Regulation

Predetermined Change Control Plans in practice

How manufacturers actually build a PCCP and what health systems should check — the three required sections turned into what you write, the five guiding principles as design constraints, and what the first authorized plans reveal. As of July 2026.

By Jonas WeirReviewed by Jonas Weir · editorial reviewUpdated

The short version

  • A PCCP has three sections the FDA reviews together: a Description of Modifications (the specific changes), a Modification Protocol (how each is validated, with pre-defined acceptance criteria), and an Impact Assessment (the benefits and risks).
  • Its whole purpose is to let a manufacturer ship pre-specified updates without a fresh marketing submission for each one — bounded flexibility, agreed in advance.
  • The FDA, Health Canada, and the UK MHRA set five guiding principles for PCCPs: Focused and Bounded, Risk-based, Evidence-Based, Transparent, and a Total Product Lifecycle perspective.
  • The first authorized plans are narrow: across 26 devices with PCCPs through May 2025, the common changes were model retraining (69%), logic updates (42%), and adding input sources (35%); 92% were 510(k) clearances and all were moderate risk.
  • Preapproval and post-market evidence are thin — seven devices were prospectively evaluated, subgroup analyses appeared for eleven with no patient-outcome data, and no post-market studies were found. Verify a device's plan and monitoring with compliance before you rely on it.

A Predetermined Change Control Plan is the FDA's answer to a structural problem with AI-enabled devices: they are built to be updated, but the traditional device framework would make many updates each trigger a new submission. The PCCP glossary entry covers what the mechanism is and where it came from. This page is the working view — what actually goes into a plan, the principles that shape it, and what the first authorized plans tell us about how the tool is being used. As of July 2026.

What a PCCP authorizes

A PCCP is documentation the FDA reviews inside a device's marketing submission so that pre-specified updates can ship without a separate submission each time. The final guidance states the point directly: the plan lets a manufacturer implement the described modifications "without necessitating additional marketing submissions for implementing each modification described in the PCCP" 1. The final guidance was originally issued December 4, 2024, and the edition on the FDA's site is dated August 18, 2025 12. The mechanism's statutory footing traces to the Food and Drug Omnibus Reform Act of 2022, as the glossary entry sets out.

The critical word is bounded. A PCCP grants flexibility inside a fence the manufacturer draws and the FDA approves — the flexibility to make the changes in the plan, and no others.

The three sections, turned into what you write

The guidance recommends three sections, reviewed together 1:

PCCP sectionWhat it doesWhat the manufacturer actually writes
Description of ModificationsNames the specific planned changesThe exact modifications — e.g., periodic retraining on new data, a logic update, or adding a compatible input source
Modification ProtocolShows how each change is developed, validated, and implemented safelyVerification and validation activities with pre-defined acceptance criteria, plus a step-by-step account of implementation
Impact AssessmentWeighs the benefits and risks of the changesAn analysis of the benefits and risks of each modification and its mitigations, individually and in combination

The Modification Protocol is where most of the engineering rigor lives. The guidance describes it as including "the verification and validation activities (including pre-defined acceptance criteria) for those modifications" and providing "a step-by-step delineation of how the modifications included in the PCCP will be implemented while ensuring the device remains safe and effective" 1. In practice, this is the difference between "we may retrain the model" and "we will retrain on data meeting these inclusion rules, validate against these metrics at these thresholds, and roll back if the thresholds are missed." The acceptance criteria are the safety rail: they are set in advance, so a change either clears the pre-agreed bar or it does not ship.

A worked example

Consider the most common case, a periodic-retraining plan. The Description of Modifications names it: the model will be retrained quarterly on newly collected data from the same clinical settings, with no change to inputs, outputs, or intended use. The Modification Protocol then makes it auditable — it specifies the data-inclusion and quality rules, the held-out test set, the performance metrics and the exact acceptance thresholds each retrained version must meet, the human review step, and the rollback procedure if a version misses. The Impact Assessment reasons about what could go wrong: could retraining on a shifted population degrade performance for a subgroup, and what mitigations catch that before release? Written this way, a single reviewed plan can cover a year of updates — which is the entire point — while every update remains inside the fence the FDA approved 1.

The five guiding principles

Before the final US guidance, the FDA, Health Canada, and the UK's MHRA jointly set out five guiding principles for PCCPs, drawing on the Good Machine Learning Practice principle that deployed models are monitored and retraining risks managed 3. They read as design constraints for anyone drafting a plan:

Guiding principleWhat it asks of a plan
Focused and BoundedDescribe specific, limited changes — the plan is a fence, with no open-ended latitude
Risk-basedMatch the depth of validation to the risk each change introduces
Evidence-BasedJustify the plan with evidence generated across the total product lifecycle
TransparentCommunicate the plan clearly to users, patients, and reviewers
Total Product Lifecycle (TPLC) perspectivePlan pre-market for the changes and monitoring that follow post-market

These principles are why a credible PCCP is narrow and specific. A plan that asks for broad latitude fails the first principle; a plan without pre-set acceptance criteria fails the third.

The approach did not appear overnight. The idea traces back to a 2019 FDA discussion paper on modifications to AI/ML-based software, which drew substantial feedback and evolved into the current three-section structure of Description of Modifications, Modification Protocol, and Impact Assessment 1. The international alignment came next: the joint principles let a manufacturer design one plan against a shared logic recognised by the FDA, Health Canada, and the MHRA, rather than three divergent national expectations 3. For a global development pipeline funneling toward multiple regulators, that shared vocabulary is part of what makes the mechanism usable at all.

What the first authorized plans look like

The theory meets reality in the plans the FDA has actually authorized. A cross-sectional analysis identified "26 AI/ML-enabled medical devices with authorized PCCPs" cleared or approved before May 30, 2025, of which "92% were cleared via the 510(k) pathway, and all were classified as moderate risk" 4. The authorized changes were modest and concentrated:

Authorized modificationShare of the 26 devices
Model retraining69%
Logic updates42%
Expansion of input sources35%

The analysis found the plans "most commonly allowing model retraining (69% of devices), logic updates (42% of devices), and expansion of input sources (35% of devices)" 4. Their character is telling: the devices "were primarily intended for use in diagnosis or clinical assessment," six carried consumer-facing components, and thirteen underwent human-factors testing 4. All were moderate-risk 510(k) clearances rather than higher-risk approvals — the tool has so far been used where the FDA is most comfortable letting it run, which is worth remembering before assuming a PCCP signals a heavily scrutinised device.

This matters for how you read the word "flexibility": the authorized plans so far cover controlled, batched changes such as periodic retraining, rather than a model that rewrites itself continuously in the field. That discipline is deliberate — unmanaged updates are how a device slides into model drift.

Where the evidence is thin

The same analysis is candid about the gaps, and this is the part a buyer should read closely. Preapproval testing was limited: "seven devices prospectively evaluated and thirteen undergoing human factors testing." On fairness, "subgroup analyses were reported for eleven devices and none included patient outcomes data." On what happens after clearance, "no postmarket studies or recalls were identified," and "user manuals could be identified online for 54% of devices, though many lacked performance details or mentioned PCCPs" 4. The authors' conclusion is worth quoting: FDA authorization of PCCPs "grants manufacturers substantial flexibility to modify AI/ML-enabled devices postmarket, while preapproval testing and postmarket transparency are limited" 4.

The reading is not that PCCPs are unsound — it is that the plan on paper is only as good as the monitoring behind it, and the public evidence that monitoring is happening has been sparse.

A due-diligence checklist for health systems

Because much of the assurance lives in documents and monitoring rather than in a one-time clearance, a deploying health system carries real responsibility. Before relying on a "PCCP-enabled" device, work through this:

  1. Read the Description of Modifications. It tells you precisely what the vendor is permitted to change without re-review. If you cannot obtain it, you cannot know what you are buying a year from now.
  2. Ask for the acceptance criteria. The Modification Protocol's thresholds are the safety rail; ask what they are and what triggers a rollback 1.
  3. Ask for subgroup and outcome evidence. Given that subgroup analyses appeared for only eleven of the first authorized devices and none reported patient outcomes 4, ask specifically whether the device was evaluated on patients resembling yours.
  4. Confirm the intended use is unchanged. A change to intended use falls outside any PCCP and requires a new submission; make sure an update has not quietly moved the device beyond what you validated.
  5. Plan local monitoring. The plan presumes surveillance. Building your own algorithmovigilance around the device is how you catch a post-update regression the vendor's summary would not show you.

The boundaries of a plan

Three boundaries keep a PCCP honest.

First, it bounds change; it does not open it up. A PCCP authorizes only the modifications written into it and validated against its protocol. Anything outside the plan — above all a change to the device's intended use — still requires a new marketing submission 1.

Second, it presumes monitoring. The acceptance criteria mean something only if the manufacturer and the deploying system actually watch real-world performance. A plan is a promise about how change will be governed, rather than a substitute for governing it.

Third, it sits inside a larger framework. The PCCP is the change-control piece of the FDA's wider lifecycle approach to AI devices, which describes the PCCP as the way to "prospectively specify and seek premarket authorization for intended modifications to an AI-DSF ... without needing to submit additional marketing submissions" 5. Read the two together: one covers the device you validate today, the other the changes you pre-authorize for tomorrow.

How to read this — and a compliance note

The counts here are dated and will move as more plans are authorized; we revisit this page every ninety days. The figures on the first authorized plans come from a single cross-sectional analysis and describe an early, small cohort 4 — informative about direction, limited as a base rate. And this is orientation, not regulatory or legal advice: whether a PCCP is appropriate for a given device, and what a specific plan permits, are determinations for regulatory affairs and counsel. Confirm any device's authorized plan, acceptance criteria, and current status with the manufacturer and your compliance team before relying on them.

Sources and method

The three sections, the Modification Protocol's acceptance-criteria language, and the "without necessitating additional marketing submissions" purpose come directly from the FDA's final PCCP guidance 1, with its edition dates confirmed on the FDA guidance page 2 and its Federal Register availability notice 6. The five guiding principles are drawn from the joint FDA–Health Canada–MHRA document 3. The figures on the first authorized plans — counts, modification types, and the limits on preapproval and post-market evidence — come from a cross-sectional analysis of the FDA's public list 4. The relationship to the wider framework is quoted from the FDA's draft lifecycle guidance 5. For the definition and origin story, see the PCCP glossary entry; for how many devices carry a plan, see the FDA AI-enabled device list tracker and the tracker of FDA-cleared AI devices by year and specialty. We update this page every ninety days and whenever the guidance or the evidence base changes.

Questions & answers

  • What are the three parts of a PCCP?

    A Description of Modifications (the specific planned changes), a Modification Protocol (how each change will be developed, validated, and implemented, including pre-defined acceptance criteria), and an Impact Assessment (the benefits and risks of the changes and their mitigations). The FDA reviews all three together within the marketing submission.

  • What changes can a PCCP authorize?

    Only the specific changes written into the plan and validated against its protocol. Across the first 26 authorized plans, the common authorized changes were model retraining (69% of devices), logic updates (42%), and expansion of input sources (35%). Changes outside the plan — above all a change to the device's intended use — still require a new submission.

  • What should a hospital check before trusting a device's PCCP?

    Read the Description of Modifications so you know what the vendor may change, ask for the acceptance criteria and any subgroup or outcome data, confirm the intended use is unchanged, and plan local performance monitoring. Preapproval and post-market evidence for these plans has been limited, so independent monitoring matters.

Sources

  1. US Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Guidance for Industry and FDA Staff (issued August 18, 2025; originally December 4, 2024). www.fda.gov/media/166704/download
  2. US Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (guidance document page, accessed July 2026). www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence
  3. US FDA, Health Canada, Medicines and Healthcare products Regulatory Agency. Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles. October 2023. www.fda.gov/media/173206/download
  4. Regulating Flexibility for Artificial Intelligence: FDA Experience with Predetermined Change Control Plans. medRxiv preprint. 2025. doi.org/10.1101/2025.08.26.25334477
  5. US Food and Drug Administration. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (Draft Guidance, issued January 2025). www.fda.gov/media/184856/download
  6. Federal Register. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions; Guidance; Availability. December 4, 2024. www.federalregister.gov/documents/2024/12/04/2024-28361/marketing-submission-recommendations-for-a-predetermined-change-control-plan-for-artificial