FDA Oversight of AI in Medicine

In the United States, the FDA regulates AI-enabled medical devices under its authority over Software as a Medical Device (SaMD). As of 2024, over 950 AI and machine learning-enabled medical devices had received FDA marketing authorization, with radiology accounting for the largest share.

The Three Main Regulatory Pathways

510(k) Clearance

The most common pathway. A manufacturer demonstrates substantial equivalence to a legally marketed predicate device. The device does not need to be clinically superior. Most AI radiology tools have entered the market through 510(k). Clearance does not mean the FDA has independently tested or validated clinical performance.

De Novo Authorization

Used for novel, low-to-moderate risk devices with no predicate. The FDA establishes new regulatory controls and a new device classification. De Novo creates a precedent that future similar devices can use as a predicate for 510(k).

Premarket Approval (PMA)

The highest-burden pathway, used for high-risk devices. Requires clinical evidence demonstrating safety and effectiveness. Few AI medical devices have gone through PMA.

What FDA Clearance Does and Does Not Mean

FDA clearance means a device has met a regulatory standard. It does not guarantee superior clinical performance, generalizability across all patient populations, or effective performance outside the conditions studied in the clearance submission.

Clinicians should be aware that:

  • Cleared devices may have been tested on datasets that differ from their own patient population
  • Post-market performance monitoring is the institution's responsibility
  • Intended use specifications in the clearance define where the device is and is not validated to perform

Locked vs. Adaptive AI Models

The FDA distinguishes between AI algorithms that are locked after training and those that continue to learn in deployment. Adaptive or continuously learning algorithms present additional regulatory complexity, and the FDA has issued draft guidance on these systems.