On January 12, 2021, FDA issued an Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan (“AI/ML Action Plan”) and a discussion paper on their Proposed Regulatory Framework for Modification to Artificial Intelligence/Machine Learning (AI/ML) – Based Software as a Medical Device (SaMD) (“Modifications Discussion Paper”). FDA started its public discussion of AI/ML in 2019, as we blogged about here, with publication of an AI/ML-based SaMD discussion paper outlining a potential framework for regulation of AI/ML-based SaMD.
The new AI/ML Action Plan describes feedback in response to the initial discussion paper and briefly describes a five-part action plan at a high level. The Modifications Discussion Paper provides more detail on FDA’s thoughts related to modification of AI/ML-based SaMD and the “Predetermined Change Control Plan” that was introduced in the 2019 discussion paper.
The AI/ML Action Plan describes the following actions and goals:
Develop an update to the proposed regulatory framework presented in the AI/ML-based SaMD discussion paper, including through the issuance of a Draft Guidance on the Predetermined Change Control Plan.
Strengthen FDA’s encouragement of the harmonized development of Good Machine Learning Practice (GMLP) through additional FDA participation in collaborative communities and consensus standards development efforts.
Support a patient-centered approach by continuing to host discussions on the role of transparency to users of AI/ML-based devices. Building upon the October 2020 Patient Engagement Advisory Committee (PEAC) Meeting focused on patient trust in AI/ML technologies, hold a public workshop on medical device labeling to support transparency to users of AI/ML-based devices.
Support regulatory science efforts on the development of methodology for the evaluation and improvement of machine learning algorithms, including for the identification and elimination of bias, and on the robustness and resilience of these algorithms to withstand changing clinical inputs and conditions.
Advance real-world performance pilots in coordination with stakeholders and other FDA programs, to provide additional clarity on what a real-world evidence generation program could look like for AI/ML-based SaMD.
AI/ML Action Plan at 8.
SaMD as a whole, and especially SaMD that uses AI/ML, is updated frequently. Under the current regulatory framework, for class II devices subject to 510(k) premarket notifications, a new submission is required whenever software changes introduce new risks, impact risk controls, or significantly affects clinical functionality or performance specifications. With AI/ML, software can be retrained using new data and, given the technology, these changes can be made more quickly than other types of design changes. These types of changes often times improve performance or allow use of the software in expanded populations which, today, require a new regulatory submission.
In the Modifications Discussion Paper, FDA proposes a total product lifecycle (TPLC) approach for regulation of AI/ML-based SaMD that would allow some modifications that currently require a new regulatory submission to be made without one. The proposed TPLC approach would start with the incorporation of GMLPs into the device manufacturer’s quality system. While the specifics are not yet defined, GMLPs would describe best practices such as data management, feature extraction, training and evaluation used in the development of AI/ML-based SaMD.
The Modifications Discussion Paper then describes a regulatory pathway including an initial premarket evaluation of the AI/ML-based SaMD that includes both an evaluation of the safety and effectiveness of the device as well as an evaluation of “SaMD pre-specifications” (SPS) and an “algorithm change protocol” (ACP).
An SPS would describe “anticipated modifications to ‘performance’ or ‘inputs,’ or changes related to the ‘intended use’ of AI/ML-based SaMD.” Modifications Discussion Paper at 10.
An ACP would define “specific methods that a manufacturer has in place to achieve and appropriately control the risks of the anticipated types of modifications delineated in the SPS.” Id. After initial FDA clearance, including the SPS and ACP, future modifications within the bounds of the cleared SPS and ACP could be made without a pre-market submission.
The Modifications Discussion Paper also proposes a process for transparency and real‑world performance monitoring of AI/ML-based SaMD, including “periodic reporting to FDA on updates that were implemented as part of the approved SPS and ACP, as well as performance metrics for those SaMD.” Id. at 14. Labeling to accurately and completely describe the modification and its impact on performance are also proposed. Again, the Modifications Discussion Paper does not provide details on the mechanisms by which transparency would be achieved as it seeks to receive input from stakeholders.
For some SaMD, we see a potential tradeoff between speed of the initial clearance and speed at which future changes may be made. If the future path of the AI/ML-based SaMD is clear, spending extra time to develop an SPS and/or an ACP before the initial submission would allow future changes to be implemented without pre-market review, which will decrease time to market. In cases where future modifications may not be well defined during initial review, developing an SPS and/or an ACP may delay initial submission without much gain down the road.
Anticipating that changes may not always be known at the time of the initial submission, the Modifications Discussion Paper proposes a step where FDA can perform a focused review of the SPS and ACP for any modification that is outside of the SPS and ACP agreed to in the initial clearance. The mechanism for this interaction and its timeframe are not yet defined. For any AI/ML-based SaMD where development of an SPS and ACP at time of initial submission does not make sense or would be likely to cause delays that are not balanced by improved speed of future modifications, the traditional framework would still remain an option.
While the unique characteristics of AI/ML may make it faster to implement a design change compared to other device types, the proposed framework may also be welcomed by industry for other types of devices. A regulatory submission is currently filed at the end of development when verification and validation data supporting the final, finished device are complete for FDA review.
Many manufacturers start with an initial regulatory application for a foundational device and then submit subsequent submissions to add new feature and uses, many of which are known from the beginning. The Modifications Discussion Paper makes us wonder, if FDA can define a mechanism by which to allow modification of an AI/ML-based SaMD based on pre‑specifications and change protocols, would they consider a similar approach for other device types?
In thinking about the current approach to FDA review, we recognize that, in using the pre-submission process, there are frequently multiple iterations of a protocol with the agency before agreement is obtained. For device submissions that include data generated by following a protocol that had been subject to pre-submission interactions with FDA, FDA still scrutinizes the data and often raises questions or concerns during review of the study report in the pre-market application.
It is hard to envision FDA becoming comfortable with granting marketing authorization based on authorization of specifications and a change plan without review of the test results to ensure the protocol was strictly followed, that all data meet the specification as it was agreed upon, that appropriate regression testing was also performed, and that any justified anomalies or adverse events are appropriate. The proposed real-world performance monitoring will need to be robust to ensure FDA feels the types of issues it may have detected in premarket review would be detected quickly via the postmarket processes.
As with the initial AI/ML discussion paper, the Modifications Discussion Paper is not a guidance or even a draft guidance, but is intended to further the AI/ML discussion by eliciting additional comments and feedback that the Agency can incorporate into guidance planned for later this year.
Despite these matters being in the discussion phase, FDA has already granted marketing authorization via the De Novo pathway of a device incorporating artificial intelligence that utilized a predetermined control plan to incorporate future modifications (DEN190040); thus, it seems the Agency is committed to going in this direction for regulation of AI/ML-based SaMD.
Interested parties are encouraged to provide feedback to the questions FDA has raised in the Modifications Discussion Paper to ensure that the process, when it is able to move out of the discussion phase, will provide a viable mechanism for today’s SaMD and innovations that are yet to come.