FDA Predetermined Change Control Plan for AI/ML Medical Devices

What a PCCP Is — and What Problem It Solves

AI/ML-based medical devices present a regulatory problem that traditional device frameworks weren't designed for: the model can change after clearance. A radiology AI trained on one hospital system's imaging data may need retraining as patient populations shift, imaging protocols update, or scanner hardware changes. Under the standard 510(k) pathway, any modification that could affect safety or effectiveness requires a new submission — which can take months and creates a disincentive to update models even when updates would improve performance.

The Predetermined Change Control Plan is FDA's mechanism for resolving this tension. It allows a manufacturer to describe, in advance and as part of their marketing submission, the specific types of modifications they anticipate making to their AI/ML device — and the controls they will apply when making those changes. If FDA accepts the PCCP, the manufacturer can implement those pre-approved modifications without filing a new 510(k), De Novo, or PMA supplement, provided the changes stay within the plan's defined scope and the specified performance protocols are met.

The concept was first formally proposed in FDA's 2021 action plan for AI/ML-based SaMD, which identified the locked vs. adaptive algorithm problem as one of the central regulatory challenges in the space. The final guidance — "Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions" — was published in December 2023 and represents the operative framework as of mid-2026.

Scope: Which Devices Can Use a PCCP

The PCCP framework applies to AI/ML-enabled device software functions (AI/ML-DSF) — a subset of Software as a Medical Device (SaMD). Not every AI device qualifies, and not every type of modification can be pre-approved.

• Devices must be subject to premarket review — 510(k), De Novo, or PMA. Devices exempt from premarket notification are outside this framework.
• The PCCP must be submitted as part of the original marketing submission or as a supplement to an existing authorization. It cannot be submitted as a standalone document.
• Both locked and adaptive (continuously learning) AI/ML algorithms are in scope, but the performance monitoring requirements differ based on whether the algorithm updates in deployment or only through discrete retraining cycles.
• Modifications that would change the device's intended use, or that would affect a different patient population than the original cleared indication, are generally outside what a PCCP can pre-authorize.

Required Components of a PCCP Submission

FDA's December 2023 guidance specifies three core elements that every PCCP must contain. Each element has sub-requirements that vary by device type and modification complexity.

1. Description of Modifications

The manufacturer must describe the specific modifications they anticipate — not as open-ended categories, but as defined change types. Examples include: retraining on expanded datasets, updating model architecture within defined bounds, adjusting decision thresholds within a specified range, or adding new imaging modalities where the clinical task remains the same.

Vague descriptions like "improving model performance" are insufficient. FDA expects the modification description to be specific enough that a reviewer can assess whether a given change falls inside or outside the plan's scope.

2. Methodology for Implementing and Validating Modifications

This is often the most technically demanding section. The manufacturer must specify the validation protocol that will govern each type of anticipated change — including dataset requirements, performance metrics, statistical thresholds, and how subgroup performance will be assessed. If the device serves populations with known demographic variation in the underlying clinical condition, the validation methodology must address how the modification will be evaluated across those subgroups.

For PMA devices, the methodology requirements are more stringent than for 510(k) submissions, consistent with the higher risk level those devices represent.

3. Performance Monitoring Protocol

The monitoring protocol specifies how the manufacturer will track device performance after a modification is implemented. This includes: what data will be collected, at what cadence, what performance thresholds trigger a safety review or rollback, and how monitoring results will be reported to FDA. The guidance does not prescribe a single monitoring approach — manufacturers have flexibility in design — but the protocol must be specific enough to be auditable.

PCCP vs. Standard Post-Market Modification Pathways

What FDA Reviews When Evaluating a PCCP

FDA's review of a PCCP is conducted as part of the overall marketing submission review. Reviewers assess whether the plan is sufficiently specific, whether the validation methodology is scientifically sound, and whether the monitoring protocol provides adequate post-market assurance. A PCCP that is accepted does not mean FDA pre-approves every future modification — it means FDA has found the plan's framework acceptable. The manufacturer still bears responsibility for executing each modification within the plan's constraints.

FDA may accept a PCCP in whole, accept it with conditions, or decline to accept it while still clearing the underlying device. A declined PCCP does not block device clearance — it simply means the manufacturer cannot use the PCCP pathway for future modifications and must file standard submissions instead.

Algorithmic Bias and Equity Considerations in PCCPs

FDA's guidance explicitly addresses the need for PCCPs to account for performance across demographic subgroups. When a manufacturer anticipates retraining or expanding a model's training dataset, the validation methodology must specify how subgroup performance will be evaluated — not just aggregate performance metrics.

This is practically significant. A model retrained on a larger dataset might show improved aggregate AUC while simultaneously degrading sensitivity for a specific demographic group. If the PCCP's validation methodology only specifies aggregate performance thresholds, that degradation might not trigger the monitoring protocol's safety review criteria. The guidance's expectation of subgroup-level evaluation is an attempt to close that gap — though it stops short of mandating specific subgroup performance thresholds, leaving those to manufacturer discretion within the plan.

Regulatory History and Current Status

The PCCP concept has a traceable regulatory history spanning several years of FDA engagement with the AI/ML device sector.

1. January 2021: FDA publishes its "Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan," identifying the need for a predetermined change control mechanism as a priority.
2. October 2021: FDA issues a discussion paper on the PCCP concept, soliciting stakeholder input on how the framework should be structured.
3. April 2023: FDA publishes a draft guidance on PCCP for AI/ML-enabled device software functions, opening a 90-day comment period.
4. December 2023: FDA finalizes the guidance — "Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions" — which is the operative document as of mid-2026.
5. 2024–2026: FDA begins reviewing and accepting PCCPs as part of 510(k) and De Novo submissions. The agency has not published a consolidated public list of accepted PCCPs, but they are visible in individual 510(k) decision summaries in the FDA database.

The PCCP is intended to provide a framework for manufacturers to proactively plan for and manage modifications to AI/ML-enabled device software functions in a manner that maintains the safety and effectiveness of the device.

FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions, December 2023

Practical Limitations of the Current Framework

The PCCP framework addresses a real problem, but it comes with constraints that manufacturers and clinical evaluators should understand before treating it as a general-purpose solution for adaptive AI.

• Specificity burden: Writing a PCCP that is specific enough for FDA to accept — while broad enough to cover the modifications a manufacturer actually anticipates — is genuinely difficult. Plans that are too vague get declined; plans that are too narrow may not cover the modifications that turn out to be needed.
• No retroactive application: A PCCP must be submitted with the original marketing application or as a supplement. Manufacturers with existing clearances who want PCCP coverage for future modifications must file a new submission.
• Intended use boundary: Modifications that would change the device's intended use — even incrementally — fall outside what a PCCP can authorize. This limits applicability for manufacturers who want to expand a device's clinical scope over time.
• Monitoring overhead: The post-market monitoring obligations in a PCCP add operational complexity. Manufacturers must collect the specified data, evaluate it against the plan's thresholds, and report results. For smaller companies, this can be a non-trivial infrastructure requirement.
• Transparency gap: There is currently no centralized FDA registry of accepted PCCPs. Clinicians and procurement staff cannot easily search for which cleared devices have accepted PCCPs and what modifications those plans authorize.

How This Connects to Broader AI/ML SaMD Oversight

The PCCP framework is one component of FDA's broader approach to AI/ML SaMD oversight, which also includes the Total Product Life Cycle (TPLC) approach, post-market surveillance requirements under 21 CFR Part 822, and the agency's ongoing work on transparency and algorithmic accountability. The PCCP addresses the pre-authorization of specific future changes; it does not replace the obligation to report device malfunctions under the Medical Device Reporting (MDR) regulation, nor does it modify the Quality System Regulation requirements for design controls.

For devices operating under a PMA, the PCCP interacts with the PMA supplement process — specifically, the 30-day notice and real-time supplement pathways that already exist for certain modifications. FDA's guidance acknowledges this overlap and notes that a PCCP may be used in conjunction with existing PMA modification pathways where appropriate.