AI medical device updates: how do we control change?

FDA Webinar Marketing Submission Recommendations for a PCCP for AI-Enabled Device Software Functions

U.S. Food and Drug Administration · 2025-04-04 · Source

English-language recording of the FDA’s January 14, 2025 session, published on April 4, 2025. It explains the PCCP approach and predates the August 2025 guidance revision. Use the guidance linked below for the revised text.

Research evidence: which version was evaluated?

When reading AI research, ask which version and conditions were evaluated. The joint MHRA, FDA and Health Canada principles emphasize evidence showing benefits and risks before and after a change. Assuming that success for the first version automatically applies to every later version skips this assessment.

Regulatory framework: a PCCP is not permission for unlimited changes

The FDA’s August 18, 2025 guidance describes three connected parts of a predetermined change control plan (PCCP): planned modifications, methods for developing and validating them, and an impact assessment. Changes consistent with an authorized plan may not require a new marketing submission each time. This is a US approach, not blanket authorization for every country or update. Guidance recommendations are distinct from binding law.

Clinical use: monitoring continues after an update

The joint principles call for bounded changes, a risk-based and evidence-based approach, transparency and a total product lifecycle perspective. Mechanisms to stop or reverse changes that fail performance criteria are part of this framework. Installing an update at a hospital does not itself establish that clinical performance has been validated.

Transparency: what should the clinician know?

The FDA’s transparency principles address communicating intended use, performance and limitations clearly to relevant users. Where, when and how information is provided also matters. A technical explanation of the model alone is insufficient: information should support decisions within the workflow of people working with AI.

Editorial perspective: start with a version record

Our practical suggestion for institutions is to record the old and new versions, changed behavior, evaluation results, responsible team and rollback path together for each update. This is an editorial working proposal, not a completed conformity assessment or jurisdiction-specific legal review. The article provides no personal treatment advice and does not establish clinical effectiveness for RAI products.