PCCP Case Examples: How AI/ML Medical Devices Are Scoping Predetermined Change Control Plans

Why PCCP Scoping Is the Hard Part Nobody Talks About

The FDA's 2024 guidance on Predetermined Change Control Plans (PCCPs) gave the industry a framework. What it did not give you is a worked example of how to draw the line between what your AI/ML device can change autonomously and what triggers a new submission. That ambiguity is where regulatory strategy either pays off or costs you six to twelve months of delay.

At ADB Consulting and CRO Inc., we work with medical device startups and growth-stage companies navigating exactly this problem. Below, we walk through three representative PCCP scoping scenarios based on the types of AI/ML Software as a Medical Device (SaMD) we see most frequently. These are not hypothetical abstractions. They reflect the real scoping decisions companies face when drafting a PCCP under 21 CFR Part 880, De Novo pathways, and PMA supplement frameworks.

The Regulatory Foundation You Must Understand First

Before scoping a PCCP, your team needs to internalize three documents:

• FDA's Marketing Submission Recommendations for a Predetermined Change Control Plan for AI/ML-Enabled Device Software Functions (2024) -- this is the primary guidance and defines the two PCCP components: the Description of Planned Changes and the Modification Protocol.
• FDA's AI/ML-Based SaMD Action Plan (2021) -- provides the philosophical underpinning for why PCCPs exist, rooted in the concept of a 'good machine learning practice' lifecycle.
• 21 CFR 814.39 and 21 CFR 807.81 -- these govern when a PMA supplement or new 510(k) is required, and your PCCP must explicitly argue why anticipated changes fall outside those triggers.

The critical legal concept: a PCCP does not exempt you from substantial equivalence or safety/effectiveness obligations. It pre-authorizes a defined change within a pre-approved modification protocol. Scope too broadly and FDA will reject it. Scope too narrowly and you negate the business value of having one.

Case Example 1: AI-Assisted Radiology Triage (510(k) Pathway, Class II)

A chest X-ray triage SaMD sought 510(k) clearance with an embedded PCCP. The intended use was detection of findings warranting urgent radiologist review. The device's manufacturer wanted flexibility to retrain the model on new patient demographics without returning to FDA.

Their scoping decision: the PCCP covered performance optimization changes, specifically retraining on expanded demographic datasets to maintain sensitivity and specificity within a pre-specified performance boundary. The Description of Planned Changes stated that model updates would not alter the intended use, the output label, or the clinical decision point. The Modification Protocol required internal validation against a locked reference dataset, bias analysis across defined subgroups, and an annual real-world performance report submitted under 21 CFR 803.

What they deliberately excluded from PCCP scope: any change to the output threshold that would reclassify findings, any expansion to new anatomical regions, and any change to how the device integrates with the clinical workflow. Those were reserved for a Special 510(k) or new submission.

The lesson: tie your PCCP scope to performance maintenance, not performance expansion. FDA is far more comfortable approving a PCCP that keeps you within a validated operating corridor than one that attempts to move the goalposts.

Case Example 2: Continuous Glucose Monitoring Algorithm Update (PMA Supplement)

A Class III CGM device with an approved PMA sought to include a PCCP that would allow iterative algorithm updates as new sensor hardware generations were introduced. This is a higher-stakes scenario because any deviation from the approved algorithm in a Class III device traditionally requires a PMA supplement under 21 CFR 814.39(a).

The PCCP scoped only one category of change: algorithm recalibration to account for sensor drift characteristics of a new but substantially equivalent sensor component already cleared separately. The Modification Protocol required bench testing per ISO 15197, clinical bridging data from a minimum subject cohort, and a predefined statistical equivalence margin. The company did not attempt to include adaptive learning or autonomous retraining in scope.

The lesson: in Class III, your PCCP scope should be surgical. One change category, rigorously controlled. Attempting to bundle multiple change types in a PMA PCCP invites a not-approvable letter.

Case Example 3: Mental Health Risk Stratification SaMD (De Novo, Class II)

A behavioral health SaMD that stratified patient suicide risk sought De Novo classification with a PCCP component. The scoping challenge here was unique: the model relied on natural language inputs, making drift detection and retraining cycles a clinical necessity, not just a commercial convenience.

The company scoped their PCCP to allow quarterly model updates triggered by documented performance drift, defined as a statistically significant change in positive predictive value measured against a prospective sentinel dataset. The Modification Protocol required an independent clinical review of flagged cases before any update deployment, a 30-day shadow mode validation period, and real-world performance monitoring per FDA's 2022 discussion paper on AI transparency.

Critically, the PCCP excluded any change to the risk tier thresholds -- low, moderate, high -- that clinicians used to make care escalation decisions. Those thresholds were locked as a predicate performance characteristic.

The lesson: for high-sensitivity mental health applications, your PCCP needs a human-in-the-loop validation step in the Modification Protocol. FDA will look for it, and its absence signals immature quality system thinking.

The Scoping Principles That Apply Across All Three Examples

• Define your performance boundaries quantitatively. Vague language like 'maintain acceptable performance' will not survive FDA review. Specify your metrics, your thresholds, and your reference datasets.
• Separate intended use from model performance. PCCP scope lives in performance space. The moment a change touches intended use or indications for use, you are outside PCCP territory.
• Build your Modification Protocol to your QMS, not the other way around. Your PCCP is a regulatory document, but it will be executed by your quality system. If your ISO 13485-compliant procedures cannot operationalize the Modification Protocol, you have a compliance gap before your first update.
• Address bias and subgroup performance explicitly. FDA's reviewer guidance on AI/ML consistently flags demographic performance disparities. Your PCCP validation steps should include subgroup analysis by sex, age, and race at minimum.

What This Means for Your Next Submission

A poorly scoped PCCP creates a false sense of regulatory flexibility. Companies that draft PCCPs without grounding them in the specific change types their development roadmap actually requires -- and the risk classification that governs their device -- end up with documents that are either rejected outright or so narrow they provide no practical benefit.

Done correctly, a PCCP is one of the most valuable regulatory assets an AI/ML device company can hold. It signals to FDA that you have a mature understanding of your model lifecycle, your risk profile, and your post-market surveillance obligations.

Ready to Scope Your PCCP the Right Way?

At ADB Consulting and CRO Inc., Andre Butler and the team have guided AI/ML SaMD companies through 510(k), De Novo, and PMA pathways with embedded PCCPs. We do not offer cookie-cutter templates. We build regulatory strategies that are specific to your device, your data, and your commercial timeline.

Book a free discovery call at adbccro.com and let us help you define a PCCP scope that FDA will approve and your engineering team can actually execute.

For related guidance, see our Software as a Medical Device regulatory support.