The FDA's Predetermined Change Control Plan for AI/ML Devices

A cleared medical device is cleared as a specific thing. Retrain the model behind it and you have, in principle, a different thing. The Predetermined Change Control Plan is the Food and Drug Administration’s answer to that: authorise the boundary in advance, and changes inside it do not require a new submission.

The problem it solves

Under the ordinary rules, a manufacturer must submit a new premarket notification when it makes a change to a legally marketed device that could significantly affect its safety or effectiveness — the test in the regulations at 21 CFR 807.81(a)(3) for a 510(k) device. The Electronic Code of Federal Regulations carries part 807. For a conventional device that is a workable rule, because changes are discrete and infrequent. For a device whose function is a learned model, it is a rule that fights the technology: any retraining on new data is a change, and the honest answer to “could it significantly affect performance” is usually that it could, because that is the point of retraining.

The result before the PCCP was that manufacturers either froze the model at clearance — which is what most did, and it is why the large majority of authorised AI-enabled devices on the market are locked rather than continuously learning — or filed repeated submissions for incremental updates. The PCCP moves the review forward: the agency assesses the change process at the original submission, and authorises it.

Where the authority comes from

The statutory basis is section 515C of the Federal Food, Drug, and Cosmetic Act, codified at 21 U.S.C. 360e-4, which was added by section 3308 of the Food and Drug Omnibus Reform Act of 2022. The United States Code is published by the Office of the Law Revision Counsel. That provision states that a change made in accordance with an approved or cleared predetermined change control plan does not require a new submission.

The operative detail is in guidance rather than in the statute. The FDA issued final guidance in December 2024, “Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions”, following a draft in April 2023. The FDA publishes it in its guidance database. Guidance is not binding on the agency or on manufacturers, and the document says so; in practice it is what reviewers work from and a submission that ignores it will be asked why.

The three parts of a PCCP

A PCCP is not a statement of intent. It is three artefacts that have to be internally consistent, and reviewers read them against each other.

Description of Modifications

A specific, enumerated list of the changes the manufacturer intends to make, with the reason for each. “We may retrain the model” is not a description of modifications. What is expected is closer to: periodic retraining on additional data from the same intended population and the same imaging modalities, to maintain or improve sensitivity, with the input specification unchanged. The guidance also asks whether each modification is to be implemented automatically or with human involvement, and whether it will be uniform across the installed base or vary by site.

Modification Protocol

The methods. This is the largest part and covers data management (where new data comes from, how it is labelled, how training, tuning and test sets are kept separate), re-training practices, performance evaluation including the acceptance criteria a modified model must meet before release, and update procedures including how the change reaches devices in the field and how it can be rolled back. The acceptance criteria are the load-bearing part: they are the thing that makes the plan reviewable, because they let the agency assess in advance what “the update was acceptable” will mean.

Impact Assessment

The argument that ties the two together: for each modification, what the benefits and risks are, how the risks are mitigated by the protocol, and how the modifications interact with each other. The interaction point is easy to underweight — two changes each individually inside the plan can compose into something the plan did not contemplate, and the impact assessment is where that is supposed to be caught.

Where the boundary binds

The authorised PCCP becomes part of what was cleared or approved. Three things follow, and they are the ones manufacturers get wrong.

• The intended use cannot move. A PCCP does not authorise changes to the device’s indications for use. Adding a patient population, a new anatomy, or a new claim is a new submission, however incrementally the model got there.
• The protocol is a commitment, not a menu. Once authorised, deviating from the modification protocol is a deviation from the cleared device. If your retraining pipeline changes its labelling procedure, you have departed from the plan even if the resulting model performs better.
• Transparency obligations attach. The guidance expects the labelling to describe the PCCP so that users know the device may change and how they will be informed. A change made silently within an authorised plan is still a change a clinician relying on the device is entitled to know about.

It is also worth being precise about what a PCCP is not: it is not authorisation for a continuously learning device that adapts on each site’s local data without control. Nothing in the mechanism forbids designing such a device, but the modification protocol would have to specify and bound the local adaptation, and the acceptance criteria would have to be evaluable without the manufacturer seeing the data. That is a hard submission to write and there are very few of them.

What falls outside the plan

The residual category is where regulatory risk concentrates. A change that is not described in the Description of Modifications is outside the plan, and the ordinary analysis applies: if it could significantly affect safety or effectiveness, it needs a new submission. Common examples are a change of base model or architecture, a change to the input specification such as accepting a new scanner or a new file format, a change to the output such as adding a confidence score that clinicians will act on, and a change in the deployment environment that alters performance.

There is also an unresolved question worth naming rather than papering over. Devices increasingly incorporate general-purpose foundation models the manufacturer does not control and cannot version-pin indefinitely. A silent upstream model update is a change to the device that the manufacturer did not make and may not immediately detect. The PCCP framework was designed around a manufacturer that controls its own training pipeline, and how it applies to a hosted third-party model is not settled by the December 2024 guidance. Manufacturers in that position generally address it by pinning versions contractually and treating an upstream change as a change requiring their own evaluation, but that is a practice rather than a rule.

For the pathway a PCCP is submitted within, see AI-enabled devices and the 510(k) pathway; for the programme that promised this mechanism, see the FDA’s AI/ML SaMD action plan.