FDA AI Medical Device Approvals Hit 521 — Complete Breakdown by Specialty

The FDA’s count of AI and machine learning-enabled medical devices passed 521 in 2023, a figure that tells only part of a more complicated regulatory story. The raw number reflects a decade of accelerating submissions through 510(k) and De Novo pathways, but the distribution across medical specialties, device types, and risk classifications reveals where the regulatory framework has found its footing — and where it has not.

How the FDA Classifies AI Devices

The FDA regulates AI-enabled medical devices under the broader Software as a Medical Device (SaMD) framework, which the agency has developed in alignment with international guidance from the International Medical Device Regulators Forum. An AI/ML device is subject to FDA oversight when it meets the definition of a medical device — meaning it is intended to diagnose, cure, mitigate, treat, or prevent a disease or condition — and when it incorporates an AI or machine learning component as part of its primary function.

Most AI/ML devices have been cleared through the 510(k) pathway, which requires a showing of substantial equivalence to a legally marketed predicate device. A smaller number have gone through De Novo — a pathway for novel, low-to-moderate risk devices without an appropriate predicate — and an even smaller number have received full Premarket Approval (PMA), reserved for high-risk devices where no predicate exists and independent demonstration of safety and effectiveness is required.

The 521 figure the FDA published represents all AI/ML-enabled devices that had received marketing authorization by that point, including 510(k) clearances, De Novo authorizations, and PMA approvals. It does not reflect the number of devices currently active in the market, which is lower due to voluntary withdrawals and product discontinuations.

Specialty Distribution

Radiology dominates the cleared device landscape by a substantial margin. Approximately 75% of all cleared AI/ML devices fall under the radiology specialty, reflecting the natural alignment between image-based AI and the large, digitized datasets available in medical imaging. Within radiology, the most common intended uses include chest X-ray analysis, CT scan triage, mammography computer-aided detection, and brain MRI segmentation.

Cardiology accounts for roughly 10% of cleared devices, covering applications including ECG interpretation, echocardiogram analysis, and atrial fibrillation detection. The remaining 15% is distributed across pathology, ophthalmology, gastroenterology, and general and plastic surgery — a long tail of applications that spans a wide range of clinical tasks.

• Approximately 75% of cleared AI devices are in radiology
• Cardiology accounts for roughly 10%
• The remaining 15% spans pathology, ophthalmology, gastroenterology, and surgery
• Most devices cleared through 510(k); De Novo and PMA used for novel or high-risk applications

The Action Plan and What It Changed

The FDA’s January 2021 AI/ML-Based Software as a Medical Device Action Plan outlined the agency’s intent to develop a regulatory framework capable of handling adaptive AI algorithms — systems whose behavior changes as they are exposed to new data after market clearance. Traditional device regulation assumes a locked software specification; a model that continues to learn after deployment challenges that assumption fundamentally.

The action plan identified five areas of focus: good machine learning practices, patient-centered approach incorporating AI transparency, regulatory science for algorithm evaluation, real-world performance monitoring, and regulatory frameworks that address the unique challenges of adaptive AI. The 2023 Marketing Submission Recommendations for AI/ML-Enabled Device Software Functions document operationalized several of these commitments, providing sponsors with more explicit guidance on what the FDA expects in submissions involving AI components.

Gaps and Unresolved Questions

The concentration of cleared devices in radiology reflects the availability of large labeled imaging datasets rather than the breadth of clinical need for AI tools. Specialties where AI could have significant impact — such as primary care, mental health, and emergency medicine — have relatively few cleared devices because the underlying data infrastructure and standardization required for regulatory submission remain underdeveloped.

Post-market surveillance for AI devices also remains a developing area. Real-world performance of an algorithm trained on one hospital system’s data may differ significantly when deployed at a geographically and demographically distinct institution. The FDA’s frameworks for monitoring and responding to performance drift are still being refined.

Key Takeaway

521 FDA-cleared AI devices signals a maturing regulatory pipeline, but the 75% concentration in radiology reflects data availability more than clinical need — and post-market surveillance frameworks for adaptive algorithms remain the most significant unresolved challenge in health AI regulation.