Discussion on “Approval policies for modifications to machine learning‐based software as a medical device: A study of bio‐creep” by Jean Feng, Scott Emerson, and Noah Simon

We applaudFeng et al. for their study ofmethods to address FDA’s proposed approach to approval of modifications to machine learning (ML) based software as a medical device (SaMD). It is an excellent example of the kinds of innovative research that can help regulatory agencies like the FDA and stakeholders in their efforts to protect and promote the public health. Feng et al. were motivated by an FDA discussion paper (U.S. Food and Drug Administration, 2019a), in which the Agency reimagined how to regulate an artificial intelligence (AI) or ML-based SaMD having the power to continuously or periodically learn as data accumulate or as the needs the algorithmaddresses change. FDAenvisioned that sponsors of an ML-based SaMD propose a predetermined change control plan consisting of two parts: an SaMD prespecification (SPS) of the scope of anticipated modifications, and an algorithm change protocol (ACP) that delineates the data and procedures to be followed so that after a modification is made the device remains safe and effective with reasonable assurance. The SPS and ACP may also be useful in clarifying when a sponsor needs to seek FDA’s market authorization with a submission for modifications to a locked algorithm implemented in a nonautomated way. The proposed framework was not intended to communicate FDA’s proposed regulatory expectations, but instead meant to seek early input from stakeholders outside the Agency to be considered towards development of a more formal framework for regulating some types of modifications to an ML-based SaMD.