Ä¢¹½tv

Evolving guidance documents & industry support focus on risk management

Throughout 2022, the Food and Drug Administration reviewed and authorized more than 90 new AI/ML-enabled devices, tracked through a public database that, although not comprehensive, logs devices that have received 510(k) clearance, been granted a De Novo request, or undergone pre-market approval.

, FDA recognized its first AI-focused document, , adding it to the federal agency's recognized consensus standards for meeting medical device requirements under the Federal Food, Drug, and Cosmetic Act. According to the Association for the Advancement of Medical Instrumentation (AAMI), FDA recognition is considered a critical endorsement as the agency moves to issue an AAMI technical information report (TIR) later this year and a British National Standard. FDA liaisons play a role in AAMI standardization efforts as part of both agencies' action plans.

To help software developers, quality managers, and regulatory affairs professionals understand the fast-moving regulatory landscape for AI/ML medical devices, AAMI hosted , conducted by the co-chair of AAMI's new AI standards committee. The 2-hour course provided an overview of the agency's guidance document regarding risk management for ML systems and a proposed process for bias management. It also discussed the current regulatory landscape for AI/ML medical devices, including the FDA, EU, and China's National Medical Products Administration, and the efforts underway with the International Medical Device Regulators Forum. AAMI will also host on the proposed European Artificial Intelligence Act and its impact on regulations for medical device manufacturers.

These recent developments build on the July 2021 Consensus Report published by AAMI on the application of its revised standard ISO 14971:2019 to medical devices using artificial intelligence/machine learning. In Consensus Reports, a select group of stakeholders offer guidance on a focused topic of high importance to the health technology community.

Both FDA and AAMI have acknowledged the urgent need for reviewing agencies to to AI/ML medical device lifecycles, which can autonomously adapt to new data sets and produce software results that could drift over time and training scenarios. With appropriate risk management procedures in place, implementing AI/ML algorithms in medical devices can be a powerful tool to improve patient outcomes; however, they also pose unique challenges, such as:

• Training, testing, and validating algorithms —  Developing the algorithm with medical, data science, and patient population stakeholders. For instance, a machine learning application for breast cancer detection could produce misleading results if the algorithm's development does not include training on breast tissue density variations within a population.
• Device interoperability and IT system communication —  Cybersecurity incidents can be difficult to detect due to the nature of many AI/ML algorithms. Traditional risk analysis does not always fit neatly within AI/ML systems and makes harm estimations more complex to evaluate. FDA also issued 2022 guidance on cybersecurity in medical devices.
• Unintended data selection bias from unrepresentative data, incomplete information — For example, an algorithm for pneumonia risk could inadvertently identify asthma patients as low risk by treating their frequent prior medical visits as effective, proactive health interventions.
• Adapting to seasonal differences, unique events —  Adaptive algorithms that continuously learn may produce ineffective modeling based on seasonal variations in ER visits or special circumstances, such as a natural disaster.