(See what I did there with the logo?)
Anyway, the Food and Drug Administration recently issued a press release and a discussion paper regarding a framework for AI based medical devices. In particular the agency wants to deal with continuously learning or adaptive algorithms, rather than ones that have been locked (and approved). They want to be able to permit what are called “software as a medical device” that incorporate artificial intelligence or machine learning so that they develop their capabilities after intial regulatory approval; this approval would continue with monitoring the software through to real world performance by ensuring that the software continues to perform in accordance with a predetermined change control plan in order to ensure safety.
All of this sounds good but what is really interesting is the 20 page discussion paper that accompanies the release (linked in the release). Recognizing that an evolving piece of software cannot be authorized just once in advance of deployment, the paper is predicated on the idea that pre-market regulatory approval no longer works and a “total product lifecycle” regulatory approach is needed to allow for improvements but protect safety. New statutory authorities may be needed. The paper has categorized the types of modifications that could be expected from AI or machine learning in the lifecycle of a product.
The regulatory approach would be based on a) clear expectations on quality systems and good machine learning practices; b) premarket review to demonstrate safety, effectiveness and establish clear expecations for managint patient risk throughout product lifecycle; c) manufacturer monitoring and risk management approach; and d) transparency to users and FDA using post-market real-world performance reporting for maintaining assurance of safety and effectiveness. Detailed discussion of these principles is included along with flow charts for considering how the evaluation and continuous monitoring could work. Appendices to the paper include hypothetical examples of products, how they could be evaluated and proposed content for an algorithm change protocol. The proposed content for that protocol could include a data management plan, protocols for retraining and optimizing the algorithm, performance evaluation protocols and update procedures that described how updated medical device algorithms will be tested, distributed and communicated when released.
The paper concludes with a series of questions that the FDA would like submissions on.
I think that what is most interesting about this is that the proposed regulatory approach and components of the propsed analysis is a well thought through and detailed set of steps and considerations. In addition, it seems like an approach that would work in other regulatory fields, not just medical software. The total life approach to continuous regulation of a software product after pre-market regulatory approval based on a principled set of evaulation criteria might well be applicable across a number of industries. Is this a model approach?
1 thought on “FDA proposes new AI regulatory scheme; is it a model?”
This is the dawning of a new age for A.I. If handled responsibly, it can help humanity with greater magnitude.