FDA Outlines AI Device Modification Scenarios Under PCCPs

Share

FDA has published two hypothetical examples illustrating how manufacturers of AI-enabled medical devices could modify their products under predetermined change control plans (PCCPs) without submitting a new marketing application for each change.

The examples, posted 10/20, describe regulatory approaches for a hospital-specific sepsis alert system and a continuously learning cardiac arrhythmia detection device.

In the first example, FDA describes how a machine-learning system designed to identify patients at risk of sepsis could be adapted to individual hospitals using local patient data. Potential modifications include model fine-tuning, adjustments to feature weighting and changes to alert thresholds, subject to predefined performance criteria.

The second example addresses a deep-learning cardiac arrhythmia detection system that undergoes periodic automated retraining and deployment of updated models. Under the hypothetical PCCP, updates could be implemented without human review or intervention if they satisfy prespecified validation and performance requirements.

The examples serve to outline FDA's approach to allowing AI-enabled medical devices to evolve after authorization while maintaining oversight of changes that could affect safety and effectiveness.

PCCPs allow manufacturers to identify anticipated device modifications, describe methods for developing and validating those changes, and assess their potential impact as part of an initial marketing submission. Once FDA authorizes a PCCP, modifications implemented within its authorized scope generally do not require a separate marketing submission.

Read more