Mandating Registry-Based AI Surveillance: De-Risking Billions

The rapid evolution of artificial intelligence in healthcare presents an unprecedented opportunity to transform patient care, yet it simultaneously introduces complex regulatory challenges. The current pre-market clearance model, while strong for static medical devices, is fundamentally insufficient to guarantee the long-term safety and efficacy of clinical AI models that learn and adapt over time. These dynamic systems are susceptible to algorithmic drift, a degradation of AI model performance over time as real-world data distributions shift away from training data, potentially leading to compromised diagnostic accuracy or treatment recommendations. Federal regulators must therefore mandate the integration of algorithm performance tracking into existing clinical registries, establishing a strong, registry-based post-market surveillance framework.

The Inadequacy of Pre-Market Clearance for Adaptive AI

The traditional regulatory pathway, often exemplified by 510(k) clearance, focuses heavily on pre-market demonstration of substantial equivalence to a predicate device. This approach is well-suited for devices with fixed functionalities. However, most cardiac AI products are SaMD (Software as a Medical Device), and many are designed to be adaptive, continuously learning and improving from new data. The FDA’s Predetermined Change Control Plan (PCCP) framework offers a path for AI/ML devices to make predefined modifications without requiring new premarket submissions, acknowledging the dynamic nature of these technologies. Yet, even with a PCCP, the potential for unforeseen algorithmic drift remains a significant concern. The challenge lies in the inherent nature of AI. An AI model trained on data from 2018 to 2020 may perform optimally at the time of clearance, but by 2026, demographic shifts, changes in clinical practice, or evolving disease phenotypes can cause model performance to degrade. This algorithmic drift is a critical blind spot in a regulatory system primarily designed for static devices. Without continuous, real-world performance monitoring, the initial promise of an AI solution can erode, potentially leading to misdiagnoses, delayed treatments, or adverse patient outcomes. The percentage of cleared AI devices currently subject to mandatory post-market studies, and the compliance rates for FDA post-market study requirements, remain areas requiring more rigorous data collection and public transparency to fully assess the scope of this gap. FDA report on post-market surveillance for AI/ML devices

Using Clinical Registries for Continuous Monitoring

Our official position and recommendation is clear: the Food and Drug Administration (FDA) must mandate the integration of post-market surveillance for high-risk clinical algorithms into established clinical registries. These registries, often managed by specialty medical societies, offer a powerful and practical infrastructure for continuous performance tracking. Consider the American College of Cardiology (ACC). The ACC manages extensive clinical registries, such as the NCDR (National Cardiovascular Data Registry), which collect vast amounts of real-world evidence (RWE) on cardiovascular procedures, patient outcomes, and device performance. These registries already capture granular data points essential for evaluating clinical effectiveness and safety. By linking the performance of cardiac AI tools to these existing data streams, regulators and developers can gain unprecedented insights into how algorithms perform in diverse, real-world clinical settings over time. For instance, a cardiac AI tool designed to detect early signs of heart failure could have its diagnostic accuracy and impact on patient management tracked within an ACC registry. Any instances of algorithmic drift, where the AI’s performance deviates significantly from its validated baseline, could be flagged for immediate investigation. This approach moves beyond periodic updates or isolated studies, providing a continuous feedback loop that is essential for adaptive AI. The Pew Charitable Trusts has long advocated for stronger post-market surveillance for medical devices, emphasizing the value of integrating device performance data into existing clinical registries to enhance patient safety. Pew Charitable Trusts report on medical device registry integration

A Framework for Mandated Registry-Based Surveillance

We propose a model that links FDA post-market mandates directly to medical specialty registries. This framework would involve several key components:

• Pre-market Commitment to Post-market Surveillance: For high-risk clinical AI, particularly those with adaptive capabilities or novel functions necessitating De Novo classification, FDA clearance should be contingent on a strong, pre-defined plan for continuous post-market surveillance through an approved clinical registry. This would be a critical extension of the FDA Postmarket Surveillance Studies Program.
• Standardized Data Elements: Clinical registries would need to incorporate standardized data elements specifically designed to capture AI algorithm performance metrics. This includes input data characteristics, algorithm outputs, clinician overrides, and patient outcomes directly influenced by the AI’s recommendations.
• Performance Thresholds and Reporting: Clear performance thresholds would be established during the pre-market phase. Should an algorithm’s performance fall below these thresholds, or if significant algorithmic drift is detected, automated alerts would trigger mandatory reporting to the FDA and prompt corrective actions by the manufacturer.
• Collaboration between FDA, Specialty Societies, and Developers: This framework necessitates close collaboration. The FDA would provide regulatory oversight and set compliance standards. Specialty medical societies, like the ACC, would manage the registries, ensuring data quality and clinical relevance. AI developers would be responsible for integrating their algorithms with the registries and responding to performance anomalies.
• Transparency and Public Reporting: Aggregate, anonymized data on AI algorithm performance within registries could be made publicly available, fostering trust and enabling researchers to identify broader trends in AI efficacy and safety.

This methodology is grounded in established medical registry models and public safety advocacy papers. It recognizes that while an AI-native company may possess a significant data moat from its proprietary datasets, ensuring public trust and long-term safety requires a broader, independently monitored data ecosystem.

The Call to Action for Federal Regulators

This publication, Healthcare AI Compliance Watch, strongly urges FDA policymakers and clinical registry administrators to prioritize the development and implementation of this registry-based post-market surveillance framework. The current regulatory field, with its emphasis on pre-market clearance, leaves a critical vulnerability for algorithmic drift in high-risk clinical algorithms. By mandating the integration of AI performance tracking into existing, trusted clinical registries, the FDA can proactively safeguard patient safety, ensure the sustained efficacy of AI-powered medical devices, and foster responsible innovation in healthcare AI. This move would not only address the unique challenges posed by adaptive AI but also provide invaluable real-world evidence that can inform future regulatory guidance and best practices for GMLP (Good Machine Learning Practice). FDA guidance on Good Machine Learning Practice The time to act is now, to ensure that the far-reaching potential of healthcare AI is realized safely and ethically for all patients.