scripts/cdrh/cfdocs/cfres/res.cfm?id=207956
- Ch. 2 — Wallace D.R., Kuhn D.R. Failure modes in medical device software: an analysis of 15 years of recall data. Int J Reliab Qual Saf Eng 8, 351–371 (2001). doi:10.1142/S021853930100058X (read via the NIST copy, tsapps.nist.gov/publication/get_pdf.cfm?pub_id=917180)
- Ch. 2 — Wallace D.R., Kuhn D.R. Lessons from 342 medical device failures (NIST; conference version, c. 1999) (read via inf.ed.ac.uk/teaching/courses/seoc/2006_2007/resources/CS_342failures.pdf)
- Ch. 2 — US Food and Drug Administration. Tandem Diabetes Care, Inc. Recalls Version 2.7 of the Apple iOS t:connect Mobile App… (used with the t:slim X2 insulin pump). Recall notice, posted 28 Aug 2024. fda.gov/medical-devices/medical-device-recalls-and-early-alerts/tandem-diabetes-care-inc-recalls-version-27-apple-ios-tconnect-mobile-app-used-conjunction-tslim-x2
- Ch. 2 — US Food and Drug Administration. Medical Device Recalls: Class 1 recall Z-1609-2024, t:connect mobile app (Tandem Diabetes Care), root cause recorded as software design change; initiated Mar 2024, posted 6 May 2024. accessdata.fda.gov/scripts/cdrh/cfdocs/cfres/res.cfm?id=206914
- Ch. 3 — Dijkstra E.W. The humble programmer (ACM Turing Lecture 1972; EWD340). Commun ACM 15, 859–866 (1972). doi:10.1145/355604.361591 (read as the EWD340 transcription, cs.utexas.edu/~EWD/transcriptions/EWD03xx/EWD340.html; journal details not checked)
- Ch. 3 — Eypasch E., Lefering R., Kum C.K., Troidl H. Probability of adverse events that have not yet occurred: a statistical reminder. BMJ 311, 619 (1995). doi:10.1136/bmj.311.7005.619 (abstract read)
- Ch. 3 — Butler R.W., Finelli G.B. The infeasibility of quantifying the reliability of life-critical real-time software. IEEE Trans Softw Eng 19, 3–12 (1993). doi:10.1109/32.210303 (read as the NASA preprint, ntrs.nasa.gov/api/citations/20040139817/downloads/20040139817.pdf)
- Ch. 3 — NASA Science. Universe overview (age of the universe, about 13.8 billion years). science.nasa.gov/universe/overview/
- Ch. 4 — IEC. IEC 62304:2006, Medical device software – Software life cycle processes, ed. 1.0 (9 May 2006). IEC Webstore page, webstore.iec.ch/en/publication/6792
- Ch. 4 — IEC. IEC 62304:2006+AMD1:2015 CSV, Medical device software – Software life cycle processes, consolidated ed. 1.1 (26 Jun 2015; stability date 2028). IEC Webstore page, webstore.iec.ch/en/publication/22794; preview pages, including the introduction to Amendment 1, read via webstore.ansi.org
- Ch. 4 — US Food and Drug Administration. Recognized Consensus Standards database: IEC 62304 ed. 1.1 2015-06 consolidated (and ANSI/AAMI IEC 62304:2006/A1:2016), recognition no. 13-79, complete standard; entry 14 Jan 2019. accessdata.fda.gov/scripts/cdrh/cfdocs/cfStandards/detail.cfm?standard__identification_no=38829
- Ch. 4 — IEC. Project records for IEC 62304 ED2, Health software – Software life cycle processes: iec:proj:122433 (stage 30.20, CD study/ballot initiated, 7 Aug 2026; next stage due 27 Nov 2026) and the earlier project iec:proj:23605 (deleted 7 Mar 2023); with iec:proj:11630 and iec:proj:21252 for editions 1.0 and 1.1. iss.rs/en/project/show/iec:proj:122433 (read via iss.rs, the Institute for Standardization of Serbia’s mirror of IEC project data; the iec.ch page was blocked)
- Ch. 4 — VDE. IEC 62304 Edition 2: Änderungen (VDE Health Fachinformation, 22 Jul 2026). vde.com/iec-62304-edition-2-aenderungen (secondary)
- Ch. 4 — Johner Institute. IEC 62304 2nd edition: all areas of application and changes (8 Jul 2026). blog.johner-institute.com/iec-62304-medical-software/iec-62304-2nd-edition-all-areas-of-application-and-changes/ (secondary)
- Ch. 4 — ISO. ISO 14971:2019, Medical devices – Application of risk management to medical devices, 3rd ed. (Dec 2019; confirmed 7 Mar 2025), catalogue page. iso.org/standard/72704.html
- Ch. 4 — International Medical Device Regulators Forum, SaMD Working Group. Characterization Considerations for Medical Device Software and Software-Specific Risk. IMDRF/SaMD WG/N81 FINAL:2025. 27 Jan 2025. imdrf.org/sites/default/files/2025-01/IMDRF_SaMD%20WG_Software-Specific%20Risk_N81%20Final_0.pdf
- Ch. 4 — Wikipedia. Software safety classification (web page quoting IEC 62304:2006+A1:2015, accessed 8 Oct 2026). en.wikipedia.org/wiki/Software_safety_classification (secondary)
- Ch. 4 — LDRA Ltd. Ease the Heartache of Medical Device Software Certification: Achieving Cost-effective Compliance with IEC 62304 – Amendment 1:2015. Technical briefing v2 (Apr 2018). ldra.com/wp-content/uploads/ldra/IEC_62304_Technical_Briefing_v2_04_18.pdf (secondary)
- Ch. 4 — Johner Institute. Software safety classes according to IEC 62304 (rewritten 15 Oct 2025). blog.johner-institute.com/iec-62304-medical-software/safety-class-iec-62304/ (secondary)
- Ch. 4 — US Food and Drug Administration. Content of Premarket Submissions for Device Software Functions. Final guidance, docket FDA-2021-D-0775. 14 Jun 2023 (supersedes the guidance of 11 May 2005). fda.gov/media/153781/download
- Ch. 4 — Matrix Requirements. IEC 62304:2015 impact on class A software (blog). matrixone.health/blog/iec-62304-2015-impact-on-class-a-software (secondary; checked in the subject review)
- Ch. 5 — US Food and Drug Administration. Infusion Pumps Total Product Life Cycle: Guidance for Industry and FDA Staff. Final guidance. 2 Dec 2014. fda.gov/media/78369/download
- Ch. 5 — US Food and Drug Administration. 510(k) K203629, IDx-DR (Digital Diagnostics, Inc.): database record and 510(k) summary; received 11 Dec 2020, decision 10 Jun 2021. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K203629; accessdata.fda.gov/cdrh_docs/pdf20/K203629.pdf
- Ch. 5 — US Food and Drug Administration. 510(k) K213037, IDx-DR v2.3 (Digital Diagnostics, Inc.): database record and 510(k) summary; received 21 Sep 2021, decision 17 Jun 2022. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K213037; accessdata.fda.gov/cdrh_docs/pdf21/K213037.pdf
- Ch. 6 — Microsoft. Windows XP, product lifecycle (Microsoft Learn, accessed 8 Oct 2026). learn.microsoft.com/en-us/lifecycle/products/windows-xp (manufacturer)
- Ch. 6 — National Audit Office (UK). Investigation: WannaCry cyber attack and the NHS. HC 414. 27 Oct 2017. nao.org.uk/reports/investigation-wannacry-cyber-attack-and-the-nhs/ (summary page read)
- Ch. 6 — Dworetzky T. ‘Ransomware’ attack hit U.S. medical devices, too. DOTmed News (18 May 2017). dotmed.com/legal/print/story.html?nid=37359 (secondary, reporting Bayer and Siemens Healthineers)
- Ch. 6 — Palo Alto Networks Unit 42. 2020 Unit 42 IoT Threat Report (10 Mar 2020). unit42.paloaltonetworks.com/iot-threat-report-2020/ (manufacturer)
- Ch. 6 — US Food and Drug Administration. Cybersecurity in Medical Devices: Quality Management System Considerations and Content of Premarket Submissions. Guidance for Industry and FDA Staff, docket FDA-2021-D-1158. 3 Feb 2026 (supersedes the version of 27 Jun 2025). fda.gov/media/119933/download
- Ch. 6 — Johner Institute. Off-the-shelf software (OTS) versus SOUP (web page, undated; quoting IEC 62304:2006+A1:2015). blog.johner-institute.com/iec-62304-medical-software/off-the-shelf-software-ots-versus-soup/ (secondary)
- Ch. 6 — US Food and Drug Administration. Off-The-Shelf Software Use in Medical Devices. Final guidance. 11 Aug 2023 (supersedes the edition of 27 Sep 2019; first issued 9 Sep 1999). fda.gov/media/71794/download
- Ch. 6 — Black Duck Software. 2026 Open Source Security and Risk Analysis (OSSRA) Report (Mar 2026). blackduck.com/resources/analyst-reports/open-source-security-risk-analysis.html (manufacturer)
- Ch. 6 — OWASP CycloneDX. CycloneDX v1.5 released (26 Jun 2023). cyclonedx.org/news/cyclonedx-v1.5-released/
- Ch. 6 — National Telecommunications and Information Administration (US Department of Commerce). The Minimum Elements For a Software Bill of Materials (SBOM). 12 Jul 2021. ntia.gov/files/ntia/publications/sbom_minimum_elements_report.pdf
- Ch. 6 — Medcrypt. Blog post on CISA’s 2026 Minimum Elements for a Software Bill of Materials (29 Jul 2026). medcrypt.com/blog/cisa-sbom-minimum-elements-2026-update (secondary; written by a device-security vendor; CISA’s own 2026 PDF was blocked and not opened)
- Ch. 6 — devops.com. CISA’s 2026 SBOM guidance adds hash requirements and AI coverage (31 Jul 2026). devops.com/cisas-2026-sbom-guidance-adds-hash-requirements-and-ai-coverage/ (secondary)
- Ch. 6 — HSToday. CISA updates software bill of materials guidance to strengthen supply chain security (2026). hstoday.us/subject-matter-areas/cybersecurity/cisa-updates-software-bill-of-materials-guidance-to-strengthen-supply-chain-security/ (secondary)
- Ch. 6 — ISO/IEC. ISO/IEC 5962:2021, Information technology — SPDX® Specification V2.2.1, ed. 1 (Aug 2021; to be revised), catalogue page. iso.org/standard/81870.html
- Ch. 6 — Ecma International. ECMA-424, CycloneDX Bill of materials specification, 1st ed. (Jun 2024) and 2nd ed. (Dec 2025). ecma-international.org/publications-and-standards/standards/ecma-424/
- Ch. 6 — National Telecommunications and Information Administration. Vulnerability-Exploitability eXchange (VEX) – An Overview (27 Sep 2021). ntia.gov/files/ntia/publications/vex_one-page_summary.pdf
- Ch. 6 — United States Code. 21 U.S.C. 360n-2, Ensuring cybersecurity of devices (FD&C Act section 524B), added by Pub. L. 117-328, div. FF, title III, sec. 3305, 29 Dec 2022; effective 29 Mar 2023. law.cornell.edu/uscode/text/21/360n-2 (read via Cornell LII)
- Ch. 6 — US Food and Drug Administration. Cybersecurity in Medical Devices Frequently Asked Questions (FAQs) (content current as of 26 Jun 2025). fda.gov/medical-devices/digital-health-center-excellence/cybersecurity-medical-devices-frequently-asked-questions-faqs
- Ch. 7 — US Food and Drug Administration. Medical Device Recalls: Class 2 recall Z-1233-2023, Monaco RTP System (Elekta), builds 5.11.00–5.11.03, root cause recorded as software design; initiated 28 Feb 2023, posted 8 Mar 2023. accessdata.fda.gov/scripts/cdrh/cfdocs/cfRes/res.cfm?id=198827
- Ch. 8 — AAMI. AAMI TIR45:2012 and AAMI TIR45:2023, Guidance on the use of AGILE practices in the development of medical device software (not opened; editions as listed in the FDA Recognized Consensus Standards database)
- Ch. 8 — US Food and Drug Administration. Recognized Consensus Standards database: AAMI TIR45:2012, recognition no. 13-36 (entry 15 Jan 2013), and AAMI TIR45:2023, recognition no. 13-143 (entry 26 May 2025; declarations of conformity to 13-36 accepted until 2 Jul 2028). accessdata.fda.gov/scripts/cdrh/cfdocs/cfStandards/detail.cfm?standard__identification_no=46298; accessdata.fda.gov/scripts/cdrh/cfdocs/cfstandards/detail.cfm?standard__identification_no=46295
- Ch. 8 — Johner Institute. TIR 45: Agile software development (2016, updated for the 2023 edition). blog.johner-institute.com/iec-62304-medical-software/tir-45-agile-software-development/ (secondary)
- Ch. 8 — AAMI. Key Updates: AAMI TIR45:2023 – Guidance on Agile Practices (training page, undated). aami.org/training/training-suites/software-cybersecurity/key-updates-aami-tir45-2023-guidance-on-agile-practices (publisher’s training page)
- Ch. 8 — ISPE. ISPE GAMP 5: A Risk-Based Approach to Compliant GxP Computerized Systems (Second Edition) (Jul 2022), product page. ispe.org/publications/guidance-documents/gamp-5-guide-2nd-edition
- Ch. 8 — Wyn S., Clark C. What you need to know about GAMP 5 Guide, 2nd Edition. Pharmaceutical Engineering (Jan/Feb 2023). ispe.org/pharmaceutical-engineering/january-february-2023/what-you-need-know-about-gampr-5-guide-2nd-edition
- Ch. 8 — OpenRegulatory. IEC 62304:2006 mapping of requirements to documents (9 Jun 2022). openregulatory.com/document_templates/iec-623042006-mapping-of-requirements-to-documents (secondary; clause titles of IEC 62304, standard not opened)
- Ch. 8 — OpenRegulatory. Doing a software release in compliance with IEC 62304 (24 May 2023, updated 1 Oct 2024). openregulatory.com/articles/software-release-iec-62304 (secondary)
- Ch. 9 — Google Cloud DORA. Accelerate State of DevOps Report 2024 (22 Oct 2024). dora.dev/research/2024/dora-report/2024-dora-accelerate-state-of-devops-report.pdf (manufacturer)
- Ch. 9 — US Food and Drug Administration. Quality Management System Regulation (QMSR) web page; and Medical Devices; Quality System Regulation Amendments, final rule, FR Doc. 2024-01709 (2 Feb 2024; effective 2 Feb 2026). fda.gov/medical-devices/postmarket-requirements-devices/quality-management-system-regulation-qmsr
- Ch. 9 — ISO. ISO/TR 80002-2:2017, Medical device software — Part 2: Validation of software for medical device quality systems, ed. 1 (13 Jun 2017), catalogue page. iso.org/standard/60044.html
- Ch. 9 — US Food and Drug Administration. Computer Software Assurance for Production and Quality System Software. Final guidance, docket FDA-2022-D-0795. 24 Sep 2025. fda.gov/regulatory-information/search-fda-guidance-documents/computer-software-assurance-production-and-quality-system-software
- Ch. 9 — US Food and Drug Administration. Computer Software Assurance for Production and Quality Management System Software. Final guidance, docket FDA-2022-D-0795. 3 Feb 2026 (supersedes the guidance of 24 Sep 2025 and Section 6 of the 2002 software validation guidance). fda.gov/media/188844/download
- Ch. 9 — Johner Institute. IT security for legacy devices (web page, undated; quoting the IEC 62304 Amendment 1 definition of legacy software). blog.johner-institute.com/iec-62304-medical-software/it-security-for-legacy-devices/ (secondary)
- Ch. 9 — Spyro-soft. How to implement a legacy software gap analysis as required by IEC 62304:2015 (blog, undated). spyro-soft.com/blog/how-to-implement-a-legacy-software-gap-analysis-as-required-by-iec-62304-2015 (secondary)
- Ch. 10 — Gulshan V., Peng L., Coram M., et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA 316, 2402–2410 (2016). doi:10.1001/jama.2016.17216 (authors include company staff)
- Ch. 10 — PyTorch. torchvision.models.resnet50, documentation (stable release, accessed 8 Oct 2026). docs.pytorch.org/vision/stable/models/generated/torchvision.models.resnet50.html
- Ch. 10 — US Food and Drug Administration. De Novo DEN190040, Caption Guidance (Bay Labs, Inc.): database record and decision summary; received 27 Aug 2019, granted 7 Feb 2020. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN190040; accessdata.fda.gov/cdrh_docs/reviews/DEN190040.pdf
- Ch. 10 — US Food and Drug Administration. 510(k) K201992, Caption Guidance (Caption Health, Inc.): 510(k) summary; decision 18 Sep 2020. accessdata.fda.gov/cdrh_docs/pdf20/K201992.pdf
- Ch. 10 — PIC/S and EMA GMP/GDP Inspectors Working Group. Draft Annex 22: Artificial Intelligence. Consultation draft (Jul 2025). picscheme.org/docview/9715
- Ch. 10 — US Food and Drug Administration. Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) – Discussion Paper and Request for Feedback. 2 Apr 2019. fda.gov/media/122535/download
- Ch. 11 — Zech J.R., Badgeley M.A., Liu M., Costa A.B., Titano J.J., Oermann E.K. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLOS Med 15, e1002683 (2018). doi:10.1371/journal.pmed.1002683 (authors include company staff)
- Ch. 11 — International Medical Device Regulators Forum, AI/ML-enabled Working Group. Good machine learning practice for medical device development: Guiding principles. IMDRF/AIML WG/N88 FINAL:2025. 27 Jan 2025. imdrf.org/sites/default/files/2025-02/IMDRF_AIML%20WG_GMLP_N88%20Final.pdf
- Ch. 12 — Krause J., Gulshan V., Rahimy E., Karth P., Widner K., Corrado G.S., et al. Grader variability and the importance of reference standards for evaluating machine learning models for diabetic retinopathy. Ophthalmology 125, 1264–1272 (2018). doi:10.1016/j.ophtha.2018.01.034 (read as arXiv:1710.01711v3; journal version not opened) (authors include company staff)
- Ch. 12 — Medicines and Healthcare products Regulatory Agency, US Food and Drug Administration, Health Canada. Good machine learning practice for medical device development: guiding principles (27 Oct 2021). gov.uk/government/publications/good-machine-learning-practice-for-medical-device-development-guiding-principles
- Ch. 12 — US Food and Drug Administration. De Novo DEN170073, ContaCT (Viz.ai, Inc.): database record and decision summary; received 29 Sep 2017, granted 13 Feb 2018. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN170073; accessdata.fda.gov/cdrh_docs/reviews/DEN170073.pdf
- Ch. 12 — Daneshjou R., et al. Diverse Dermatology Images (DDI) dataset, project page (accessed 8 Oct 2026). ddi-dataset.github.io/
- Ch. 13 — NIST/SEMATECH. e-Handbook of Statistical Methods, section 7.2.4.1, Confidence intervals (web page, undated). itl.nist.gov/div898/handbook/prc/section2/prc241.htm
- Ch. 13 — MedCalc Software. Sensitivity and specificity (MedCalc manual, web page, undated). medcalc.org/en/book/sensitivity-specificity.php (secondary)
- Ch. 13 — Wong A., Otles E., Donnelly J.P., Krumm A., McCullough J., DeTroyer-Cooley O., et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern Med 181, 1065–1070 (2021). doi:10.1001/jamainternmed.2021.2626
- Ch. 14 — Wolf R.M., Channa R., Liu T.Y.A., et al. Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial. Nat Commun 15, 421 (2024). doi:10.1038/s41467-023-44676-z (authors include the company’s founder; youth accuracy of the earlier SEE study as reported in this paper)
- Ch. 14 — Heaven W.D. Google’s medical AI was super accurate in a lab. Real life was a different story. MIT Technology Review (27 Apr 2020). technologyreview.com/2020/04/27/1000658/google-medical-ai-accurate-lab-real-life-clinic-covid-diabetes-retina-disease/ (secondary)
- Ch. 14 — Beede E. Healthcare AI systems that put people at the center. Google Keyword blog (25 Apr 2020). blog.google/technology/health/healthcare-ai-systems-put-people-center/ (manufacturer)
- Ch. 14 — Beede E., Baylor E., Hersch F., Iurchenko A., Wilcox L., Ruamviboonsuk P., et al. A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (2020). research.google/pubs/a-human-centered-evaluation-of-a-deep-learning-system-deployed-in-clinics-for-the-detection-of-diabetic-retinopathy/ (abstract read) (authors include company staff)
- Ch. 14 — Ruamviboonsuk P., Tiwari R., Sayres R., Nganthavee V., Hemarat K., Kongprayoon A., et al. Real-time diabetic retinopathy screening by deep learning in a multisite national screening programme: a prospective interventional cohort study. Lancet Digit Health 4, e235–e244 (2022). doi:10.1016/S2589-7500(22)00017-6 (abstract read, via DOAJ) (authors include company staff; funded by Google and Rajavithi Hospital)
- Ch. 15 — Daneshjou R., et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. arXiv:2203.08807 v1 (Mar 2022). arxiv.org/pdf/2203.08807 (preprint)
- Ch. 15 — Daneshjou R., Vodrahalli K., Novoa R.A., et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Sci Adv 8, eabq6147 (2022). doi:10.1126/sciadv.abq6147 (abstract read)
- Ch. 15 — Seyyed-Kalantari L., Zhang H., McDermott M.B.A., Chen I.Y., Ghassemi M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med 27, 2176–2182 (2021). doi:10.1038/s41591-021-01595-0 (text read; tables not accessible)
- Ch. 15 — US Food and Drug Administration. Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests. Final guidance. 13 Mar 2007. fda.gov/media/71147/download
- Ch. 15 — Emergo by UL. India CDSCO finalizes guidance on medical device software (30 Jul 2026). emergobyul.com/news/india-cdsco-finalizes-guidance-medical-device-software (secondary)
- Ch. 15 — Regulation (EU) 2024/1689 of the European Parliament and of the Council (Artificial Intelligence Act). OJ L (12 Jul 2024). artificialintelligenceact.eu/article/3/ (Articles 3, 4, 6, 8–15, 26, 43 and 113, as amended, read via artificialintelligenceact.eu, a transcription of the OJ text)
- Ch. 15 — Emergo by UL. Decoding MDCG 2025-6: interplay between the MDR/IVDR and the AI Act (24 Jun 2025). emergobyul.com/news/decoding-mdcg-2025-6-interplay-between-mdrivdr-and-aia (secondary)
- Ch. 16 — Dratsch T., Chen X., Rezazade Mehrizi M., Kloeckner R., Mähringer-Kunz A., Püsken M., et al. Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology 307, e222176 (2023). doi:10.1148/radiol.222176 (abstract read)
- Ch. 16 — US Food and Drug Administration. Clinical Decision Support Software. Final guidance, docket FDA-2017-D-6569. 6 Jan 2026, corrected 29 Jan 2026 (supersedes the September 2022 version); with the guidance web page. fda.gov/media/109618/download; fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
- Ch. 16 — US Food and Drug Administration. Clinical Performance Assessment: Considerations for Computer-Assisted Detection Devices Applied to Radiology Images and Radiology Device Data in Premarket Notification (510(k)) Submissions. Final guidance, docket FDA-2009-D-0503. 28 Sep 2022. fda.gov/media/77642/download
- Ch. 16 — US Food and Drug Administration, Health Canada, Medicines and Healthcare products Regulatory Agency. Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles (13 Jun 2024). fda.gov/medical-devices/software-medical-device-samd/transparency-machine-learning-enabled-medical-devices-guiding-principles
- Ch. 16 — Goddard K., Roudsari A., Wyatt J.C. Automation bias: a systematic review of frequency, effect mediators, and mitigators. J Am Med Inform Assoc 19(1), 121–127 (2012). doi:10.1136/amiajnl-2011-000089 (abstract read; added in the review round of 14 Oct 2026)
- Ch. 17 — Digital Diagnostics. LumineticsCore product page (last modified 5 Aug 2026; accessed 8 Oct 2026). digitaldiagnostics.com/idx-dr/ (manufacturer)
- Ch. 18 — NEMA, DICOM Standards Committee. DICOM PS3.6 2026d, Data Dictionary, chapter 6: Registry of DICOM Data Elements. dicom.nema.org/medical/dicom/current/output/chtml/part06/chapter_6.html
- Ch. 18 — HL7 International. HL7 Version 2 Product Suite, product brief (web page, undated). hl7.org/implement/standards/product_brief.cfm?product_id=185
- Ch. 18 — HL7 International. FHIR Publication (Version) History (web page, accessed 8 Oct 2026). hl7.org/fhir/history.html
- Ch. 18 — HL7 International. FHIR R5 (v5.0.0), Resource (26 Mar 2023). hl7.org/fhir/R5/resource.html
- Ch. 18 — HL7 International. FHIR v6.0.0-ballot5, CodeSystem FHIR-version (page generated 17 Jul 2026). hl7.org/fhir/6.0.0-ballot5/codesystem-FHIR-version.html
- Ch. 18 — IHE Radiology Technical Committee. IHE Radiology Technical Framework Supplement: AI Results (AIR), Rev. 1.3, Trial Implementation. 8 Aug 2025. ihe.net/uploadedFiles/Documents/Radiology/IHE_RAD_Suppl_AIR.pdf
- Ch. 18 — IHE Radiology Technical Committee. IHE Radiology Technical Framework Supplement: AI Workflow for Imaging (AIW-I), Rev. 1.1, Trial Implementation. 6 Aug 2020. ihe.net/uploadedFiles/Documents/Radiology/IHE_RAD_Suppl_AIW-I.pdf
- Ch. 18 — Martinez-Gutierrez J.C., Kim Y., Salazar-Marioni S., et al. Automated large vessel occlusion detection software and thrombectomy treatment times: a cluster randomized clinical trial. JAMA Neurol 80, 1182–1190 (2023). doi:10.1001/jamaneurol.2023.3206 (abstract read, via scholars.aku.edu)
- Ch. 18 — UTHealth Houston, McWilliams School of Biomedical Informatics. News release on the study of artificial intelligence software and endovascular thrombectomy treatment times (2023). sbmi.uth.edu/news/story/uthealth-houston-study-artificial-intelligence-software-improves-endovascular-thrombectomy-treatment-times-for-stroke-patients (secondary)
- Ch. 18 — Medical Device Coordination Group. MDCG 2019-16 rev.1, Guidance on Cybersecurity for medical devices. Dec 2019, rev.1 Jul 2020 (quoting Regulation (EU) 2017/745, Annex I, sections 17.2 and 17.4). health.ec.europa.eu/system/files/2022-01/md_cybersecurity_en.pdf
- Ch. 18 — IEC. IEC 80001-1:2021, Application of risk management for IT-networks incorporating medical devices – Part 1: Safety, effectiveness and security in the implementation and use of connected medical devices or connected health software, ed. 2.0 (21 Sep 2021). IEC Webstore page, webstore.iec.ch/en/publication/34263; ISO catalogue page (stage 90.92, to be revised, 4 Feb 2026), iso.org/standard/72026.html
- Ch. 18 — ISO. ISO/TR 80001-2-6:2014, Application of risk management for IT-networks incorporating medical devices – Part 2-6: Application guidance – Guidance for responsibility agreements (catalogue entry read via implementer.digitalhealth.gov.au, Australian Digital Health Agency)
- Ch. 19 — Brady A.P., Allen B., Chong J., Kotter E., Kottler N., Mongan J., et al. Developing, purchasing, implementing and monitoring AI tools in radiology: practical considerations. A multi-society statement from the ACR, CAR, ESR, RANZCR & RSNA. Can Assoc Radiol J 75, 226–244 (2024). doi:10.1177/08465371231222229
- Ch. 19 — Kwong J.C.C., Erdman L., Khondker A., Skreta M., Goldenberg A., McCradden M.D., et al. The silent trial: the bridge between bench-to-bedside clinical AI applications. Front Digit Health 4, 929508 (2022). doi:10.3389/fdgth.2022.929508
- Ch. 19 — IntuitionLabs. GAMP 5 categories explained (16 Oct 2025, updated 8 Aug 2026). intuitionlabs.ai/articles/gamp-5-categories-explained (secondary)
- Ch. 19 — US Code of Federal Regulations (eCFR). 21 CFR Part 11, Electronic Records; Electronic Signatures (62 FR 13464, 20 Mar 1997). ecfr.gov/current/title-21/chapter-I/subchapter-A/part-11
- Ch. 19 — Blumenthal R., Erdmann N., Heitmann M., Lemettinen A.-L., Stockton B.M. Machine learning risk and control framework. Pharmaceutical Engineering (Jan/Feb 2024). ispe.org/pharmaceutical-engineering/january-february-2024/machine-learning-risk-and-control-framework
- Ch. 19 — ISPE. ISPE GAMP Guide: Artificial Intelligence (Jul 2025), product page. ispe.org/publications/guidance-documents/gamp-guide-artificial-intelligence
- Ch. 19 — Stockton B., Staib E., Heitmann M. New GAMP Guide addresses challenges posed by AI-enabled computerized systems. Pharmaceutical Engineering (Sep/Oct 2025). ispe.org/pharmaceutical-engineering/september-october-2025/new-gampr-guide-addresses-challenges-posed-ai
- Ch. 19 — European Commission, DG SANTE. Stakeholders’ Consultation on EudraLex Volume 4 – Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22 (7 Jul to 7 Oct 2025). health.ec.europa.eu/consultations/stakeholders-consultation-eudralex-volume-4-good-manufacturing-practice-guidelines-chapter-4-annex_en
- Ch. 19 — European Commission. EudraLex – Volume 4, Good Manufacturing Practice guidelines (web page, accessed 8 Oct 2026). health.ec.europa.eu/medicinal-products/eudralex/eudralex-volume-4_en
- Ch. 20 — Michigan Medicine Health Lab. Study of 24 U.S. hospitals shows onset of COVID-19 led to spike in sepsis alerts (6 Dec 2021), reporting Wong A., et al. Quantification of sepsis model alerts in 24 US hospitals before and during the COVID-19 pandemic. JAMA Netw Open 4, e2135286 (2021). doi:10.1001/jamanetworkopen.2021.35286. michiganmedicine.org/health-lab/study-24-us-hospitals-shows-onset-covid-19-led-spike-sepsis-alerts (secondary; the paper itself was blocked and not opened)
- Ch. 20 — Davis S.E., Lasko T.A., Chen G., Siew E.D., Matheny M.E. Calibration drift in regression and machine learning models for acute kidney injury. J Am Med Inform Assoc 24, 1052–1061 (2017). doi:10.1093/jamia/ocx030 (abstract read)
- Ch. 20 — US Food and Drug Administration, CDRH Digital Health Center of Excellence. Request For Public Comment: Measuring and Evaluating Artificial Intelligence-enabled Medical Device Performance in the Real-World. Docket FDA-2025-N-4203. 30 Sep 2025. fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device
- Ch. 21 — US Food and Drug Administration. Artificial Intelligence in Software as a Medical Device (web page with milestones, content current as of 25 Mar 2025). fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-software-medical-device
- Ch. 21 — US Food and Drug Administration, Health Canada, Medicines and Healthcare products Regulatory Agency. Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles (Oct 2023; web page republished 18 Aug 2025). fda.gov/medical-devices/software-medical-device-samd/predetermined-change-control-plans-machine-learning-enabled-medical-devices-guiding-principles
- Ch. 21 — US Food and Drug Administration. Predetermined Change Control Plans for Medical Devices. Draft guidance, docket FDA-2024-D-2338. Aug 2024. fda.gov/regulatory-information/search-fda-guidance-documents/predetermined-change-control-plans-medical-devices
- Ch. 21 — Dayma K., Patel P., Hildreth K., Jamaspishvili T. Predetermined change control plan adoption and documentation transparency in U.S. Food and Drug Administration–cleared radiology artificial intelligence/machine learning devices. Radiol Artif Intell 8(5) (online 29 Jul 2026). doi:10.1148/ryai.260385 (abstract read)
- Ch. 21 — US Food and Drug Administration. 510(k) K253281, UpDoc (Updoc, Inc.): database record and 510(k) summary; received 29 Sep 2025, decision 23 Dec 2025. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K253281; accessdata.fda.gov/cdrh_docs/pdf25/K253281.pdf
- Ch. 22 — Abbott. Important Cybersecurity Advisory: pacemaker firmware update, letter to physicians (US), 28 Aug 2017. cardiovascular.abbott/content/dam/cv/cardiovascular/pdf/reports/Pacemaker-Firmware-Update-Doctor-Letter-Aug2017-US.pdf (manufacturer)
- Ch. 22 — US Food and Drug Administration. Medical Device Recalls: recall event 78093 (St. Jude Medical pacemakers, Merlin PCS programmer software and Merlin@home software; records Z-0029 to Z-0038-2018, Class 2), initiated 28 Aug 2017, posted 12 Jun 2018. accessdata.fda.gov/scripts/cdrh/cfdocs/cfres/res.cfm?start_search=1&event_id=78093
- Ch. 22 — American College of Cardiology. FDA approves firmware addressing cybersecurity vulnerabilities in Abbott implantable pacemakers (31 Aug 2017). acc.org/latest-in-cardiology/articles/2017/08/31/12/13/fda-approves-firmware-addressing-cybersecurity-vulnerabilities-in-abbott-implantable-pacemakers (secondary)
- Ch. 22 — CSO Online. 465,000 Abbott pacemakers vulnerable to hacking, need a firmware fix (4 Sep 2017). csoonline.com/article/3222068/465000-abbott-pacemakers-vulnerable-to-hacking-need-a-firmware-fix.html (secondary)
- Ch. 22 — US Food and Drug Administration. FDA warns patients and health care providers about potential cybersecurity concerns with certain Medtronic insulin pumps. Press release, 27 Jun 2019. biospace.com/fda-warns-patients-and-health-care-providers-about-potential-cybersecurity-concerns-with-certain-medtronic-insulin-pumps (read via the BioSpace reprint; the FDA safety communication returned 404)
- Ch. 22 — US Food and Drug Administration. Medical Device Recalls: Class 2 recall Z-1581-2020, MiniMed insulin pump MMT-508 (Medtronic), initiated 27 Jun 2019, posted 26 Mar 2020, status open; and recall event 83433 (16 records, MiniMed 508 and Paradigm models). accessdata.fda.gov/scripts/cdrh/cfdocs/cfRES/res.cfm?id=175194; accessdata.fda.gov/scripts/cdrh/cfdocs/cfRes/res.cfm?start_search=1&event_id=83433
- Ch. 22 — Cybersecurity and Infrastructure Security Agency (NCCIC). ICS Medical Advisory ICSMA-19-178-01: Medtronic MiniMed 508 and Paradigm Series Insulin Pumps (27 Jun 2019, date inferred from the advisory number). cisa.gov/news-events/ics-medical-advisories/icsma-19-178-01
- Ch. 22 — Cybersecurity and Infrastructure Security Agency (NCCIC). ICS Medical Advisory ICSMA-18-219-02: Medtronic MiniMed 508 and Paradigm Series insulin pumps, remote controllers (7 Aug 2018; updated). cisa.gov/news-events/ics-medical-advisories/icsma-18-219-02
- Ch. 22 — US Food and Drug Administration. Cybersecurity (Digital Health Center of Excellence web page, content current as of 6 Jul 2026), listing the Playbook for Threat Modeling Medical Devices (30 Nov 2021). fda.gov/medical-devices/digital-health-center-excellence/cybersecurity
- Ch. 22 — US Food and Drug Administration. Cybersecurity Vulnerabilities with Certain Patient Monitors from Contec and Epsimed: FDA Safety Communication. 30 Jan 2025, updated 2 Jul 2025. fda.gov/medical-devices/safety-communications/cybersecurity-vulnerabilities-certain-patient-monitors-contec-and-epsimed-fda-safety-communication
- Ch. 22 — MITRE and Medical Device Innovation Consortium. Playbook for Threat Modeling Medical Devices (Nov 2021; FDA-funded). Release: mitre.org/news-insights/news-release/mitre-and-medical-device-innovation-consortium-create-playbook-threat; mdic.org/resource/playbook-for-threat-modeling-medical-devices/
- Ch. 23 — Regenscheid A. Platform Firmware Resiliency Guidelines. NIST SP 800-193. May 2018. csrc.nist.gov/pubs/sp/800/193/final (abstract page read)
- Ch. 23 — IEC. IEC 81001-5-1:2021, Health software and health IT systems safety, effectiveness and security – Part 5-1: Security – Activities in the product life cycle, ed. 1.0 (Dec 2021). standards.iteh.ai/catalog/standards/iso/227a3148-4206-42e3-ad21-05f80dda507a/iec-81001-5-1-2021 (standard’s front matter read via a reseller preview)
- Ch. 23 — ISO. IEC 81001-5-1:2021, catalogue page (published 21 Dec 2021; stage 90.92, to be revised, since 19 Sep 2025). iso.org/standard/76097.html
- Ch. 23 — Blue Goat Cyber. Did AAMI SW96 replace TIR57? (blog post, 21 Jul 2026). bluegoatcyber.com/blog/did-aami-sw96-replace-tir57-fda-2026 (secondary)
- Ch. 23 — AAMI. ANSI/AAMI SW96:2023, Standard for medical device security – Security risk management for device manufacturers (not opened; year and alignment with ISO 14971 as reported by AHA News and MedTech Intelligence)
- Ch. 23 — AHA News. FDA recognizes ANSI/AAMI medical device standard to enhance cybersecurity (8 Nov 2023). aha.org/news/headline/2023-11-08-fda-recognizes-ansiaami-medical-device-standard-enhance-cybersecurity (secondary)
- Ch. 23 — MedTech Intelligence. FDA recognizes AAMI SW96 cybersecurity guidance document (15 Nov 2023). medtechintelligence.com/news_article/fda-recognizes-aami-sw96-cybersecurity-guidance-document/ (secondary)
- Ch. 24 — Armis. URGENT/11 (research page; updated 1 Oct 2019 and 15 Dec 2020). armis.com/research/urgent-11/ (manufacturer)
- Ch. 24 — US Food and Drug Administration. FDA informs patients, providers and manufacturers about potential cybersecurity vulnerabilities for connected medical devices and health care networks that use certain communication software. Press release, 1 Oct 2019. fda.gov/news-events/press-announcements/fda-informs-patients-providers-and-manufacturers-about-potential-cybersecurity-vulnerabilities
- Ch. 24 — Whooley S. B. Braun, Baxter, Carestream, Green Hills affected by Ripple20 cyber vulnerabilities. MassDevice (29 Jun 2020). massdevice.com/b-braun-baxter-carestream-green-hills-affected-by-ripple20-cyber-vulnerabilities/ (secondary)
- Ch. 24 — FIRST. FIRST has officially published the latest version of CVSS (v4.0). Press release, 1 Nov 2023. first.org/newsroom/releases/20231101
- Ch. 24 — FIRST. Common Vulnerability Scoring System v4.0: Specification Document (document version 1.2). first.org/cvss/v4-0/specification-document
- Ch. 24 — Chase M., Christey Coley S. Rubric for Applying CVSS to Medical Devices. MITRE (web page dated 21 Oct 2020). mitre.org/md-cvss-rubric
- Ch. 24 — US Food and Drug Administration. New Medical Device Development Tool (MDDT) Qualification for Cybersecurity. Bulletin, 20 Oct 2020. content.govdelivery.com/accounts/USFDA/bulletins/2a6cf79
- Ch. 24 — US Food and Drug Administration. Postmarket Management of Cybersecurity in Medical Devices. Final guidance. 28 Dec 2016. fda.gov/files/medical%20devices/published/Postmarket-Management-of-Cybersecurity-in-Medical-Devices---Guidance-for-Industry-and-Food-and-Drug-Administration-Staff.pdf
- Ch. 24 — ISO/IEC. ISO/IEC 29147:2018, Information technology — Security techniques — Vulnerability disclosure, ed. 2 (23 Oct 2018; to be revised since 26 Sep 2025), catalogue page. iso.org/standard/72311.html
- Ch. 24 — ISO/IEC. ISO/IEC 30111:2019, Information technology — Security techniques — Vulnerability handling processes, ed. 2 (1 Oct 2019; to be revised since 26 Sep 2025), catalogue page. iso.org/standard/69725.html
- Ch. 24 — European Commission. Guidance – MDCG endorsed documents and other guidance (web page, accessed 8 Oct 2026). health.ec.europa.eu/medical-devices-sector/new-regulations/guidance-mdcg-endorsed-documents-and-other-guidance_en
- Ch. 25 — US Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices (list; page updated 4 Sep 2026). fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices
- Ch. 25 — IntuitionLabs. FDA-approved AI medical devices list: complete 2026 guide (19 Jul 2026), citing counts by TheImagingWire and Innolitics. intuitionlabs.ai/articles/fda-approved-ai-medical-devices-list (secondary)
- Ch. 25 — US Food and Drug Administration. Digital Health Advisory Committee meeting of 6 Nov 2025, Generative Artificial Intelligence-Enabled Digital Mental Health Medical Devices: agenda, discussion questions and brief summary. fda.gov/media/189389/download; fda.gov/media/189392/download; fda.gov/media/190450/download
- Ch. 25 — GE HealthCare. GE HealthCare drives growth with investment in AI-enabled medical devices and tops FDA’s list of AI authorizations for 4th year with 100. Press release, 23 Jul 2025. gehealthcare.com/en/about/newsroom/press-releases/ge-healthcare-drives-growth-with-investment-in-ai-enabled-medical-devices-and-tops-fdas-list-of-ai-authorizations-for-4th-year-with-100 (manufacturer)
- Ch. 25 — US Food and Drug Administration. 510(k) K200921, qER (Qure.ai, Mumbai): database record; received 6 Apr 2020, decision 17 Jun 2020. accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K200921
- Ch. 25 — DAIC. Qure.ai’s chest X-ray reporting tool receives additional FDA clearances (26 Feb 2026). dicardiology.com/content/qureais-chest-x-ray-reporting-tool-receives-additional-fda-clearances (secondary, reporting the manufacturer)
- Ch. 25 — Qure.ai. Regulatory and privacy (web page, accessed 8 Oct 2026). qure.ai/regulatory-and-privacy (manufacturer)
- Ch. 25 — HLTH. Aidoc wins FDA nod for comprehensive foundation model AI (26 Jan 2026). hlth.com/insights/news/aidoc-wins-fda-nod-for-comprehensive-foundation-model-ai-2026-01-26 (secondary, reporting the manufacturer)
- Ch. 25 — Tijori Alerts. Lords Mark receives India’s first Biomescan SaMD manufacturing licence (10 Aug 2026). tijorialerts.com/company-updates/lords-mark-receives-indias-first-biomescan-samd-manufacturing-licence-139868/ (secondary, reporting the manufacturer)
- Ch. 25 — US Food and Drug Administration. November 20–21, 2024: Digital Health Advisory Committee Meeting Announcement (Total Product Lifecycle Considerations for Generative Artificial Intelligence-Enabled Medical Devices). fda.gov/advisory-committees/advisory-committee-calendar/november-20-21-2024-digital-health-advisory-committee-meeting-announcement-11202024
- Ch. 25 — McGuireWoods. A pathway for clinical AI developers opens: FDA clears first software as a medical device with patient-facing LLM. Legal alert, 6 Jul 2026. mcguirewoods.com/client-resources/alerts/2026/7/a-pathway-for-clinical-ai-developers-opens-fda-clears-first-software-as-a-medical-device-with-patient-facing-llm/ (secondary, reporting the manufacturer’s claim)
- Ch. 25 — Aguilar M. A ‘historic’ FDA clearance raises the question: is the LLM an interface or the decision-maker? STAT (2 Jul 2026). statnews.com/2026/07/02/fda-clearance-raises-questions-updoc-use-generative-ai-diabetes-treatment/ (secondary; read in part, paywalled)
- Ch. 26 — Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI). OJ L (24 Jul 2026); in force 27 Jul 2026. eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32026R1744 (read in two parts, with gaps)
- Ch. 26 — Taylor Wessing. Update: AI-enabled medical devices and IVDs confirmed as high risk (Life Sciences Legal Lens, vol. 2; 2026). taylorwessing.com/en/international-life-sciences-newsletter/life-sciences-legal-lens-vol-2/update-ai-enabled-medical-devices-and-ivds-confirmed-as-high-risk (secondary)
- Ch. 26 — Covington & Burling. 5 key takeaways from FDA’s revised clinical decision support (CDS) software guidance (8 Jan 2026). cov.com/en/news-and-insights/insights/2026/01/5-key-takeaways-from-fdas-revised-clinical-decision-support-cds-software-guidance (secondary)
- Ch. 26 — OpenRegulatory. MDCG 2019-11 explained (web page dated 12 Aug 2026). openregulatory.com/mdcg/mdcg-2019-11 (secondary)
- Ch. 26 — GMP Insiders. MDCG 2019-11 Rev.1: qualification and classification of software under the MDR and IVDR (web page). gmpinsiders.com/mdcg-2019-11-rev1-mdr-ivdr/ (secondary)
- Ch. 26 — Sobande S. European revision of primary software guidance MDCG 2019-11, revision 1. Emergo by UL (20 Jun 2025). emergobyul.com/news/european-revision-primary-software-guidance-mdcg-2019-11-revision-1-small-changes-meaningful (secondary)
- Ch. 26 — Rao A. CDSCO draft guidance on medical device software. India Briefing (12 Nov 2025). india-briefing.com/news/cdsco-draft-guidance-medical-software-40691.html (secondary)
- Ch. 26 — Nishith Desai Associates. From code to compliance: the CDSCO guidance on medical device software (3 Aug 2026) (secondary)
- Ch. 26 — US Food and Drug Administration. How to Determine if Your Product is a Medical Device (content current 29 Sep 2022). fda.gov/medical-devices/classify-your-medical-device/how-determine-if-your-product-medical-device
- Ch. 26 — US Food and Drug Administration. Classify Your Medical Device (content current 26 Aug 2026). fda.gov/medical-devices/overview-device-regulation/classify-your-medical-device
- Ch. 27 — US Food and Drug Administration. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft guidance, docket FDA-2024-D-4488. Cover date 7 Jan 2025 (web page published 6 Jan 2025). fda.gov/media/184856/download; fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing (PDF read to partway through Performance Validation)
- Ch. 27 — Medical Device Coordination Group. MDCG 2020-1, Guidance on Clinical Evaluation (MDR) / Performance Evaluation (IVDR) of Medical Device Software. Mar 2020. ec.europa.eu/docsroom/documents/40323
- Ch. 27 — Commission Implementing Decision (EU) 2021/1182 on harmonised standards for medical devices, consolidated text to 17 Jun 2026. eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:02021D1182-20260617
- Ch. 27 — Critical Software. IEC 62304 Edition 2 changes (web article, undated). asd.criticalsoftware.com/en/newsroom/iec-62304-edition-2-changes-august-2026 (secondary)
- Ch. 28 — US Food and Drug Administration. Deciding When to Submit a 510(k) for a Software Change to an Existing Device. Final guidance, docket FDA-2016-D-2021. 25 Oct 2017. fda.gov/media/99785/download
- Ch. 28 — Medical Device Coordination Group. MDCG 2020-3, Guidance on significant changes regarding the transitional provision under Article 120 of the MDR. Mar 2020 (rev.1 of Sep 2023 not opened). ec.europa.eu/docsroom/documents/40301
- Ch. 28 — Central Drugs Standard Control Organisation, India. IVD Medical Devices FAQ, Doc No. CDSCO/IVD/FAQ/04/2022, with addendum of 28 Mar 2025. cdsco.gov.in/opencms/export/sites/CDSCO_WEB/Pdf-documents/FAaddendum.pdf (read by the G-06 researcher)
- Ch. 28 — Artixio. CDSCO medical device post approval change application (19 Aug 2026). artixio.com/post/cdsco-medical-device-post-approval-change-application (secondary)
- Ch. 28 — Central Drugs Standard Control Organisation, India. Addendum No. 02 to Doc No. CDSCO/IVD/FAQ/04/2022 (13 Mar 2026). cdsco.gov.in/opencms/export/sites/CDSCO_WEB/Pdf-documents/Addendum-Doc-No-CDSCOIVDFAQ-042022dated-13032026.pdf (read by the G-06 researcher)
- Ch. 28 — Regulation (EU) 2017/745 on medical devices, Annex X section 5 (changes to the approved type). OJ L 117 (5 May 2017). Text read as reproduced at advisera.com/13485academy/mdr/conformity-assessment-based-on-type-examination
- Ch. 29 — DLA Piper. FDA issues revised cybersecurity premarket submission guidance (27 Feb 2026). dlapiper.com/en/insights/publications/2026/02/fda-issues-revised-cybersecurity-premarket-submission-guidance (secondary)
- Ch. 29 — NSF. FDA updates cybersecurity guidance: shift toward QMSR alignment rather than new requirements (3 Feb 2026). nsf.org/life-science-regulatory-news/fda-updates-cybersecurity-guidance-shift-toward-qmsr-alignment-rather-than-new-requirements (secondary)
- Ch. 29 — AAMI. Food and Drug Administration adds AAMI cybersecurity guidance to Recognized Consensus Standards Database (AAMI CR515:2025, Cybersecurity considerations unique to machine learning-enabled medical devices). Press release via Newsfile, 24 Mar 2026. newsfilecorp.com/release/289622/Food-and-Drug-Administration-Adds-AAMI-Cybersecurity-Guidance-to-Recognized-Consensus-Standards-Database (standards developer’s press release)
- Ch. 29 — Regulation (EU) 2024/2847 of the European Parliament and of the Council of 23 October 2024 (Cyber Resilience Act). OJ L (20 Nov 2024); in force 10 Dec 2024. eur-lex.europa.eu/eli/reg/2024/2847/oj (recital 25 read on EUR-Lex; Articles 2 and 71 read via springlex.eu)
- Ch. 29 — Pure Global. India CDSCO medical device software guidance 2026 (9 Aug 2026). pureglobal.com/news/india-cdsco-medical-device-software-guidance-2026 (secondary)
- Ch. 29 — BioSpectrum India. CDSCO releases guidance document on medical device software (27 Jul 2026). biospectrumindia.com/news/93/28210/cdsco-releases-guidance-document-on-medical-device-software.html (secondary)
- Ch. 29 — LexCounsel. CDSCO publishes guidance on medical device software under MDR 2017. Mondaq (5 Aug 2026). mondaq.com/cdsco-publishes-guidance-on-medical-device-software-under-mdr-2017/1826286 (secondary)
- Ch. 29 — Press Information Bureau, Government of India. Digital Personal Data Protection Rules, 2025: explainer (17 Nov 2025). static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf
- Ch. 29 — S.S. Rana & Co. MeitY notifies final Digital Personal Data Protection Rules, 2025 (14 Nov 2025). ssrana.in/articles/meity-notifies-final-digital-personal-data-protection-rules-2025/ (secondary)
- Ch. 29 — Hogan Lovells. India’s Digital Personal Data Protection Act 2023 brought into force (Nov 2025). ca.hoganlovells.com/en/publications/indias-digital-personal-data-protection-act-2023-brought-into-force- (secondary)
Take the PDF with you.
Reading online needs no sign-up. For the PDF edition, laid out for print and offline reading, tell us where to send it and we will email you.
Check your inbox.
The download link is on its way to your email. If it has not arrived in a few minutes, check your spam folder or write to info@unplex.tech.
Preview only — nothing was sent
Let's Talk