Regulatory Challenges and Opportunities: A Review of U.S. Food and Drug Administration-Approved Artificial Intelligence and Machine Learning-Enabled Cardiovascular Devices.

The integration of artificial intelligence (AI) and machine learning (ML) into medical devices has revolutionized healthcare, enhancing diagnostic accuracy and clinical decision-making. However, their rapid evolution poses challenges to traditional regulatory frameworks, particularly in ensuring safety and effectiveness. This review examines current Food and Drug Administration (FDA) regulatory pathways for AI/ML-enabled cardiovascular medical devices. It assesses the pre- and postmarket criteria evaluated by the FDA and discusses strengths and gaps in the regulatory pathway. We conducted a systematic review of FDA-authorized AI/ML-enabled cardiovascular devices from November 1995 to May 2025. Inclusion criteria required devices to utilize AI/ML technology, be FDA-authorized via 510(k) pathway, and be intended for cardiovascular applications. Data extracted included device purpose, primary users, indications, pre- and postmarket evaluations, and identified limitations. Out of 1,247 FDA-authorized AI/ML devices, we selected 96 cardiovascular devices. All were cleared through 510(k) pathway. Common premarket evaluations included clinical validation, bench testing, and algorithm performance assessments. However, postmarket surveillance requirements beyond baseline FDA regulations, Medical Device Reporting (MDR) and the Quality System Regulation (QSR), were rarely identified, especially before recent FDA guidelines. Key limitations included restricted patient demographics and lack of transparency in AI capabilities. Public summaries showed no evidence of autonomous, in-field model retraining without human oversight; changes were managed through existing mechanisms or proposed Predetermined Change Control Plans (PCCPs). FDA's framework has enabled broad authorization, but stronger transparency and device-appropriate postmarket monitoring tailored to AI/ML-specific risks (e.g., dataset shift, bias, evolving performance) would enhance long-term safety and effectiveness.