Last week, the FDA published a discussion paper outlining a proposed regulatory framework for artificial intelligence/machine learning software as a medical device (AI/ML SaMD). Specifically, the proposal identifies the circumstances under which FDA would allow modifications without additional FDA review to the algorithm of software that continuously learns and adapts—thereby addressing some lingering unanswered questions for the SaMD community. (Software as a medical device, or SaMD, is defined by the International Medical Device Regulators Forum (IMDRF) as "software intended to be used for one or more medical purposes that perform these purposes without being part of a hardware medical device.")

Currently, modifications to SaMD that: (1) affect safety, efficacy or performance; (2) alter intended use; or (3) are major modifications to the software algorithm, likely would require FDA review prior to implementation. See “Deciding When to Submit a 510(k) for a Software Change to an Existing Device.” (FDA Guidance Document (October 2017). The issue for AI/ML SaMD, however, is that these types of changes, e.g., increased performance, use of additional inputs or expansion of intended use or indications for use may be precisely the types of adaptations that the software developer intended to take place through the algorithm’s ability to continuously learn and adapt.

While the Digital Health Software Precertification Program working model for SaMD shows: (1) FDA’s recognition of the speed and frequency of software updates in today’s world; and (2) its efforts to obviate the need for premarket review of SaMD modifications for companies that demonstrate a culture of qualify and organizational excellence and a commitment to monitoring real-world performance, FDA has yet to satisfactorily address how it would regulate continuously learning AI/ML SaMD. More specifically, there has been no guidance as to the extent FDA would require premarket review of algorithm changes resulting from software adaptations due to new data inputs, such that the outputs may be different from those originally submitted. Instead, AI/ML SaMD developers have been left to speculate as to what approach FDA would accept. 

The new paradigm outlined in FDA’s discussion paper forms a preliminary answer to these questions and solicits input from key stakeholders in order refine the details. The proposal takes a total product lifecycle (TPLC) regulatory approach and envisions firms submitting, as part of their premarket submission package, pre-specified performance objectives for the AI/ML-based SaMD and a defined algorithm change plan (ACP). 

• Approved pre-specified performance objectives (SaMD Pre-Specifications or SPS) would form the boundaries within which the algorithm would be allowed to change, as a result of its learning and adaption, without FDA review so long as the manufacturer follows its FDA-approved ACP. The SPS can include changes such as modifications to performance, inputs and, in some cases, intended uses or indications for use. FDA would evaluate each SPS based upon the AI/ML SaMD’s particular functions, IMDRF risk levels and intended uses.
• The ACP would outline the data and procedures that the manufacturer will follow to control risk while the software adapts in the manner specified in the SPS. The ACP would need to include components that address how the AI/ML SaMD developer would handle data management, retraining, performance evaluation and update procedures. 

FDA’s premarket review of AI/ML SaMD would include:

• Review of the SaMD’s performance 
• The manufacturer’s anticipated modifications
• The ability of the manufacturer to manage and control any risks resulting from those modifications

Anticipated Adaptations Included in the SaMD Pre-Specifications 

During its premarket review, FDA would determine, perhaps on a case-by-case basis, the types of changes that could be pre-specified in the SPS and managed through an ACP, based on the benefits and risks to patients. Once FDA approves a product’s SPS and ACP, future modifications based on real-world data that fall within the approved SPS and ACP, including any pre-approved changes to intended use, could be made without further FDA review (i.e., through a “Letter to File”). In FDA’s view, this approach, together with other requirements that FDA would place on manufacturers, discussed in more detail below, would provide a reasonable assurance of safety and efficacy.

Adaptations Not Included in the SaMD Pre-Specifications

Even for modifications that fall outside of the approved SPS and ACP, if there is no change intended to use, and the manufacturer is able to refine the SPS or ACP based upon real-world learning and training, then FDA may conduct a “focused review” of the SPS and ACP and approve new versions. However, FDA notes that refinement of the SPS and ACP would not be appropriate where the modification introduces significant changes to risk or a change to intended use. In those instances, FDA instead would require a new premarket submission. To assist AI/ML-based SaMD manufacturers in determining whether the modification falls within its existing SPS and ACP, manufacturers would be able to contact the applicable FDA division to confirm its decision or request a pre-submission meeting to engage with FDA in a discussion regarding the modification. 

Additional Requirements

As a TPLC approach, FDA expects manufacturers of AI/ML SaMD to commit to transparency and post-market real-world performance monitoring, including periodic reports to FDA regarding released updates and performance metrics. Transparency may require updates to labeling, for instance, to describe changes to performance, new inputs, or specifications. In addition, FDA would expect manufacturers of AI/ML-based SaMD to:

• Embrace a cultural of qualify and organizational excellence 
• Use good machine learning practices (for which FDA will establish clear expectations) and demonstrate analytical and clinical validation
• Incorporate a risk management approach in the development, validation and execution of algorithm changes. 

The idea of allowing boundaries within which an algorithm can adapt is not surprising. Validation of performance within tolerance levels is not a new concept in the life sciences industry. What does seem revolutionary, or at least evolutionary depending on how it is implemented, is the idea that FDA may allow refinement of the SPS and ACP in certain circumstances, rather than require a full premarket submission, when the modification expands beyond their boundaries. If adopted, it will be interesting to see how companies leverage this "focused review" in their strategy to bring products to market faster. 

All stakeholders should review the framework proposed in the discussion paper, which fleshes out many more details and provides useful examples of the types of modifications that may and may not be acceptable. FDA is seeking comments and feedback by June 3, 2019, through a public docket on the Federal Register: FDA-2019-N-1185. FDA will use comments received to help inform a draft guidance that it plans to release in the future.