Pre-Cert tried to regulate the software maker, not each version

The Food and Drug Administration’s January 2019 Pre-Cert working model proposed trusting appraised software makers and their real-world data in place of some product review. The law kept the product as the unit, so health systems still own product evidence and local monitoring.

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

The Food and Drug Administration’s January 2019 Pre-Cert working model proposed trusting appraised software makers and their real-world data in place of some product review. The law kept the product as the unit, so health systems still own product evidence and local monitoring.

  1. The week On January 7, 2019, the Food and Drug Administration’s device center published version 1.0 of “Developing a Software Precertification Program: A Working Model.”
  2. The unit of trust moves from the version to the maker The working model’s introduction states the problem plainly.
  3. What a software maker would have had to show The Excellence Appraisal was meant to be demanding in a specific way.
  4. Where real-world evidence comes from The Real-World Performance component is the part of the model that reaches into hospitals and clinics.
  5. The legal ceiling the framework already named The January 2019 package was candid about the law.
  6. What a health system’s compliance office can use Pre-Cert did not become law, but the working model remains a detailed public statement of what a regulator thought good software makers do.
  7. Objections and limits The argument has weak points, and a skeptical reader will find them.
  8. With hindsight The pilot ran on the schedule FDA’s program page records.

The week

On January 7, 2019, the Food and Drug Administration’s device center published version 1.0 of “Developing a Software Precertification Program: A Working Model.” The release notes inside the document date the version “20190107 v1.0.” The working model says the agency announced the Software Precertification (Pre-Cert) Pilot Program in July 2017. The same release notes list two earlier versions, dated 20180426 (v0.1) and 20180619 (v0.2). Version 1.0 was the one the agency set out to test.

Two companion documents came with it: a “2019 Test Plan” and a paper titled “Regulatory Framework for Conducting the Pilot Program within Current Authorities.” Both are marked January 2019. The working model said its approach “may require additional statutory authority to implement fully,” and that the agency intended to run the pilot under the authorities it already had. The agency asked for public feedback on version 1.0 by March 8, 2019, through docket FDA-2017-N-4301.

The model described a voluntary pathway. Its stated goal was more streamlined and efficient oversight of software-based medical devices from manufacturers that had shown a culture of quality and organizational excellence and were committed to monitoring real-world performance. For version 1.0 it was limited to software as a medical device (SaMD), defined as software intended for one or more medical purposes that performs them without being part of a hardware medical device. Software functions excluded from the device definition by the 21st Century Cures Act, such as administrative support for a health care facility and certain electronic patient records, stayed outside the pilot.

The model had four components. An Excellence Appraisal would evaluate the organization against five Excellence Principles: product quality, patient safety, clinical responsibility, cybersecurity responsibility and proactive culture. A Review Pathway Determination would use the risk categories of the International Medical Device Regulators Forum (IMDRF), which cross the state of the healthcare situation or condition with the significance of the software’s information to the healthcare decision. A Streamlined Review would cover the product-specific elements the appraisal did not. A Real-World Performance component would have the organization collect and analyze data on the product in use and give the regulator access to analytics.

The model sketched two levels of precertification for a future program. Level 1 was designed to let an organization with a limited track record market certain lower-risk software without review. Level 2 was designed to let an organization with a proven track record market certain lower and moderate risk software without review. Other software would still get a streamlined review. The agency said it did not intend to establish precertified companies, or levels, during the 2019 testing, and that it would be primarily responsible for all Excellence Appraisals that year, while it considered accredited third-party appraisers for the future.

The Test Plan set out how the agency would check the idea. Internally, staff would take SaMD submissions that had already been reviewed, extract what a Streamlined Review package would contain, and test whether appraisal elements could answer the questions that remained. With volunteers, sponsors would file traditional De Novo requests or 510(k) submissions; CDRH staff would assemble a mock Streamlined Review package, and an independent review team would judge whether it was enough. The official regulatory decision would rest on the traditional submission.

The Regulatory Framework explained the legal route. The agency would use the De Novo classification process under section 513(f)(2) of the Federal Food, Drug, and Cosmetic Act. Appraisal results would be documented in a device master file. If a device was classified into class II, its special controls might include Excellence Appraisal elements or postmarket data collection requirements, and subsequent “Pre-Cert 510(k)” submissions from appraised sponsors would draw on that file.

Pre-Cert is easy to file away as a fast lane for technology companies that never opened. That reading misses what the January 2019 working model proposed. It proposed to change the unit of regulation. Under the device law, FDA reviews a product, and a significant change to that product can need a new review. The working model proposed that FDA would also judge the organization that makes the software, once, and then rely on that judgment and on data about the software in use to review each product more lightly, or in some cases not at all.

Our argument is that the unit is what mattered, and that the unit did not move. FDA’s own pilot concluded that its approach was not practical to implement under existing law and that an appraisal of the maker could not stand in for device-specific clinical and cybersecurity review across the board. The authority Congress did add was tied to a device and to a plan for that device’s changes. For a health system, that leaves two jobs where they were: the evidence for each product it deploys, and the monitoring of that product in its own setting. What Pre-Cert does leave behind is a well-specified set of questions about software makers that a health system can put to its vendors.

The analysis has six parts:

  • The unit of trust: what the working model moved from the product version to the organization, and what it left with the product.
  • What a maker would have had to show: the Excellence Principles, the performance indicators, the levels and the triggers for a fresh appraisal.
  • Where real-world evidence comes from: the analytics the model asked for, and how much of that data originates in clinical settings.
  • The legal ceiling: what the companion Regulatory Framework already said about the limits of current authority.
  • What a compliance office can use: the parts of the model that translate into a health system’s own controls.
  • Objections and limits: where the argument is weakest.

The opening four parts read the January 2019 documents. The fifth turns them into practice for a provider organization, and the sixth tests the argument.

The unit of trust moves from the version to the maker

The working model’s introduction states the problem plainly. It says FDA’s traditional approach for hardware-based devices is not well suited to the faster, iterative design, development and validation used for software. It contrasts hardware makers, who modify their products every few months to years, with software developers, who modify theirs in response to real-world performance and user feedback every few weeks to months. It also says that evaluating code alone may not give a full understanding of a SaMD product’s safety and effectiveness, partly because its impact on patients can be indirect.

What would be judged once

The model’s answer was to evaluate certain things at the level of the organization. The model says FDA proposed that certain elements traditionally reviewed in a premarket submission could be evaluated at the organization level during the Excellence Appraisal, and at the product level during Review Pathway Determination and Streamlined Review. The Test Plan put the same idea more concretely. Elements that are reviewed product by product but applied systematically across all of a company’s SaMD products could be evaluated once, at the organization level. Elements that vary by product would stay in the product’s submission.

That split is the heart of the proposal. A software maker’s design controls, testing practices, risk management process and cybersecurity process do not change from one product to the next. Under the product model, FDA sees evidence of them again in each submission. Under the Pre-Cert model, FDA would see them once in depth and then check, through performance data, that they kept working.

What would stay with the product

The model did not drop the product. The Review Pathway Determination used the IMDRF grid, which places a product in one of four categories according to how serious the patient’s condition is and how much the software’s output drives the healthcare decision. A product that treats or diagnoses a critical condition sits in the highest category. A product that informs clinical management of a non-serious condition sits in the lowest. The organization’s precertification level and the product’s category together would decide whether a product needed review and how much.

The model also kept the statutory standard. It said software products from precertified companies must continue to meet the same statutory standards as products that took the traditional path. Precertification was described as a pathway or method to access and evaluate the necessary information at different points in a product’s lifecycle. The safety and effectiveness standard was not lowered. The timing and the source of the evidence were what changed.

What a buyer would have seen

For a hospital or health system buying software, the change would have shown up in the signals it relies on. Under the product model, a buyer can look for a product’s clearance, approval or De Novo classification, each tied to a described device. Under the Pre-Cert model, a buyer of a lower-risk product from a precertified maker might have seen no product review at all. What it would have seen instead was the maker’s precertification status and a set of product-level elements.

The working model listed those product-level elements. They included the significance of the information the software provides to the healthcare decision, the state of the condition, the core functionality, a device description, and a description of the software’s performance. The device description could include how the software works, its significant security, technical and safety risks, supporting platforms and compatibility, instructions and limitations for use, inputs and customer support. The model said FDA intended, in a future state, to post these elements on its website. For the pilot, it said that at this time it intended to post them on its website.

That is a different kind of evidence for a buyer. A clearance tells the buyer that FDA reviewed a described product. A precertification status would have told the buyer that FDA had appraised the maker and expected it to monitor the product in use. A buyer relying on the second kind would need to know more about the maker and about the monitoring, because those were now carrying the weight.

What a software maker would have had to show

The Excellence Appraisal was meant to be demanding in a specific way. The model said it would let organizations demonstrate excellence based on outcomes achieved by their own processes, and that it would apply a least burdensome approach by observing the organization’s current processes. It named existing standards it would recognize, including the Quality System Regulation and ISO 13485, IEC 62304, ISO 14971, ISO 9001 and ISO/IEC/IEEE 12207, and the outcomes achieved by following them.

Five principles, measured

The five Excellence Principles each had a short definition in the model. Product quality covered development, testing and maintenance. Patient safety covered a safe patient experience and patient safety as a factor in decisions. Clinical responsibility covered clinical evaluation and patient-centric issues, including labeling and human factors. Cybersecurity responsibility covered protecting cybersecurity and addressing issues through active engagement with stakeholders and peers. Proactive culture covered surveillance, assessment of user needs and continuous learning.

The model tied the principles to measures. One example in the model concerned how a company reviews software changes. It listed key performance indicators such as traceability audit results, the number of passed and failed regression tests post-release, the number of risks, how frequently the risk analysis is updated, the number of design reviews and the number of critical bug reports submitted by customers. It said these indicators are reviewed during release cycles and expected to trend down, and that no change or an increase indicates that errors are being introduced in updates.

Those are not regulatory abstractions. They are numbers a well-run software team already has, or can produce. The model’s stated goal was to develop a library of activities, processes and indicators that high-performing organizations use. The model said it would give an organization flexibility to show how its own processes and measures of performance track to the program’s elements and, in the end, to the Excellence Principles.

Levels and maintenance

The two proposed levels separated organizations by track record. Both required objectively demonstrated excellence in all five principles. Level 1 was for an organization with a limited track record in developing, delivering and maintaining products. Level 2 was for one with a proven track record. The model presented the levels for exploration only, and said FDA would not establish them during the 2019 testing.

The model was explicit that precertification would not last indefinitely. In its response to public comments, FDA wrote that the program was not intended to be an indefinite designation, and that organizations would commit to share key performance metrics with the agency as part of maintaining it. Section 3.6 listed examples of events that might trigger an additional appraisal:

  • Significant recurring patient and product issues.
  • Significant restructuring of the organization that affects leadership, team and unit activities.
  • A merger or acquisition that affects the evaluated quality system and processes.
  • Continuous improvement efforts that may change performance capability.
  • Performance in new activities or a new clinical domain not part of the original assessment.
  • Continued recurrence of a safety signal observed in real-world performance monitoring.

This list is the part of the model with the longest shelf life for a buyer. It names the moments when trust in a software maker should be re-earned. A health system that relies on a vendor’s processes, rather than on a single reviewed version, has the same exposure FDA described. When the vendor is acquired, restructures its engineering team or moves into a clinical area it has not served before, the processes the buyer evaluated may no longer be the ones producing the software.

Consequences of a failed review

The model also said what would happen when a precertified maker’s product did not pass. During 2019 testing, if FDA did not authorize a product, the organization and FDA would complete an after-action review to find gaps in the evidence and plan a future submission. The model said FDA expected to implement a process in which repeated unsuccessful streamlined reviews would trigger a reassessment of the organization’s precertification. The status, in other terms, was meant to respond to evidence about products. A provider can apply the same logic to a vendor: a run of failed local validations, unexplained defects or poorly documented updates is information about the vendor’s processes, and it can feed the vendor’s next review.

Who would do the appraising

The model said FDA would be primarily responsible for all appraisals during 2019, while its vision for the future included accredited third parties. It said appraisal information collected by third parties would be used in FDA’s decisions, similar to a conformity assessment against a recognized consensus standard. That matters for how a buyer would have read a precertification. An appraisal by FDA staff and an appraisal by an accredited third party are both forms of assurance, but a buyer would want to know which stood behind a status, and when it was done.

Where real-world evidence comes from

The Real-World Performance component is the part of the model that reaches into hospitals and clinics. The model defined real-world performance analytics (RWPA) as systematic computational analyses of data relevant to the safety, effectiveness and performance of a SaMD product in real-world settings, marketed by a precertified organization. It named three types.

Three kinds of analytics

Real-world health analytics were analyses of clinical outputs and outcomes related to the product’s intended use. The model said they could inform changes to intended use, support expanded functions and populations, and identify emerging safety issues. For lower-risk products, the sources could include user complaints, search analytics and monitoring of use errors. For higher-risk products, the model said precertified organizations might seek external sources through activities such as registries, partnerships with healthcare systems, or data commons and other structured postmarket collection.

User experience analytics were analyses of how users experience the product in real use. The model mentioned use patterns, download rates, user retention, satisfaction and complaints, and said excellent organizations would seek feedback from diverse sources.

Product performance analytics were analyses of the product’s real-world accuracy, reliability and security. The model expected organizations to track reliability and availability, and, for connected products, to monitor and address security vulnerabilities and threats. It pointed to FDA’s guidance on postmarket management of cybersecurity in medical devices, including participation in Information Sharing and Analysis Organizations.

The examples point to the clinic

The model’s examples show where the data would come from. In one, a precertified organization markets a moderate-risk computer-aided detection product designed to capture both the software’s output and the radiologist’s final read. The manufacturer monitors false positives, defined as cases where the software detects disease and the radiologist disagrees. When the trend worsens, it investigates, finds a software defect and issues an update, with FDA marketing authorization if appropriate.

That example only works if the radiologist’s final read flows back to the manufacturer. The read is made by a clinician, in a health system’s workflow, on the health system’s systems. The model treated it as a data element the manufacturer would analyze. It did not say how the data would travel or under what terms.

In another example, a security researcher tells a precertified organization about an open communication port in a marketed product. The organization judges the residual risk of patient harm acceptable, treats the fix as a routine cybersecurity update, pushes a validated patch remotely and communicates with end users. It then looks for other open ports, changes requirements, trains developers and adds test cases. The model presented this as behavior that supports maintenance of precertification.

From the health system’s side, that remote patch lands in its own environment. Its clinical engineering, information security and change management teams are the end users the model mentions. A program that rewards fast, remote fixes also creates a stream of changes the provider has to absorb, record and, where needed, revalidate.

What the plan would contain

The model said pilot participants would develop an RWPA plan before introducing a product to market. For all products, a plan should include the proposed data elements, the intended frequency of collection, the intended data structure and format, and commit and stretch goals for each element. FDA said it intended to focus on trends and summary analytics rather than raw data, and a footnote said existing requirements, including adverse event reporting, would continue to apply. The model also said the RWPA framework would not apply to SaMD products cleared before the organization was precertified.

Two points follow for providers. One is that the model treated real-world performance as the maker’s obligation and the maker’s data. The other is that much of the underlying signal, from clinician overrides to outcomes to network behavior, originates with the provider. Under any model, a health system that wants to know how a product performs in its own population cannot assume that the maker’s summary analytics answer that question. A trend computed across all customers can look healthy while performance at one site drifts.

Signals that cross products

The model also described a use of analytics that no single maker could provide. In one example, a precertified organization’s product analyzes a signal from a third-party wearable sensor that is not itself a medical device. Through its analytics, the organization notices more user complaints about performance and traces the cause to a change in the sensor’s sampling rate, which degraded the signal the algorithm relied on. The maker adjusts its algorithm to keep its cleared performance, and the model says FDA would be able to notify makers of other products relying on that sensor that additional hardware testing may be needed.

The example names a dependency that a provider’s inventory can miss. Software that is a device can rely on hardware, platforms or data feeds that are not devices and that change on their own schedules. The model treated those dependencies as part of the maker’s monitoring. A provider that runs the software also runs, buys or supports some of those dependencies, and a change on its side, such as a new device model, an operating system update or a different imaging protocol, can alter the software’s inputs. Recording those dependencies next to each clinical product gives the provider a way to ask the maker, before a local change, whether it falls within what the product was validated for.

Data that flows back

If a health system agrees to send clinical data to a software maker so that the maker can compute such analytics, the arrangement raises its own compliance questions. What data leaves, in what form, under what agreement, and for what permitted uses are questions for the privacy officer and counsel, and the answers depend on the facts. The working model was silent on them because it addressed the maker’s relationship with FDA. A provider has to address them in its relationship with the maker.

The January 2019 package was candid about the law. The working model said the approach may require additional statutory authority to implement fully. The Test Plan said the model’s section 5 proposed a future state, envisioning additional legislative authority, in which lower-risk products from precertified manufacturers could enter the market directly. The Regulatory Framework said the pilot would also inform FDA whether it should seek other regulatory authorities.

De Novo as the vehicle

The Regulatory Framework chose the De Novo classification process as the vehicle. De Novo is the route by which a new type of low to moderate risk device can be classified into class I or class II rather than remaining in class III, which would require premarket approval. Under the framework, an appraised sponsor’s “Pre-Cert De Novo Request” would contain the elements not already received through the appraisal and documented in the device master file, plus the sponsor’s certification that it was following the processes FDA had reviewed.

The framework said that if FDA classified the device into class II, it would establish special controls. For organizations that opted to pursue certification, these could include certain appraisal elements, or requirements to notify FDA if appraisal elements changed. The special controls could also include postmarket data collection requirements. And certain special controls, such as appraisal elements or real-world performance requirements, might permit certain modifications without premarket review where they effectively mitigated the risks of those modifications.

Controls that bind a class

A De Novo classification creates a device type, not a company status. The framework said so in its own terms. The De Novo order would establish a classification that enables 510(k) submissions for devices of the same type. Sponsors that had not been appraised could market a device in that class through a traditional 510(k), including a demonstration that they met all applicable special controls. The effect is that the controls written for a pilot participant’s device would become the controls for its category.

That is the legal ceiling in miniature. The statute FDA was working within sorts devices into classes and attaches controls to classes. A program built around trust in a particular organization had to express that trust through controls on a device type, which then reach every maker of that type. A reader in January 2019 had the text needed to see the tension, and the hindsight section sets out how FDA described it in 2022.

Modifications stayed under existing guidance

On changes to a device, the framework kept to existing rules. It said modifications would be assessed under 21 CFR 807.81(a)(3) for whether they could significantly affect safety or effectiveness or changed the intended use, and it cited FDA’s guidance on deciding when to submit a 510(k) for a software change to an existing device. If a device’s special controls effectively mitigated the risk of a modification that did not change intended use, the framework said a new 510(k) would not be required. Otherwise one would be, though an appraised sponsor could file a “Pre-Cert 510(k)” drawing on its documented appraisal.

The working model’s comment table made the same point. In response to commenters on how modifications should be handled, FDA said it would use the current software modifications guidance for SaMD modifications in the 2019 pilot. The pilot never had authority to free a precertified maker’s updates from the product-by-product test. It could only explore what such a system might look like.

What this meant for regulatory status

For a health system’s compliance office, the practical conclusion was available in 2019. Whatever the pilot learned, the regulatory status of a product the provider bought would still be a product’s status: a clearance, a De Novo classification, an approval, or a determination that the software was not a device or was under an enforcement policy. The Test Plan itself said the official regulatory decision for each test case would be based on the traditional submission. A precertification of the maker would not appear in a product’s legal status during the pilot.

What a health system’s compliance office can use

Pre-Cert did not become law, but the working model remains a detailed public statement of what a regulator thought good software makers do. That makes it useful as a template for a provider’s own controls, provided the provider is clear that these are its own choices and not requirements imposed by the program. The steps below are good practice as we see it, not obligations, and an organization should check with counsel where a contract, a privacy rule or a device regulation is involved.

Know which software is a device

The working model restated the categories of software functions excluded from the device definition under section 520(o) of the FD&C Act, as added by the 21st Century Cures Act. They include functions intended for administrative support of a health care facility, for maintaining or encouraging a healthy lifestyle, to serve as certain types of electronic patient records, for transferring, storing, converting formats or displaying data without interpreting or analyzing it, and to provide certain limited clinical decision support.

A provider’s software inventory can record, for each clinical product, which side of that line the maker says the product falls on, and on what basis: a clearance, a De Novo classification, an approval, an exclusion, or an enforcement policy. That record tells the compliance office which products come with FDA-reviewed claims and which do not. It also tells the clinical teams where the provider’s own validation carries the whole weight.

Ask vendors the appraisal questions

The five Excellence Principles give a procurement team a structure for vendor diligence that goes beyond a security questionnaire. A provider can ask how the vendor reviews software changes, what indicators it tracks on release quality, defects and rollbacks, how it handles customer-reported critical bugs, and how it monitors security vulnerabilities in fielded products. The model’s example indicators, such as regression test results and critical bug reports from customers, are things a mature vendor can describe without disclosing trade secrets.

The answers do not replace product evidence. They tell the provider how much to trust the vendor’s change process between the provider’s own checks. A vendor that cannot describe its release indicators is asking the provider to accept each new version on faith.

Write the triggers into the relationship

The model’s triggers for a new appraisal translate directly into contract and vendor-management terms. A provider can ask to be told of a merger or acquisition affecting the vendor’s quality system, a significant restructuring of the team that builds the product, a move into a new clinical domain, and recurring safety or product issues. On notice, the provider can reassess the vendor. Without notice, the provider learns of these changes from the news or not at all.

Agree what real-world data goes where

If a product’s monitoring depends on clinical data the provider generates, such as clinician overrides, final reads or outcomes, the provider can agree in advance what data it will share, in what form and under what terms, and what analytics it will receive back. A provider that receives site-level performance, in addition to a vendor-wide trend, can compare the product’s performance in its own population with the claims it relied on at purchase. Whether sharing that data needs a business associate agreement or other terms depends on the facts and is a question for the privacy officer and counsel.

Treat vendor patches as changes

The working model’s security example ended with a patch pushed remotely. Remote updates can close vulnerabilities fast, and a provider has reason to want that. They are also changes to clinical software in the provider’s environment. A provider can require that each update arrive with a version identifier and a description of what changed, can record updates in its own change log, and can decide in advance which kinds of change trigger local revalidation. The hindsight section notes that FDA’s 2026 classification order for one type of machine learning-based imaging software made a version history and a description of how users are informed part of the required labeling for devices of that type, each of which has a change control plan.

Keep local monitoring

The model’s analytics were the maker’s. A provider’s own monitoring of the products it runs, including override rates, downtime, complaints from clinicians and, where relevant, outcomes in its own population, stays with the provider under any regulatory model. The working model described real-world performance as a way for FDA to keep confidence in a maker. A provider needs the same confidence about a product at its own site, and the provider is the party positioned to measure that directly.

Objections and limits

The argument has weak points, and a skeptical reader will find them.

A pilot that ended teaches little

One objection is that Pre-Cert was a pilot, never a regulation, and that studying it is history rather than compliance. The answer is that the program’s questions outlived it. The working model set out, in public, what FDA thought distinguished software makers it could trust, what data it would want from them, and when it would look again. Those are the same questions a provider faces when it buys software that changes. The law did not adopt FDA’s answer, but the questions did not go away.

A provider cannot appraise a vendor the way a regulator can

A second objection is bargaining power. FDA proposed appraisals by its own staff or accredited third parties, and ongoing access to analytics. A single hospital buying a niche product has none of that. The objection is fair and argues for proportion. A provider can scale its questions to the product’s risk, using the IMDRF logic the model used: more for software that drives or makes treatment and diagnosis decisions in serious or critical conditions, less for software that informs management of non-serious ones.

Organizational trust was a sound idea

A third objection is that judging the maker is a sound idea and that the law, not the idea, was the problem. That is partly right, and FDA’s 2022 report, discussed in the hindsight section, said organization-level information about development processes can be considered across multiple devices. But the same report found that organizational appraisals would not be sufficient to take the place of device-specific clinical performance and cybersecurity reviews for all moderate-risk devices. That finding concerns the evidence, not the statute. For a provider, it supports the position here: organizational diligence informs how much to trust the change process, and product evidence still decides whether to deploy.

Where the argument is weakest

The weakest point is that we cannot show what the pilot changed in vendor behavior or provider practice. The public documents describe FDA’s plans and findings, not the practices of the participants or their customers in detail, and we have no data on how health systems responded. The argument rests on the documents’ logic and on the regulatory record that followed, not on observed outcomes. A provider that adopts these controls should measure them on its own terms.

The January 2019 working model asked a regulator to trust an organization and verify that trust with data from the field. The law kept asking about the product, and the authority that followed was attached to a device and its planned changes. For a health system, the useful inheritance is the questions FDA wrote down about software makers, put to vendors as diligence, alongside product evidence and the provider’s own monitoring, which no maker’s status can replace.

With hindsight

The pilot ran on the schedule FDA’s program page records. FDA announced an opportunity for volunteers to take part in the Test Plan on May 22, 2019, published a mid-year update on July 18, 2019, and published a 2020 update on September 14, 2020. On September 26, 2022, it issued “The Software Precertification (Pre-Cert) Pilot Program: Tailored Total Product Lifecycle Approaches and Key Findings.” The program page says the pilot was completed with the issuance of that report. As checked on November 15, 2026, the page carries a “Content current as of” date of 09/26/2022.

The report was direct about the legal result. FDA said it encountered challenges implementing the approach under its current statutory authorities, and it described them. Participation was limited to 9 pilot participants; a footnote points to the Paperwork Reduction Act rules, under which a collection of information imposed on ten or more persons generally needs Office of Management and Budget approval. Together with the need to limit formal implementation to De Novo classification, that left few devices available for consideration under the pilot. For devices where De Novo was appropriate, pilot-specific special controls would have applied to every device that could show substantial equivalence to the participant’s device, whoever developed it. FDA also could not require participants to provide information beyond what the statute already required. The report said participants nonetheless provided significant additional information voluntarily, but that it was challenging to harmonize across participants.

FDA concluded that the approach in the working model “is not practical to implement under our current statutory and regulatory authorities.” It said rapidly evolving technologies could benefit from a new regulatory paradigm, which would require a legislative change, and that new legislative authority could be supplemental to, and not replace, the established pathways.

The report’s findings on substance matter as much for providers. FDA found that organization-level information about product development processes can be considered in a cross-cutting way for multiple devices. It also found that further development was needed before identifying low-risk devices where an organizational appraisal alone could be relied on, and that organizational appraisals would not be sufficient to take the place of device-specific clinical performance reviews and cybersecurity reviews for all moderate-risk devices. It said early appraisals, run without pre-specified criteria, produced outputs less specific than regulatory decisions would need. From its research and the pilot, FDA identified 250 KPIs and metrics and grouped commonly used ones into 9 categories: complaints, data quality, defects, device activations or user adherence, regression testing, releases, rollbacks, services and security.

The authority Congress did add was narrower and attached to devices. On December 29, 2022, section 3308 of the Food and Drug Omnibus Reform Act of 2022, part of the Consolidated Appropriations Act, 2023, added section 515C, “Predetermined Change Control Plans for Devices,” to the FD&C Act, with provisions for devices requiring premarket approval or 510(k) clearance. FDA’s notice of its final guidance on change control plans for AI-enabled device software functions appeared in the Federal Register on December 4, 2024, following a draft noticed on April 3, 2023. The guidance recommends that a plan describe the planned modifications, the methodology to develop, validate and implement them, and an assessment of their impact. FDA reviews the plan as part of a device’s marketing submission, so that each modification it describes does not need its own submission.

The De Novo route also produced a device-level form of the idea. On June 17, 2026, FDA published an order classifying radiological machine learning-based quantitative imaging software with a predetermined change control plan into class II. The order says FDA received the De Novo request on September 28, 2022, and that the classification was applicable on February 24, 2023. Its special controls require documentation of planned modifications and the methods for developing, verifying and validating them, an assessment of their risks, and labeling that includes a statement that the device has a change control plan, a version history and a description of how users will be informed of modifications.

What held up is the diagnosis in the 2019 documents: software changes faster than a product-by-product review was built for, and data from the field can support oversight. What did not hold up is the unit. The legal tools that followed attach a plan for change to a specific device, and FDA’s own pilot found that an appraisal of the maker could not stand in for device-specific clinical and cybersecurity review across the board. For a health system, the evidence for each product and the monitoring of that product at its own site remained its work.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • FDA
  • Pre-Cert
  • software as a medical device
  • real-world performance
  • vendor risk
  • change control