Predetermined Change Control Plans: How PCCPs Are Reshaping AI/ML Device Submissions

Traditional medical device regulation assumes a device is fixed at the moment of clearance or approval. Software that learns doesn’t work that way — a model retrained on new data six months post-launch is, in a meaningful sense, a different algorithm than the one FDA reviewed. Predetermined Change Control Plans (PCCPs) are FDA’s answer to that mismatch, and by 2026 they have become one of the most consequential regulatory tools available to AI/ML device developers.

What a PCCP Actually Does

A PCCP allows a sponsor to describe, in advance and as part of the original marketing submission, the specific types of modifications an AI/ML-enabled device may undergo after clearance — algorithm retraining, performance recalibration, expanded input data sources — along with the methodology used to implement and validate each type of change. If FDA authorizes the PCCP, the sponsor can then make those pre-specified changes without submitting a new 510(k) or PMA supplement for each one.

This matters because the alternative — filing a new submission every time a model is retrained — is fundamentally incompatible with how adaptive algorithms are meant to improve over time. Without a PCCP mechanism, sponsors face a choice between freezing model performance at launch or accepting significant regulatory friction on every improvement cycle.

What Belongs in a Compliant PCCP

FDA’s guidance frames a PCCP around three core components, and reviewers evaluate all three together rather than in isolation:

1. Description of Modifications. The sponsor must specify exactly what types of changes are covered — for example, retraining on an expanded dataset to improve sensitivity for a specific subpopulation, or recalibrating a risk score threshold. Vague or open-ended modification descriptions are a common reason PCCPs get rejected or scoped back during review; FDA expects specificity about what will change and what triggers a change.

2. Modification Protocol. This is the methodology section — how will each described modification be developed, validated, and implemented before deployment? This typically includes the data management approach, retraining methodology, and the performance evaluation methods and acceptance criteria that will be applied before a modified version is released. The protocol needs to be detailed enough that FDA can assess, at authorization time, whether future changes made under it will still meet safety and effectiveness expectations — without seeing the actual future data.

3. Impact Assessment. The sponsor must assess how the described modifications, implemented per the stated protocol, will affect the device’s benefits and risks, and confirm the device will continue to provide a reasonable assurance of safety and effectiveness after each authorized change.

Why This Sits Inside a Total Product Lifecycle Approach

PCCPs don’t exist in isolation — they’re one expression of FDA’s broader Total Product Lifecycle (TPLC) approach to AI-enabled device software, first articulated comprehensively in FDA’s January 2025 draft guidance on lifecycle management and marketing submission recommendations for AI-enabled device software functions. The TPLC framing treats premarket review as the beginning of an ongoing oversight relationship rather than a single gate — real-world performance monitoring, transparency to users about how the algorithm behaves, and bias mitigation are expected to continue throughout the device’s marketed life, not just at initial clearance.

This is also where international alignment is accelerating. The FDA, Health Canada, and the UK’s MHRA jointly authored Good Machine Learning Practice (GMLP) guiding principles, and IMDRF finalized its own GMLP principles in 2025 — signaling that PCCP-style mechanisms and lifecycle expectations are converging across major regulators rather than remaining a US-only construct. For sponsors planning multi-market AI device launches, this convergence is a genuine strategic opportunity: a PCCP-style change protocol built to a harmonized standard travels better across jurisdictions than one built narrowly to satisfy a single regulator.

The Scale of the Opportunity

The growth in AI-enabled device authorizations gives a sense of how central this mechanism has become. By early 2026, FDA had authorized more than 1,350 AI-enabled devices — roughly double the number authorized in 2022. A meaningful and growing share of these submissions now include a PCCP, particularly in imaging, diagnostics, and monitoring categories where iterative model improvement is central to the product’s competitive value.

Common Pitfalls Sponsors Should Plan Around

Treating the PCCP as boilerplate. FDA reviewers scrutinize the specificity of the modification protocol closely. A PCCP written in general terms — “the model may be retrained periodically to improve performance” — without defined triggers, datasets, and acceptance criteria is unlikely to be authorized as submitted.

Underestimating the validation infrastructure required. A PCCP is only as credible as the sponsor’s ability to actually execute the modification protocol it describes — including maintaining the data pipelines, retraining infrastructure, and performance monitoring systems needed to prove each authorized change meets its stated acceptance criteria before release.

Failing to connect the PCCP to post-market surveillance. Real-world performance monitoring isn’t a separate obligation from the PCCP — it’s the evidence base that demonstrates the modification protocol is working as intended. Sponsors that build these as disconnected workstreams tend to struggle to produce a coherent picture during FDA interactions or audits.

What This Means for Regulatory Strategy

For AI/ML device developers, the PCCP decision should be made early — ideally during initial submission planning, not retrofitted after a device is already cleared. A well-constructed PCCP can be a genuine competitive advantage: it converts a source of ongoing regulatory friction into a predictable, pre-authorized improvement cycle. Sponsors that treat it as a checkbox exercise, rather than an extension of their quality and data infrastructure, tend to find that out the hard way — during FDA review, not after.

This article reflects FDA guidance and publicly available data as of 2026 and is intended for general regulatory and educational purposes. It does not constitute regulatory or legal advice for a specific device or submission.