Copy to clipboard 
In response to the constantly evolving fields of Artificial Intelligence (AI) and Machine Learning (ML) in medical devices, FDA has struggled to produce a regulatory framework that keeps pace with these fast-moving technological advancements. However, earlier this month, the Agency released a five-part action plan meant to update the proposed regulatory framework that it published in 2019.
FDA’s action plan comprises five relatively brief but direct responses to stakeholder feedback on the Agency’s initial paper. Further guidance is expected to be published later in 2021, according to FDA. The following summary outlines the five points explained in the action plan paper:
1. Tailored Regulatory Framework for AI/ML-based SaMD
This action is based off stakeholders’ suggestions to further develop and clarify the regulatory framework for AI/ML-based Software as a Medical Device (SaMD), which includes the Predetermined Change Control Plan. In response to the suggestions, FDA has committed to updating the proposed framework for AI/ML-based SaMD, which includes offering Draft Guidance on the Predetermined Change Control Plan.
FDA reported stakeholder support for the Agency’s approach to regulating algorithms that learn and change over time, and offered specific feedback for improving safety and effectiveness in this area. While there was a generally positive consensus relating to modifications to AI/ML software devices, additional modifications were suggested for inclusion within the framework. FDA said that it would create draft guidance for public comment that proposes what should be included when a technology has pre-specification (SPS) that describe “what” aspects the manufacturer intends to change through learning and algorithm change protocol (ACP) that explain “how” the algorithm will learn and change, while remaining safe and effective.
The Agency set the goal of releasing Draft Guidance within the year, but did not commit to a specific timeline. It did commit to refining the identification of modifications that would fall under the framework and specifics regarding the focused review. The submission and review process and the content of submissions are included.
2. Good Machine Learning Practice (GMLP)
In response to the desire to strengthen GMLP through consensus standards and community initiatives, FDA committed to encourage harmonization within best practices in AI and ML. This applies to data management, feature extraction, training, interpretability, evaluation and documentation according to the action plan.
Within the plan, FDA touted its connection with groups such as the Institute of Electrical and Electronics Engineers and the Xavier AI World Consortium, International Medical Device Regulators Forum’s IMDRF Artificial Intelligence Medical Devices Working Group and others. It committed to furthering its knowledge of AI and ML technologies by deepening its work with these communities.
3. Patient-Centered Approach Incorporating Transparency to Users
FDA announced that it would continue discussions based on how patients interact with AI and ML technologies and product transparency. It committed to holding a workshop that will explore how device-labeling supports transparency for users and builds trust in AI/ML-based devices.
In acknowledging unique patient issues and concerns involved with AI and ML’s integration into medical devices, such as usability, equity, trust and accountability, FDA doubled down on its commitment to transparency with patients, and highlighted the importance of showing the public how these technologies work. “The Agency is committed to supporting a patient-centered approach including the need for a manufacturer’s transparency to users about the functioning of AI/ML-based devices to ensure that users understand the benefits, risks, and limitations of these devices,” read a line from FDA’s action plan.
4. Regulatory Science Methods Related to Algorithm Bias & Robustness
The administration announced efforts for identifying and addressing bias in algorithms and to increase robustness. “Because AI/ML systems are developed and trained using data from historical datasets, they are vulnerable to bias – and prone to mirroring biases present in the data,” FDA explained. Recognizing the wide disparities that exist in healthcare based on a person’s race, gender and socioeconomic status, FDA stressed the importance of AI- and ML-driven medical device technology to be accessible to and designed for “a racially and ethnically diverse intended patient population.”
In addition to removing bias from these technologies, FDA said that current technologies needed to be robust and resilient enough to handle shifting clinical inputs and conditions. To aid in these efforts, the administration announced the support of numerous regulatory science research efforts, including collaborations with the Centers for Excellence in Regulatory Science and Innovation (CERSIs) at the University of California San Francisco (UCSF), Stanford University and Johns Hopkins University.
5. Real-World Performance (RWP)
FDA committed to collaborating with stakeholders who are piloting the RWP process for AI/ML-based SaMD. It highlighted the importance of collecting and monitoring real-world data to help AI and ML medical device manufacturers “understand how their products are being used, identify opportunities for improvements, and respond proactively to safety or usability concerns,” and said it was an effective way to mitigate risk. Responding to questions and concerns on the subject, FDA announced that it would work to support RWP monitoring with stakeholders on a volunteer basis. It plans to construct a framework for real-world data-gathering with information gleaned from the project. A vague public engagement initiative was also announced by the administration.
While the orthopedic device industry waits for FDA to produce a comprehensive regulatory framework, it’s moving ahead with AI and ML technologies in anticipation of future adoption.
“We needed more advanced computational tools,” said Peter Verrillo, CEO of the surgery startup Enhatch, “so we pursued different technologies that would help us understand human anatomy. Every five years, developers have about 100x more computational power available, so that opens up areas that we were never able to explore before.”
AI’s ability to comb through vast assortments of data and develop insights based on what it observes is what makes it such a valuable asset in orthopedics. Complex and laborious tasks that used to take weeks are now being accomplished “in seconds,” according to Verrillo.
