The Food and Drug Administration (FDA) has released an eagerly awaited update outlining its approach to regulating changes to artificial intelligence-enabled medical devices post-authorization. The final guidance focuses on pre-determined change control plans (PCCPs), a framework that allows for modifications to devices after they have entered the market.
PCCPs were first introduced in a 2019 discussion paper, aiming to address the limitations of AI or machine learning-based devices that are currently “locked,” preventing them from adapting or evolving over time. These plans enable manufacturers to propose performance enhancements through iterative changes, which are included in premarket submissions to the FDA.
The FDA initially outlined PCCPs for AI-enabled devices in a 2023 draft guidance and later expanded the concept to other medical devices, such as in-vitro diagnostics, in an August draft guidance. The final guidance, while consistent with the 2023 draft, includes a new section on version control and maintenance. It allows for revisions to PCCPs before the FDA can determine the safety and effectiveness of the device, with the possibility of authorization even if deficiencies remain unresolved.
Although the FDA has yet to authorize any adaptive AI-enabled devices, manufacturers have submitted PCCPs across various pathways, including 510(k), de novo, and premarket approval. These plans give manufacturers the flexibility to update AI-enabled devices within defined limitations, ensuring that modifications align with the authorized PCCP and the device’s intended use.
Key elements of a PCCP include a description of planned modifications, a modification protocol outlining safety and effectiveness criteria, and an impact assessment detailing benefits, risks, and mitigation plans. Manufacturers must specify the nature of planned changes, whether automatic or manual, and whether they will be uniform across devices or tailored to specific clinical settings or patient characteristics.
Device labeling should communicate the incorporation of machine learning and an authorized PCCP, alerting users to potential software updates and performance changes. An example provided by the FDA involves re-training an AI model in intensive care units to improve alarm accuracy without requiring a new marketing submission.
However, significant changes outside the scope of a PCCP, such as predicting physiological instability in advance, would necessitate a new submission. Similarly, modifications to user interfaces, like adapting algorithms for patient-facing applications, would require separate premarket applications.
In conclusion, the FDA’s final guidance on PCCPs for AI-enabled devices sets a clear framework for manufacturers to implement iterative enhancements while maintaining safety and effectiveness standards. This regulatory approach paves the way for continued innovation in the evolving landscape of medical device technology.