In April 2019, the US Food and Drug Administration (FDA) issued a white paper, “Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device,” announcing steps to consider a new regulatory framework to promote the development of safe and effective medical devices that use advanced AI algorithms. AI, and specifically ML, are “techniques used to design and train software algorithms to learn from and act on data.” FDA’s proposed approach would allow modifications to algorithms to be made from real-world learning and adaptation that accommodates the iterative nature of AI products while ensuring FDA’s standards for safety and effectiveness are maintained.
Under the existing framework, a premarket submission (i.e., a 510(k)) would be required if the AI/ML software modification significantly affects device performance or the device’s safety and effectiveness; the modification is to the device’s intended use; or the modification introduces a major change to the software as a medical device (SaMD) algorithm. In the case of a PMA-approved SaMD, a PMA supplement would be required for changes that affect safety or effectiveness. FDA noted that adaptive AI/ML technologies require a new total product lifecycle (TPLC) regulatory approach and focuses on three types of modifications to AI/ML-based SaMD:
- performance, which modify clinical and analytical performance
- inputs, which are used by the algorithm and their clinical association with the SaMD output, or
- intended use, which is described through the significance of information provided by the SaMD for the state of the healthcare situation or condition
Traditional SaMD is evaluated on a risk-based approach, by weighing the significance of the information provided by the SaMD to the healthcare decision (treat or diagnose, drive clinical management, inform clinical management) against the state of the healthcare situation or condition (critical, serious, non-serious). AI/ML-based SaMDs, however, introduce a new variable—whether the software exists on a spectrum from “locked” to continuously learning or adaptive—posing a particular challenge in determining a threshold for when modifications to such devices should undergo premarket review. As such, FDA proposed four general principles to balance benefits and risks of AI/ML-based SaMD while minimizing regulatory burdens and allowing software to continue to learn and evolve over time to improve patient care:
- Establishing clear expectations on quality systems and good ML practices (GMLP)
- Conducting premarket review for those SaMD that require premarket submission to demonstrate reasonable assurance of safety and effectiveness and establishing clear expectations for manufacturers of AI/ML-based SaMD to manage patient risks throughout the lifecycle of the software (i.e., relying on the principle of a “predetermined change control plan” that anticipates certain modifications, the “SaMD Pre-Specifications,” and associated methodology for those changes, the “Algorithm Change Protocol” in a controlled manner that manages risks to patients)
- Expecting manufacturers to perform continuous monitoring on their AI/ML devices and incorporate a risk management approach and other approaches outlined in FDA’s “Deciding When to Submit a 510(k) for a Software Change to an Existing Device” guidance in the development, validation, and execution of the algorithm changes
- Enabling increased transparency to users and FDA using post-market real-world performance reporting for maintaining continued assurance of safety and effectiveness
The comment period for the white paper closed on June 3. It is unclear whether the agency will issue a guidance document that formalizes its proposed approach or whether it will informally draw from the principles it established in the white paper.