The FDA has published the final version of a guidance for predetermined change control plans (PCCPs) for artificial intelligence (AI) software. The final guidance is limited in terms of scope, but is a critical development in the agency’s ability to manage the volume of premarket applications it receives for AI software.
The AI PCCP guidance, which is distinct from a draft guidance for PCCPs for non-software devices, is broadened to all AI device software functions. In contrast, the draft version was directed toward machine learning (ML) software functions to
the exclusion of other AI software. The category of FDA-regulated AI devices is poised to expand significantly in the near future, given the agency’s two-day hearing on generative AI devices. There is no indication as to when the FDA might develop any guidance for generative AI, however.
The AI PCCP guidance is formally directed toward premarket submissions for AI PCCP applications that describe a series of modifications that can be made to the AI over time. This approach allows the developer to plan a series of modifications that would ordinarily require a new regulatory filing, but which will not under the terms of the guidance. The PCCP concept will reduce the number of premarket applications the agency receives for a given device type developed by a specific manufacturer/developer, but requires more up-front work on the part of the applicant to describe all the planned changes to the index device.
The draft guidance suggested a reluctance on the part of the FDA to allow manufacturers to use an AI PCCP to make changes to the indication for use statement, but the final guidance is more permissive. However, the FDA stated that the indication for use cannot stray beyond changes as would be necessary to accommodate the addition of a device or device component to the sponsor’s PCCP.
The guidance includes AI that is updated manually, automatically or a combination of both. The PCCP filing should describe how the sponsor will validate the described future changes to the AI software, which should be encoded in the developer’s quality system procedures. Risk management practices will have to be updated to accommodate the changes described in the PCCP as well.
The FDA requires that the developer use separate data sets to train and validate the AI. Developers may have to update the data they use to validate each iteration of the device described in the PCCP. The developer’s description of the protocol should include a description of software retraining practices used to ensure that the AI will function in a safe and effective manner across the adaptations described in the PCCP.
There are several types of modifications seen as appropriate for a PCCP, such as a change related to quantitative measures of the software’s performance specifications. Another acceptable modification would be to device inputs and the compatibility of those inputs with the AI. The AI may also be modified to address the performance for a subpopulation that is already within a cleared or approved indication.
The list of premarket application types that are eligible for the AI PCCP policy includes class III devices under the PMA program and class II devices under the de novo program for novel, moderate-risk devices. Sponsors can use the 510(k) program to develop a class II AI device based on a non-PCCP predicate device, although only two prongs of the 510(k) program are eligible, the traditional and abbreviated 510(k)s. Devices granted market access under the special 510(k) program are not eligible to be used as predicates for the AI PCCP program.
Copyright © 2025 - Medmarc