On April 2, 2019, the Food and Drug Administration (FDA) issued a statement by Commissioner Scott Gottlieb on steps toward a new review framework specifically tailored to promote the development of medical devices that use artificial intelligence algorithms. The statement announced that FDA released a discussion paper and request for feedback, Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD), as a first step toward developing this new approach and helping innovators and developers bring artificial intelligence devices to market. The paper is for discussion purposes only and not meant as a draft guidance for industry.
This Bulletin provides a high-level overview of FDA’s proposed framework; it is not intended to provide a detailed explanation of the proposal, which is very technical.
Artificial intelligence (AI) is the science and engineering of creating an intelligent machine. AI can use various techniques to create intelligent behavior. FDA explains that AI algorithms are software that can learn from and act on data. Artificial intelligence/machine learning (AI/ML)-based software refers to software that is able to create intelligent behavior by learning from, and acting on, data. AI/ML-based software intended to treat, diagnose, cure, mitigate, or prevent disease or other conditions is regulated as a medical device and falls under the category of Software as a Medical Device (SaMD). The agency has already cleared some SaMDs, such as a software that can detect diabetic retinopathy.
To date, FDA’s clearance or approval of AI/ML-based SaMD has typically involved algorithms that are locked prior to marketing and that do not continually adapt or learn and improve when the algorithm is used. These types of algorithms are modified by the manufacturer at intervals and may require FDA premarket review for modifications beyond the original market authorization if the changes affect the safety or effectiveness of the device. In contrast, adaptive AI/ML algorithms continuously learn and adapt over time without manual modification. Devices that use these algorithms are able to learn from new real-world user data presented to the algorithm and may result in a different output compared to the output initially cleared by FDA.
The discussion paper focuses on a framework that would allow for modifications to AI/ML-based SaMD to be made from real-world learning and adaptation, while also ensuring safety and effectiveness. This framework is meant to be more suited for AI/ML-based software than the existing medical device regulatory framework for SaMD. As Dr. Gottlieb notes, “we’re working to develop an appropriate framework that allows the software to evolve in ways to improve its performance while ensuring that changes meet our gold standard for safety and effectiveness throughout the product’s lifecycle.”
FDA’s Proposed Framework
The proposed framework would give manufacturers an option to submit a plan for modification during the premarket review. FDA’s determination regarding the acceptability of the plans would include a consideration of the algorithm’s performance, the manufacturer’s plan for modifications, and the ability of the manufacturer to manage and control risks of the modifications. The framework also proposes that manufacturers would have the option to submit a plan for modifications during the initial premarket review so that FDA’s determination would provide reasonable assurance of safety and effectiveness, including the manufacturer’s ability to manage and control resultant risks of the modifications. The proposed framework relies on a principle of what the agency refers to as a “predetermined change control plan,” which would allow FDA the option to review the types of anticipated modifications based on the algorithm’s re-training and update strategy, and the methodology being used to implement those changes in a way that manages risks to patients.
The proposed framework includes an assessment of the manufacturer’s quality and organizational excellence to ensure that FDA has reasonable assurance of the high quality of the software development, testing, and performance monitoring. The discussion paper notes that manufacturers can work to assure the safety and effectiveness of their software by using appropriate mechanisms to support transparency about the function and modifications of medical devices (including updates to FDA, device companies and collaborators of the manufacturer, and the public, such as clinicians) and real-world performance monitoring. According to FDA, many of the modifications to AI/ML-based SaMD may be supported by analysis of real-world data. Gathering this data may allow manufacturers to understand how to improve and respond proactively to safety concerns. The agency would also expect manufacturers to use a quality assurance system and demonstrate analytical and clinical validation, as described in the Software as a Medical Device (SaMD): Clinical Evaluation Guidance for Industry.
The paper concludes with examples of AI/ML-based SaMD modifications that may and may not be permitted without additional agency review. FDA plans to issue draft guidance based on input received from experts and stakeholders.
- FDA’s proposed framework represents a focus on promoting advancements in AI/ML-based technologies, while minimizing risks to public safety due to the adaptive nature of these technologies.
- FDA’s thinking and position will continue to evolve as the agency issues guidance documents based on public feedback received in response to the discussion paper.
- The agency expects to see more medical devices using advanced AI algorithms to improve the products’ performance and safety. Companies intending to use such technology should continue to follow FDA’s approach and thinking as they develop their regulatory strategy for their own proposed medical devices.
For more information about the discussion paper and FDA’s proposed regulatory framework, please contact Alan Minsk or Genevieve Razick.