PCCP vs 510(k) Supplement for AI Algorithm Changes — When to Use Each

Cleared. Now what? For AI medical device companies, clearance is not the end of the regulatory journey — it's the beginning of a long-term relationship with FDA that plays out every time your algorithm changes. And AI algorithms change. They retrain on new data, expand to new populations, improve with clinical feedback, and drift with distribution shift.

Understanding which regulatory pathway to use for each type of change is essential. The options are: no submission required (if within PCCP scope), Special 510(k), Traditional 510(k), or De Novo for truly novel changes. Here's how to navigate them.

The Hierarchy of Algorithm Change Pathways

Option 1: No Submission Required (PCCP-Covered Change)

If you have an approved PCCP and your algorithm change falls within its specified scope — and you can document that you've met the pre-specified performance goals — you don't need to submit anything to FDA before implementing the change. You document it in your quality system and update your Device History File.

This is the gold standard for AI device companies because it allows continuous improvement without regulatory lag. The catch: you need to have designed your PCCP carefully upfront. PCCPs that are too vague don't provide meaningful coverage. PCCPs that are too narrow don't cover the changes you'll actually want to make.

Option 2: Special 510(k)

For changes to a cleared device that don't raise new questions of safety or effectiveness, a Special 510(k) is faster than a traditional submission. FDA aims to review Special 510(k)s in 3 months rather than the 9-14 months typical for Traditional submissions.

Special 510(k) is appropriate for algorithm changes when: the change is well-characterized, you can rely on design controls and risk analysis to demonstrate the change doesn't raise new issues, and the change doesn't affect the intended use or alter the fundamental scientific technology.

For AI devices, Special 510(k) is most appropriate for: performance improvements where the algorithm architecture doesn't change, threshold adjustments within validated ranges, and software updates that fix bugs without changing the AI's decision logic.

Option 3: Traditional 510(k)

Significant algorithm changes that raise new performance questions require a Traditional 510(k). This is the full submission with new clinical performance data, predicate comparison, and special controls compliance documentation.

For AI devices, Traditional 510(k) is typically required when: the algorithm architecture changes significantly, training data expands to new patient populations or clinical settings, the intended use expands to new indications, or the device's outputs change in ways that affect clinical decision-making.

The PCCP Decision Framework

Before your initial submission, ask:

• Do we plan to retrain our model on new data? → PCCP needed
• Will we expand to new hospital systems or patient populations? → PCCP needed
• Will we update our algorithm as clinical evidence accumulates? → PCCP needed
• Is our device a "lock" algorithm that won't change post-clearance? → PCCP optional but not critical

If any of the first three apply, design your PCCP before you submit. It's far harder to add a PCCP to a cleared device than to include one in your original submission.

What a Good PCCP Looks Like

Approved PCCPs in FDA's public records share common characteristics. They specify: the exact types of modifications anticipated (not vague categories), quantitative performance thresholds that must be maintained, the statistical methods used to assess performance, and the site and patient diversity requirements for validation.

The most common PCCP rejection: insufficient specificity. "Retraining on new data" is not a sufficient description. "Retraining on additional CT imaging data from new clinical sites, maintaining sensitivity ≥ 94% and specificity ≥ 87% on a holdout set of ≥ 500 cases with defined demographic distribution" is getting closer.

Enforcement Risk Without a PCCP

FDA has issued warning letters to companies that updated their AI algorithms post-clearance without appropriate regulatory coverage. The enforcement position is clear: if your change would have required a 510(k) if submitted as a new device, it requires a 510(k) as a post-market modification — unless you have PCCP coverage.

This means every AI device company should have an internal algorithm change control procedure that maps each type of potential change to its regulatory pathway before implementing it.