The 2017 decision support proposal drew the regulatory line at whether a clinician could check the reasoning
On December 8, 2017, the Food and Drug Administration proposed how it would read the Cures Act carve-out for clinical decision support software. In our reading, it made explainability the test, turning a design choice into a compliance question for health systems.
Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.
Two-minute summaryThe dek and the first sentence of each section, quoted from the article
On December 8, 2017, the Food and Drug Administration proposed how it would read the Cures Act carve-out for clinical decision support software. In our reading, it made explainability the test, turning a design choice into a compliance question for health systems.
- The week On Friday, December 8, 2017, the Food and Drug Administration published three guidance documents on medical software in the Federal Register.
- Inputs: information stayed outside, signals stayed in The first statutory condition is a carve-out from the carve-out.
- Audience: professionals were excluded, patients got a policy The third condition limits the statutory exclusion to recommendations made to a health care professional.
- Basis: the clinician had to be able to check the recommendation The fourth condition decided the hard cases on the information side, and the draft’s reading of it is the core of the document.
- Plumbing: records, data and administration moved outside FDA’s reach The companion draft dealt with the software around decision support.
- How the drafts applied to different health organizations The two drafts addressed developers.
- Objections, in order Five objections deserve an answer.
- What a leader could check These checks follow from the four parts.
- With hindsight FDA did not finalize the 2017 decision support draft as written.
The week
On Friday, December 8, 2017, the Food and Drug Administration published three guidance documents on medical software in the Federal Register. Two were proposals explaining how the agency read the software provisions of the 21st Century Cures Act. The third, in final form, covered how to evaluate software that is itself a regulated medical product. The decision support notice was dated December 4.
The Cures Act was enacted on December 13, 2016. Its section 3060(a), titled Clarifying Medical Software Regulation, added a new subsection, 520(o), to the Federal Food, Drug, and Cosmetic Act. That subsection listed software functions taken out of the statutory definition in section 201(h). A conforming change to 201(h) said the definition does not cover the functions excluded under 520(o).
The first proposal, Clinical and Patient Decision Support Software, dealt with the fifth category in the list, section 520(o)(1)(E). That provision takes such software out of the definition only when it meets four conditions. The software must not be intended to acquire, process or analyze a medical image, a signal from an in vitro diagnostic product, or a pattern or signal from a signal acquisition system. It must be intended to display, analyze or print medical information. It must be intended to support or provide recommendations to a health care professional. And it must be intended to let that professional independently review the basis for the recommendations, so that the professional does not rely primarily on them.
The proposal defined clinical decision support, or CDS, as software that meets the first three conditions. It said CDS is excluded only when it also meets the fourth. The agency read that condition to describe software that clearly explains four things: its purpose or intended use, its intended user, the inputs used to generate the recommendation, and the rationale or support for it. The text said the intended user should be able to reach the same recommendation without relying primarily on the software. It said the sources should be identified, understandable to the user and publicly available, such as clinical practice guidelines and published literature.
The proposal also created a second category, patient decision support software, for functions intended for patients and caregivers who are not health care professionals. The statute’s exclusion did not cover these. The agency said it intended to adopt an enforcement discretion policy for them that paralleled the professional exclusion, using four matching factors. Its example of a function it would still oversee was software that recommends warfarin dose changes from a home blood test result without the patient consulting a provider.
The second proposal, Changes to Existing Medical Software Policies Resulting from Section 3060 of the 21st Century Cures Act, covered the other four categories. These were software for administrative support of a health care facility, for maintaining a healthy lifestyle, for serving as electronic patient records, and for transferring, storing, converting formats or displaying laboratory and instrument data. The agency proposed edits to four existing guidances, on general wellness, mobile medical applications, off-the-shelf software and data systems. It proposed withdrawing a fifth, on premarket notifications for medical image management products.
That text drew several specific lines. Laboratory information management systems fell outside the definition. Electronic patient record functions were excluded under three conditions. The records were created or reviewed by professionals. They were part of health IT certified under the program of the Office of the National Coordinator for Health Information Technology. And the function did not interpret or analyze them. The agency said it did not intend to enforce its requirements against record functions that lacked that certification but met the other conditions. Software that only transferred, stored, converted or displayed instrument data fell outside, but software that generated alarms or prioritized a multi-patient display was not excluded. FDA added that it intended to focus its oversight on alarms meant to prompt immediate clinical action.
The final guidance, Software as a Medical Device: Clinical Evaluation, had been developed through an international forum of regulators, the IMDRF. The notice said it set out harmonized principles on when and what type of clinical evaluation is appropriate, based on the risk of the software. Those principles covered scientific validity, clinical performance and analytical validity.
The same issue of the Federal Register announced a public workshop on the agency’s software precertification pilot, set for January 30 and 31, 2018. Comments on the decision support proposal were requested by February 6, 2018, under docket FDA-2017-D-6569. Both proposals carried the standard statement that guidance documents describe the agency’s current thinking and do not create legally enforceable obligations.
Clinical decision support runs through a modern health system. It fires drug interaction alerts in the order entry screen, suggests a test before a new prescription, scores a patient’s risk and ranks options for a care team. In December 2017, FDA said which of those functions it would treat as medical devices. The statute had already fixed the frame. The draft guidance decided how the frame would be read.
Our argument is that the draft put the regulatory line at a property of the software that a buyer can inspect: whether a clinician can see and check the basis for what it recommends. In our reading, that made explainability a compliance attribute. A rule engine that cites a guideline sits on one side of the line. A model that gives the same answer without showing its reasoning sits on the other, even if the clinical stakes are identical.
That had consequences beyond the companies FDA regulates. Health systems build their own decision support inside their electronic health records. They buy apps and analytics that make recommendations. Their clinical informatics teams choose how an alert is worded and what it shows. Each of those choices bears on the fourth statutory condition, as FDA read it.
We take the two drafts in four parts. Together they cover what a function reads, whom it addresses, how it explains itself, and the software around it that the Cures Act took out of FDA’s reach:
- Inputs: the line between information and signals, and why images and waveforms stayed regulated.
- Audience: the split between professionals and patients, and the enforcement policy FDA proposed for the second group.
- Basis: the fourth condition, and FDA’s reading of what it means for a clinician to review a recommendation independently.
- Plumbing: the records, data transfer and administrative software the companion draft placed outside the device definition.
After the four parts we look at how the drafts applied to different health organizations, take the main objections in order, and list what a leader could check. A note on status: both 2017 documents were drafts. They did not bind FDA or anyone else. We describe what they said and give our own reading of what followed from them for the organizations that use the software.
Inputs: information stayed outside, signals stayed in
The first statutory condition is a carve-out from the carve-out. Software intended to acquire, process or analyze a medical image or a diagnostic or physiological signal cannot qualify for the decision support exclusion. The draft explained why. Products that acquire and analyze physiological signals had long been regulated as devices, and the fidelity with which a signal is captured and processed matters to performance.
What counted as a signal
The draft gave a broad reading. Physiological signals were those that need an in vitro diagnostic device or a signal acquisition system to obtain. A signal acquisition system was the electronic circuitry and processor that receives inputs from sensors inside, attached to or outside the body, or from a sample taken from it. The draft’s examples covered electrocardiograph and electroencephalograph machines, imaging technologies and digital pathology.
It also named the kinds of processing that stayed regulated. Algorithms that derive new data points from physiological data, such as ST-segment measurements from an ECG, were one kind. Algorithms that identify features within an image were another. Software that analyzes genomic data to estimate a patient’s risk of disease was a third.
What counted as information
On the other side was medical information, the second condition. The draft read it to include patient-specific information such as demographics, symptoms and test results. It also included reference information such as practice guidelines, peer-reviewed studies, textbooks, approved drug labeling and government recommendations. FDA described this as the kind of information professionals use to make decisions for an individual patient.
The distinction is subtle. A laboratory result is information once it is reported. The analyzer that produced it is a device. Software that reads the reported value and compares it with a guideline threshold works on information. Software that reads the raw instrument output and interprets it works on a signal. The draft’s examples put a diabetes diagnosis aid on the information side, because a clinician enters glucose and hemoglobin A1c results and the software compares them with established criteria.
Why the input decided so much
For a health system, the input condition is the easiest to apply and the hardest to work around. An inventory of decision support can sort functions by what they read before anyone debates how they explain themselves. A sepsis alert that reads charted vital signs and lab values is reading information. A tool that reads the bedside monitor’s waveform is reading a signal.
That sorting matters for procurement. A function on the signal side is a device in FDA’s reading, whatever its vendor calls it. The buyer can ask for its clearance or approval status, and can expect labeling that states its intended use. A function on the information side may or may not be a device, and the other three conditions decide.
A design pressure the condition created
In our reading, the input condition gave developers a reason to design around signals. A function that consumes a value already charted by a clinician is in a different position from one that ingests the same measurement straight from the instrument. The clinical content may be identical. The regulatory status may differ.
That pressure is not necessarily bad. Charted values have passed through a professional’s hands. But it means the architecture of an integration can bear on the status of the function. Data may flow from the device to the EHR flowsheet and then to the rule, or straight to the analytic engine. In our view, few integration teams would have treated interface design as a regulatory decision.
Audience: professionals were excluded, patients got a policy
The third condition limits the statutory exclusion to recommendations made to a health care professional. The draft said plainly what followed. Software that supports or provides recommendations to patients, not professionals, is not excluded from the device definition.
The patient decision support category
FDA did not stop there. It said some decision support for patients and caregivers was low risk and outside the functions on which it intended to focus. It proposed to adopt an enforcement discretion policy for this patient decision support software that generally paralleled the professional exclusion.
The four factors mirrored the statute. The software would not read images or signals, would display or analyze medical information, would support recommendations to patients or lay caregivers, and would let them independently review the basis. The explanation would cover the same four elements: purpose, intended user, inputs and rationale. The sources would be identified, understandable and publicly available.
The draft added a point that matters for design. It said the kinds of explanation a professional can understand and apply differ from those a patient can, given the difference in clinical education and experience. So the same function could meet the review condition for a physician and fail it for a patient.
The legal difference between exclusion and discretion
For a professional-facing function, meeting the four conditions meant the software was not a device at all. For a patient-facing function, meeting the four factors meant it was still a device, but one against which FDA did not intend to enforce its requirements. In our reading, that is a significant difference for anyone relying on it.
An exclusion is set by statute. Enforcement discretion is an agency policy stated in guidance, and the agency can change it. The draft itself was not final, so the policy was a proposal on top of a proposal. A health system that offered patients a symptom checker or a medication guide in its portal was relying on that policy, not on the statute.
The examples on each side
The draft’s patient examples were modest. Software reminding a patient how or when to take a prescribed drug, consistent with labeling, fell within the policy. So did software helping a patient choose an over-the-counter cold or allergy medicine, with warnings about overlapping active ingredients. The warfarin dosing example fell outside it, because the software would recommend dose changes without the patient consulting a provider.
For a health system, these examples map onto real portal features. Medication reminders and refill prompts sit near the first example. Triage tools that tell a patient whether to seek care are a harder case, because the recommendation bears on diagnosis. The draft did not address triage directly. An organization offering one would have had to apply the factors itself.
When a provider tool reaches the patient
The audience line can move inside one product. A health system may build a tool for its clinicians and later decide to show the same output to patients in the portal. Under the draft, that change moves the function from the professional exclusion to the proposed patient policy. The explanation that satisfied a physician may not satisfy a patient, because the draft said the two audiences understand different kinds of explanation.
So a change of audience is a change that needs review. A governance process that reviews new tools but not new audiences for old tools would miss it.
Who in the organization owned the audience question
The audience question crosses several teams. Clinical informatics owns the provider-facing tools. Digital or patient experience teams own the portal and apps. Marketing may own content that recommends services. The draft’s distinction means each team needs to know which side of the line its tools are on, and who decides when a provider tool is extended to patients.
Basis: the clinician had to be able to check the recommendation
The fourth condition decided the hard cases on the information side, and the draft’s reading of it is the core of the document. The statute says the software must be intended to enable the professional to independently review the basis for its recommendations, so that the professional does not rely primarily on them. The draft turned that into a test of explanation.
Four things the software had to explain
FDA said it read the condition to describe software that clearly explains its purpose or intended use, its intended user, the inputs used to generate the recommendation and the rationale or support for it. The draft’s examples of inputs were simple, such as patient age and gender. Its examples of intended users were specific, such as ultrasound technicians or vascular surgeons.
Then came the decisive sentence. The intended user should be able to reach the same recommendation on his or her own without relying primarily on the software. The software, in this reading, was a convenience that saved the clinician time. It was not a source of judgment the clinician could not reproduce.
Public, understandable sources
The draft set three tests for the sources behind a recommendation. They had to be identified and easily accessible. They had to be understandable by the intended user, such as data points whose meaning the user understands well. And they had to be publicly available, such as clinical practice guidelines and published literature.
FDA spelled out the consequence. A practitioner could not independently evaluate the basis of a recommendation that rested on non-public information, or on information whose meaning the practitioner could not be expected to understand. Two of the draft’s device examples made the point. One used an algorithm undisclosed to the user to pick an antihypertensive drug class. Another used a proprietary algorithm to recommend radiation treatment where the basis was unavailable for review.
What that meant for models
The draft did not mention machine learning by name. But its tests had obvious implications for statistical and learned models. A model trained on a health system’s own data rests on non-public information. Its weights are not something a clinician can review. Its output may be a score whose meaning depends on the training population.
In our reading, the draft put such models on the device side of the line unless their basis could be explained in terms a clinician could check against public sources. A logistic score built from published coefficients might qualify. A model trained on proprietary data and shown only as a number would not. That reading did not depend on how accurate the model was.
Display as part of the test
The fourth condition made the screen part of the analysis. The same rule can be presented as a bare instruction or as a recommendation with its inputs, its source and a link to the guideline. The draft’s footnotes asked that software keep its reference information current, display its source prominently and give users a way to reach current information, such as a link to drug labeling.
So the clinical informatics team that builds an alert makes a regulatory choice when it decides what the alert shows. Hiding the rationale to reduce clutter is a usability decision with a compliance side. A health system that wanted its homegrown tools to stay outside the device definition had a reason to show its work on screen.
The record a health system would keep
If explainability decides status, the evidence of explainability is a record someone should keep. For a locally built rule, that record is short. It names the rule’s purpose, the clinicians it addresses, the data elements it reads, the guideline or study it relies on, and the version of that source. It shows what the clinician sees when the rule fires.
For a purchased function, the record is the vendor’s. A health system can ask for the intended use statement, the list of inputs, the sources and a sample of the screen. It can keep those with the contract. If a question about the product’s status ever arises, from a regulator, an auditor or a malpractice claim, that file is where the answer starts.
The record also has to stay current. The draft asked software to keep its reference information up to date. A rule built on a guideline that has since been revised is still transparent, but its source is stale. A review cycle that checks each rule’s source against the current edition is the operational form of that request.
The examples on the excluded side
The draft listed functions that met all four conditions. They included software that matches a diagnosis to current treatment guidelines and cites the source, and drug interaction and allergy alerts based on FDA-approved labeling. Software suggesting liver function tests before a statin was another. So were chemotherapy suggestions consistent with labeling, parenteral nutrition calculators, an anion gap report based on practice guidelines and a ventilator guideline suggestion based on blood gas readings.
Each of these rests on a rule a clinician could find and apply. That was the common thread. The draft’s device examples, by contrast, involved images, signals, undisclosed algorithms or recommendations a clinician could not independently derive.
Why explainability, and why it was a hard test
There is a case for the approach. A clinician who can see why the software recommends something can catch errors, and the professional’s judgment stays the safeguard. That is the logic of the statute’s wording, and the draft followed it.
The test was also demanding. Some questions in medicine have no single guideline answer. Weighing several inputs together is the reason to use software at all. A test that asks whether the user could reach the same answer alone favors simple rules over complex ones, whatever their relative safety. We return to this in the objections.
Plumbing: records, data and administration moved outside FDA’s reach
The companion draft dealt with the software around decision support. It interpreted the first four exclusions in section 520(o) and set out edits to existing guidance. Its effect, in our reading, was to remove a large part of a health system’s data infrastructure from any question of FDA device regulation, provided it did not interpret or analyze data.
Administrative software
The statute excluded software for administrative support of a health care facility, with a long list of examples. These included financial records, claims and billing, scheduling, business analytics, population health management and laboratory workflow. FDA said it had not historically considered the bulk of these to be devices. Its one concrete change was to remove a section on laboratory information management systems from its off-the-shelf software guidance, because such systems were now outside the device definition.
Wellness software
The healthy lifestyle exclusion covered software unrelated to diagnosing, curing, mitigating, preventing or treating a disease. FDA tied this to the first category of its existing general wellness guidance: products that encourage a general state of health or healthy activity. Software making claims about weight management, fitness, relaxation, stress, sleep and similar topics was not a device when not related to disease.
The second category of the wellness guidance was different. It covered products that link healthy lifestyle choices to reducing the risk or impact of chronic diseases. Because those claims relate to disease, FDA said they were not excluded. It said it would continue not to enforce its requirements for them where they presented low risk.
Electronic patient records
The records exclusion had three conditions. The records had to be created, stored, transferred or reviewed by health care professionals or people working under their supervision. They had to be part of health IT certified under the ONC program. And the function could not be intended to interpret or analyze records for diagnosis or treatment.
FDA added a policy for the second condition. It said it did not intend to enforce its requirements for record functions that lacked ONC certification but met the other two conditions. It also said software letting patients keep their own records, not intended for professional review, was a personal health record and not a device if not intended for diagnosis or treatment. And it flagged that an EHR could contain both excluded functions and device functions, which it would address in separate guidance on products with multiple functions.
Device data and images
The fourth exclusion covered software that transfers, stores, converts formats or displays laboratory and other device data, unless it interprets or analyzes them. FDA said medical device data systems, medical image storage devices and medical image communications devices were therefore not devices. It proposed withdrawing its guidance on premarket notifications for medical image management devices.
The draft drew one line carefully. Software that generates alarms or alerts, or prioritizes a multi-patient display, was not excluded, because those functions analyze or interpret device data. Its example was a function that prioritizes patients in an intensive care unit by clinical status, which it said interprets or analyzes device data. A dashboard that only displayed the same data in a fixed order would, in our reading, fall within the exclusion. The draft also said FDA intended to focus its oversight on alarms and prioritized displays meant to prompt immediate clinical action, not on low-risk notifications that need none.
Functions, not platforms
The draft made a structural point that applied across all four exclusions. Section 3060 created a function-specific definition, independent of the platform the software runs on. FDA proposed changing mobile application to software function, where appropriate, in its mobile guidance. For a health system, this means the analysis is done one function at a time. One EHR module can contain an excluded record function, an excluded display function and a regulated alarm.
Products with several functions
The statute anticipated mixed products. Section 520(o)(2) addresses a product with at least one device function and at least one function that is not a device. The draft said FDA would give recommendations on such products in a separate guidance. Until then, a health system looking at a large platform had to apply the exclusions function by function, without settled guidance on how FDA would treat the regulated parts of an otherwise excluded system. In our reading, that made the inventory approach more useful, not less: the unit of analysis was the function, whatever the product boundary.
What the carve-outs left in place
The exclusions removed functions from the device definition. They did not remove other obligations. A health system’s records and data systems still carry the duties that apply to them under other law. Other federal and state rules on health data, privacy and security still reach those systems, whatever their FDA status. Certified health IT remains subject to the conditions of the ONC certification program the records exclusion itself relies on.
The statute also kept two doors open. The draft’s footnotes noted that a function described in the first four exclusions would still be a device if FDA found it reasonably likely to have serious adverse health consequences, under conditions the statute set. And none of the exclusions applied to software that met the criteria for the highest risk class. So the carve-outs were wide, but not unconditional.
How the drafts applied to different health organizations
The two drafts addressed developers. But their consequences landed differently on the organizations that build, buy and use the software.
Health systems that build their own decision support
A large health system’s informatics team writes rules, order sets and alerts in its EHR. Some of those rules recommend actions to clinicians. Under the drafts, a rule that matches charted data to a public guideline and shows its source fits the excluded side. A homegrown risk model trained on the system’s own data, shown as a score, raises the questions the draft’s device examples raised.
Whether a health system that builds software for its own use is a device manufacturer in FDA’s eyes is a separate legal question the drafts did not answer. Organizations should check with counsel. What the drafts did change is the vocabulary of internal governance. A clinical decision support committee could now ask of each new tool whether it reads signals, whom it addresses and whether its basis can be reviewed.
Ambulatory networks and physician groups
An ambulatory network buys more than it builds. Its exposure runs through contracts. The drafts gave it questions to put to vendors: which exclusion the vendor relies on, how the product meets the review condition, and what happens if FDA’s reading changes. A practice using an app that recommends treatment from a proprietary score has an interest in knowing whether that app should have been cleared.
Academic medical centers and research programs
Academic centers develop predictive models, sometimes for publication and sometimes for clinical deployment. The drafts did not address research use. In our reading, the fourth condition starts to bite when a model moves from a study into clinical workflow and begins to recommend care for individual patients.
Digital front doors and patient tools
Organizations offering patient apps and portals faced the audience question. The professional exclusion did not help them. They relied on the proposed enforcement policy for patient decision support, on the wellness guidance, or on the records exclusion. Each had its own conditions, and the first was a draft.
Community and rural hospitals
A smaller hospital with a thin informatics staff builds little and configures a lot. Its decision support arrives as vendor content inside the EHR, switched on or off by a small team. For such an organization, the drafts mattered mainly through the vendor. The hospital could ask which content the vendor treated as excluded and why, and it could keep the vendor’s explanation of each function’s sources.
The configuration choices still belonged to the hospital. Turning off the display of a rule’s source to shorten an alert, or changing a threshold without updating the cited guideline, alters the facts the fourth condition turns on. A small team that knows this can avoid undoing a vendor’s careful design with a local setting.
Health plans and payer analytics
The statute’s list of administrative functions named analysis of historical claims data to predict future utilization or cost-effectiveness, determination of health benefit eligibility and population health management. Those are the core analytic functions of a health plan and of a provider organization that bears financial risk. The statute excluded them when intended for administrative support of a health care facility, and the companion draft said FDA had not historically considered most such functions to be devices.
The boundary to watch is the move from administration to care. A model that predicts which members will be costly is an administrative tool. The same model used to recommend a clinical intervention to a physician for an individual patient raises the decision support questions. The function, not the data or the owner, decides.
Life sciences companies
Drug and device companies that distribute decision support with their products had separate issues. The draft said it did not address requirements that may apply to software disseminated by or on behalf of a sponsor for use with its drugs or biologics, such as labeling or combination product rules. That boundary is narrow but real: a dosing calculator from a drug maker is not the same regulatory object as one built by a hospital.
Health IT vendors
EHR vendors and analytics companies were the direct audience. For them, the drafts changed design incentives. Showing the inputs, the guideline and the reasoning on screen was the route to staying outside the device definition. In our reading, that put a premium on transparency features that customers could see in demonstrations and that FDA could see in labeling.
Objections, in order
Five objections deserve an answer. We take them in the order they bear on the argument.
1. The drafts were non-binding, so they changed nothing
The drafts did not bind anyone, and both said so. But the statute had been enacted on December 13, 2016, and FDA’s reading of it shaped how the agency would act. A developer that ignored the reading took the risk that FDA would treat its product as an uncleared device. A health system buying the product inherited part of that risk. Non-binding guidance still sets expectations.
2. Explainability is a poor proxy for safety
This is the strongest objection. A transparent rule can be wrong and a well-validated model can be safe. A test based on whether a clinician could reproduce the answer may push developers toward simpler tools rather than better ones. We think the objection is right about the incentive. But the test came from the statute’s wording, which turns on independent review and primary reliance. FDA could interpret that wording. It could not replace it with a risk test.
3. Clinicians do not read the basis anyway
Busy clinicians may accept recommendations without opening the source. The statute’s phrase is about intent, not observed behavior: the software must be intended to enable independent review so that the clinician does not rely primarily on it. Still, the objection has force for design. Software that technically exposes its basis but buries it may meet the words and miss the purpose. That is why display choices mattered.
4. Health systems are users, not manufacturers
Health systems that buy software are not FDA’s regulated parties for that software. In our reading, that is true as a starting point. But the analysis in the drafts reaches health systems in three ways: through the tools they build themselves, through the contracts they sign and through the patient-facing tools they offer. The drafts gave each of those a set of questions to answer, whoever is formally responsible.
5. The line would move, so designing to it was wasted effort
FDA revised its reading more than once after 2017, as the hindsight section shows. A developer or health system that designed to the 2017 draft’s exact words would have redone some work. But the questions the draft posed did not change: what the software reads, whom it addresses and whether its basis can be reviewed. An organization that recorded those answers for each tool could apply each new reading quickly. One that had not would have to start from nothing each time.
What a leader could check
These checks follow from the four parts. They are our suggestions, not requirements.
- Does the organization keep an inventory of decision support functions, recording for each one what it reads, whom it addresses and what it cites?
- For each function that reads images, waveforms or instrument output, is its regulatory status known and recorded?
- Do provider-facing alerts show their inputs and the source of their recommendation, or only an instruction?
- Has someone decided which patient-facing tools rely on an FDA enforcement policy rather than a statutory exclusion?
- Do vendor contracts state the regulatory basis the vendor relies on and require notice if it changes?
The checks need owners. The inventory belongs to the clinical decision support committee or its equivalent, with clinical informatics doing the work. Status questions for purchased products belong to supply chain and legal. Display standards belong to informatics and the clinicians who use the tools. Patient-facing tools belong to the digital team, with legal review.
With hindsight
FDA did not finalize the 2017 decision support draft as written. On September 27, 2019, it published a revised draft, now titled simply Clinical Decision Support Software. The notice said the revision described a risk-based approach to oversight of decision support functions that remain devices, using risk categories from an International Medical Device Regulators Forum framework. The same issue announced the final version of the companion guidance on changes to existing software policies.
On September 28, 2022, FDA announced the final Clinical Decision Support Software guidance. The notice said the final version focused solely on the criteria for non-device decision support. In response to comments, it no longer contained the regulators forum risk categories. It added explanation of how a software function, regardless of its complexity, can be intended to let a professional independently review the basis for its recommendations. It also said FDA’s existing digital health policies continued to apply to software functions that meet the device definition, including those intended for patients or caregivers.
The guidance changed again in 2026. FDA issued a revised version on January 6, 2026, and a superseding version on January 29, 2026, under the same docket number as the 2017 draft. As checked on November 15, 2026, the January 29 document is the current guidance on FDA’s site.
The current guidance keeps the four statutory criteria and keeps the fourth as the hinge. Its reading of that criterion has moved, in our reading, from a test of public sources to a test of disclosure. The 2017 draft said sources should be publicly available, such as guidelines and published literature. The 2026 guidance says that, regardless of the software’s complexity and whether or not it is proprietary, the software or its labeling should describe its underlying sources in language the intended user can follow. Its examples of acceptable sources include information the developer has communicated to the user.
The 2026 guidance also asks for things the 2017 draft did not mention. It recommends a plain-language description of how the underlying algorithm was developed and validated. It recommends identifying the required inputs, how to obtain them and their data quality requirements. It says FDA considers the level of automation and whether the decision is time-critical. It explains this with automation bias, the tendency of people to over-rely on a suggestion from an automated system, which it says increases when urgent action leaves no time to consider other information.
On the audience question, the 2026 guidance does not contain the separate patient decision support policy the 2017 draft proposed. It says functions that support or provide recommendations to patients or caregivers meet the device definition, and points to FDA’s existing policies on device software, mobile applications, clinical evaluation, device data systems and general wellness. The clinical evaluation document it names carries the same title as the guidance FDA finalized in the week of December 4, 2017.
One more change bears on design. The 2026 guidance says that where only one recommendation is clinically appropriate and the function otherwise meets all four criteria, FDA intends to exercise enforcement discretion. It gives examples, along with sub-examples that would not qualify.
What held up from the 2017 drafts is the structure. The input line between information and signals remains. The professional and patient audiences are still treated differently. The fourth criterion, whether the professional can independently review the basis, still decides the hard cases. The companion draft’s carve-outs for administrative software, records and device data display were finalized in 2019.
What did not hold up is the specific reading of review. The demand that sources be public, which in our reading put trained models built on private data on the device side, gave way to a demand for plain-language disclosure of sources, development and validation. That shift matters for health systems. It means the documentation a vendor provides, and what the screen shows, now carry more of the weight that the public literature carried in 2017.
For a health system leader in 2026, the practical point from the December 2017 week is unchanged. The regulatory status of a decision support tool depends on what it reads, whom it addresses and how it explains itself. Those are properties an organization can inventory, ask vendors about and design for. The details of the test have changed more than once, so the inventory needs to be rechecked against the current guidance, not the one in force when a tool was built.
Spotted an error? Tell us and we will check it and correct the article where needed.