Danielle Besal, MS, Principal Consultant, MRC Global LLC09.10.21
One of the fastest-growing segments of the medical technology market is that of artificial intelligence. Artificial intelligence/machine learning (AI/ML) has wide-ranging applications within the healthcare industry—from its use in automating internal business performance, improving health analytics, or incorporation within medical devices. With 80 percent of healthcare executives planning to adopt AI tools1 and the global AI in healthcare market projected to grow upwards of 40 percent from 2021-2028,2 there is an ever-increasing need for clarity on how to achieve U.S. market access for medical devices incorporating AI algorithms.
In the United States, the U.S. Food and Drug Administration (FDA) pathway for device marketing authorization relies on the demonstration of safety and effectiveness of the final device design (regardless of a product’s classification). This presents an obvious problem for AI/ML technologies that are designed to learn and adapt over time in response to real-world data. How can a medical device manufacturer remain in regulatory compliance when its product is constantly evolving and changing in real time?
The FDA has proposed a new regulatory framework for AI/ML devices following a total product lifecycle (TPLC) approach.3,4 The FDA has long advocated for a total product lifecycle approach for medical devices. It is the basis for FDA’s Software Precertification Pilot Program and certain device-specific FDA Guidance (e.g., infusion pumps). The framework for AI-based devices centers on four pillars:
These requirements may seem familiar, as they are essentially analogous to standard requirements such as good software engineering and quality system practices. In fact, FDA has stated it will encourage this harmonization of GMLP development to standard practices and actively participates in industry and standards working groups for AI.
AI-based medical devices follow the same FDA premarket review processes [i.e., 510(k), PMA, De Novo, if required] for the purposes of demonstrating the safety and effectiveness of the product. Due to the nature of AI algorithms, FDA has stated the premarket review process will also focus on the expectations for continual risk management throughout a device’s lifecycle. Another focus of the premarket review process specific to AI devices is the emphasis on algorithm bias and robustness. Whether the algorithm incorporated within the device is a locked or adaptive design, the hallmark of AI/ML algorithms is their initial training from historical data sets. Thus, if biases were present in the data used to train the algorithms, the device will parrot these in its performance. The FDA expects that manufacturers justify the methods used to address bias and evaluate robustness for the algorithm.
Perhaps the most notable difference in the proposed framework comes from the concepts of “pre-specifications” and a predetermined algorithm change protocol. Pre-specifications are a boundary of anticipated modifications to the performance, inputs, or intended use of the device that may be implemented as the algorithm learns over time. The algorithm change protocol is the predetermined set of methods a manufacturer implements to ensure the device and its anticipated modifications (i.e., pre-specifications) remain safe and effective over time. Device manufacturers are used to evaluating product changes and seeking FDA authorization for that change (when necessary) prior to implementation. For devices that incorporate adaptive AI/ML algorithms, this presents a regulatory challenge, as the device inputs or performance may change in real time and, thus, it may not be possible to submit the change to FDA for authorization prior to its implementation. The intent of submitting pre-specifications and an algorithm change control plan with the device’s initial premarket application is obtaining FDA consensus on which anticipated modifications are allowable in real time without the need for a subsequent application. Thus, contrary to the traditional regulatory framework, the AI-based device may learn and adapt over time and, as long as those changes remain within the boundaries of the FDA-agreed pre-specifications and are validated according to the algorithm change plan, the modifications may be documented without the need for a new FDA submission.
The final pillar of the proposed framework is that of transparency and performance monitoring. The obvious benefits of an AI-based device being able to change and improve its performance over time are accompanied by inherent risks of user trust—how is one able to understand the device performance at any point in time? Adaptive AI algorithms necessitate a high level of post-market monitoring by the manufacturer. The manufacturer must commit to frequent labeling updates in order to fully convey the extent of, rationale for, and new performance resulting from any algorithm changes. In communicating the changes, both the device user as well as the patient should be included for audience considerations. FDA expects that reports of the device’s real-world performance and changes implemented (in accordance with the approved pre-specifications and algorithm change protocol) be submitted to the agency at regular intervals.
Although there are many parallels in the proposed AI-based device framework to the traditional regulatory framework, medical products incorporating AI algorithms present unique risks worthy of the slightly increased regulatory burden. The concepts of pre-specifications and algorithm change protocols provide a feasible means for device manufacturers to remain in regulatory compliance while maintaining the inherent benefits of artificial intelligence algorithms. FDA has indicated a draft guidance document for Predetermined Change Control Plans is forthcoming this year and will include recommendations on the scope of pre-specifications and algorithm change protocols. In its action plan published earlier this year, FDA has committed to a continued focus on AI/ML technologies and plans to conduct many public workshops to continue collecting stakeholder input. Much like AI algorithms themselves, the regulatory framework for AI-based devices will continue to evolve over time.
References
As a principal consultant at MRC Global, Danielle Besal is an expert in leading and managing regulatory affairs projects. This includes managing regulatory affairs teams in multiple classes of trade through the completion of submissions, change control, business development activities, and maintaining global regulatory compliance. Danielle has also spearheaded a transition of an acquisition-related manufacturer to change all global product registrations, as well as chaired the operations-based team through the implementation of corresponding rebranding activities. Prior to MRC Global, Danielle was a Regulatory Affairs manager at Medtronic plc, where she managed regulatory teams and large-scale projects. Prior to Medtronic, Danielle was a Regulatory Affairs project manager at MicroPort Orthopedics and Wright Medical Technology. Danielle earned a bachelor of science degree in physics from Rhodes College and completed a master of science degree in biomedical engineering at the University of Memphis & University of Tennessee Health Science Center.
In the United States, the U.S. Food and Drug Administration (FDA) pathway for device marketing authorization relies on the demonstration of safety and effectiveness of the final device design (regardless of a product’s classification). This presents an obvious problem for AI/ML technologies that are designed to learn and adapt over time in response to real-world data. How can a medical device manufacturer remain in regulatory compliance when its product is constantly evolving and changing in real time?
The FDA has proposed a new regulatory framework for AI/ML devices following a total product lifecycle (TPLC) approach.3,4 The FDA has long advocated for a total product lifecycle approach for medical devices. It is the basis for FDA’s Software Precertification Pilot Program and certain device-specific FDA Guidance (e.g., infusion pumps). The framework for AI-based devices centers on four pillars:
- Good machine learning practices (GMLP)
- Premarket review for safety and effectiveness
- Established pre-specifications and algorithm change protocol
- Patient-focused transparency and real-world performance monitoring
These requirements may seem familiar, as they are essentially analogous to standard requirements such as good software engineering and quality system practices. In fact, FDA has stated it will encourage this harmonization of GMLP development to standard practices and actively participates in industry and standards working groups for AI.
AI-based medical devices follow the same FDA premarket review processes [i.e., 510(k), PMA, De Novo, if required] for the purposes of demonstrating the safety and effectiveness of the product. Due to the nature of AI algorithms, FDA has stated the premarket review process will also focus on the expectations for continual risk management throughout a device’s lifecycle. Another focus of the premarket review process specific to AI devices is the emphasis on algorithm bias and robustness. Whether the algorithm incorporated within the device is a locked or adaptive design, the hallmark of AI/ML algorithms is their initial training from historical data sets. Thus, if biases were present in the data used to train the algorithms, the device will parrot these in its performance. The FDA expects that manufacturers justify the methods used to address bias and evaluate robustness for the algorithm.
Perhaps the most notable difference in the proposed framework comes from the concepts of “pre-specifications” and a predetermined algorithm change protocol. Pre-specifications are a boundary of anticipated modifications to the performance, inputs, or intended use of the device that may be implemented as the algorithm learns over time. The algorithm change protocol is the predetermined set of methods a manufacturer implements to ensure the device and its anticipated modifications (i.e., pre-specifications) remain safe and effective over time. Device manufacturers are used to evaluating product changes and seeking FDA authorization for that change (when necessary) prior to implementation. For devices that incorporate adaptive AI/ML algorithms, this presents a regulatory challenge, as the device inputs or performance may change in real time and, thus, it may not be possible to submit the change to FDA for authorization prior to its implementation. The intent of submitting pre-specifications and an algorithm change control plan with the device’s initial premarket application is obtaining FDA consensus on which anticipated modifications are allowable in real time without the need for a subsequent application. Thus, contrary to the traditional regulatory framework, the AI-based device may learn and adapt over time and, as long as those changes remain within the boundaries of the FDA-agreed pre-specifications and are validated according to the algorithm change plan, the modifications may be documented without the need for a new FDA submission.
The final pillar of the proposed framework is that of transparency and performance monitoring. The obvious benefits of an AI-based device being able to change and improve its performance over time are accompanied by inherent risks of user trust—how is one able to understand the device performance at any point in time? Adaptive AI algorithms necessitate a high level of post-market monitoring by the manufacturer. The manufacturer must commit to frequent labeling updates in order to fully convey the extent of, rationale for, and new performance resulting from any algorithm changes. In communicating the changes, both the device user as well as the patient should be included for audience considerations. FDA expects that reports of the device’s real-world performance and changes implemented (in accordance with the approved pre-specifications and algorithm change protocol) be submitted to the agency at regular intervals.
Although there are many parallels in the proposed AI-based device framework to the traditional regulatory framework, medical products incorporating AI algorithms present unique risks worthy of the slightly increased regulatory burden. The concepts of pre-specifications and algorithm change protocols provide a feasible means for device manufacturers to remain in regulatory compliance while maintaining the inherent benefits of artificial intelligence algorithms. FDA has indicated a draft guidance document for Predetermined Change Control Plans is forthcoming this year and will include recommendations on the scope of pre-specifications and algorithm change protocols. In its action plan published earlier this year, FDA has committed to a continued focus on AI/ML technologies and plans to conduct many public workshops to continue collecting stakeholder input. Much like AI algorithms themselves, the regulatory framework for AI-based devices will continue to evolve over time.
References
As a principal consultant at MRC Global, Danielle Besal is an expert in leading and managing regulatory affairs projects. This includes managing regulatory affairs teams in multiple classes of trade through the completion of submissions, change control, business development activities, and maintaining global regulatory compliance. Danielle has also spearheaded a transition of an acquisition-related manufacturer to change all global product registrations, as well as chaired the operations-based team through the implementation of corresponding rebranding activities. Prior to MRC Global, Danielle was a Regulatory Affairs manager at Medtronic plc, where she managed regulatory teams and large-scale projects. Prior to Medtronic, Danielle was a Regulatory Affairs project manager at MicroPort Orthopedics and Wright Medical Technology. Danielle earned a bachelor of science degree in physics from Rhodes College and completed a master of science degree in biomedical engineering at the University of Memphis & University of Tennessee Health Science Center.