Drafting a Predetermined Change Control Plan for AI/ML Medical Devices: A Practical Guide

Why Your AI/ML Device Needs a PCCP Before It Hits the Market

If you are developing an artificial intelligence or machine learning-based medical device, you already know the technology does not stand still. Models retrain. Performance drifts. Clinical data shifts. The core regulatory challenge has always been this: how do you continuously improve an adaptive algorithm without filing a new 510(k) every six months?

The FDA's answer is the Predetermined Change Control Plan, or PCCP. Introduced formally in the Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions guidance issued in December 2024, the PCCP framework gives manufacturers a structured, pre-approved pathway to implement certain modifications to AI/ML device software functions without submitting a new premarket application for each change. Done right, it is a genuine competitive advantage. Done poorly, it invites a 510(k) hold, a Refuse to Accept, or worse, post-market enforcement action.

This guide walks you through what a defensible PCCP actually looks like in practice.

The Regulatory Foundation: What You Are Actually Agreeing To

A PCCP is submitted as part of your premarket submission, whether that is a 510(k) under 21 CFR Part 807, a De Novo request, or a PMA under 21 CFR Part 814. When FDA clears or approves your device with a PCCP, the plan becomes a binding part of your marketing authorization. Deviating from it without prior authorization is a violation of your clearance conditions, full stop.

The statutory authority sits in Section 515C of the FD&C Act, added by the 21st Century Cures Act and reinforced by the AI/ML Action Plan FDA published in January 2021. The December 2024 guidance operationalizes those provisions with concrete expectations for what each PCCP section must contain.

The Three Core Components of a PCCP

1. The Statement of Planned Changes (SPS)

The SPS defines the specific modifications you anticipate making to the AI/ML device software function. This is not a vague wish list. FDA expects you to be concrete and bounded. Acceptable planned changes typically fall into categories such as:

• Modifications to the model architecture or algorithm type within a defined scope
• Retraining on expanded or updated datasets from specified sources or patient populations
• Changes to input feature sets within pre-described boundaries
• Performance threshold adjustments within a pre-approved range

Each planned change must be described with enough specificity that a reviewer can determine whether a future modification falls inside or outside the approved envelope. Vague language like 'periodic model updates as needed' will not pass review. Be explicit about what changes are in scope and, critically, what is out of scope.

2. The Impact Assessment Protocols (IPS)

For every planned change in your SPS, you must describe how you will evaluate whether that change maintains safety and effectiveness before implementation. The IPS is essentially your pre-approved validation protocol. It must specify:

• The performance metrics that will be evaluated (e.g., sensitivity, specificity, AUC, calibration)
• The datasets to be used for evaluation, including their provenance, diversity, and relevance to the intended use population
• Pre-specified acceptance criteria that must be met before the change is implemented
• Any human factors or usability considerations triggered by the change

FDA pays close attention to whether your acceptance criteria are clinically meaningful, not just statistically convenient. If your device is used in a high-stakes clinical decision context, performance thresholds need to reflect the real-world consequences of false positives and false negatives.

3. The Methodology for Implementing and Monitoring Changes

This section describes your post-implementation surveillance plan. How will you detect performance degradation in the real world? What triggers a rollback or a new submission? FDA expects you to reference your existing Quality Management System processes under 21 CFR Part 820 and align your monitoring approach with your complaint handling and CAPA procedures. For companies operating under ISO 13485, your change control SOPs should map directly to this section.

Common Drafting Mistakes That Sink PCCPs

After reviewing dozens of AI/ML submissions, the patterns of failure are predictable. The most common problems include planned changes that are defined so broadly they are meaningless, acceptance criteria that are not pre-specified but instead described as 'to be determined,' and impact assessment protocols that rely on internal datasets without acknowledging distributional shift risks. Another frequent gap is failing to address what happens when a planned change produces an unexpected result during the IPS evaluation. FDA wants to see a clear decision tree: pass, fail, or escalate to a new submission.

Practical Sequencing: Build the PCCP in Parallel with Your Design

The worst time to draft a PCCP is after your model architecture is locked and your 510(k) is already in front of a reviewer. The best PCCP submissions are developed alongside the device design, because the planned changes you describe should reflect your actual product roadmap, not an afterthought. Involve your clinical, software, and data science teams early. The regulatory team's job is to translate their development intent into FDA-acceptable language, and that translation takes time.

One Final Consideration: Transparency in Labeling

Under the December 2024 guidance, FDA expects that your labeling discloses the existence of a PCCP and provides users with sufficient information to understand that the device may change over time. This is not just a regulatory checkbox. For clinicians relying on your device for patient care, transparency about algorithmic updates is an ethical obligation as much as a regulatory one.

Ready to Build a PCCP That Will Hold Up to FDA Scrutiny?

At ADB Consulting and CRO Inc., Andre Butler and the team work directly with AI/ML device developers to draft PCCPs that are specific, defensible, and aligned with your real product strategy. We have helped startups and established manufacturers navigate the intersection of adaptive algorithms and FDA expectations, and we know exactly where submissions get stuck.

Do not wait until your 510(k) is on hold to figure out your change control strategy. Book a free discovery call with our team today at adbccro.com and let us help you get this right the first time.