1,500 FDA-cleared AI medical devices, but clinical evidence lagging behind

Radiology in the lead with 76% of authorisations, massive hospital adoption: clinical evaluation is struggling to keep pace with market launches.

More than 1,500 medical devices incorporating artificial intelligence had FDA approval as of April 2026, 68% of which have been granted since 2022 alone. The distribution by specialty is very unbalanced: radiology accounts for 76% of authorizations, ahead of cardiology (10%) and neurology (4%).

Adoption is following the same pace. According to the Office of the National Coordinator for Health Information Technology, 71% of U.S. hospitals reported using predictive AI integrated into their electronic patient records in 2024, up from 66% the previous year.

The gap between authorization and demonstration

The problem lies in the nature of the evidence required. Authorization is most often based on technical performance measured on datasets, not on a demonstrated benefit for the patient. The example of Epic's sepsis detection model has become emblematic: a sensitivity of 33%, which allows two out of three cases to pass, and a positive predictive value of 12%, i.e. seven false alarms for one relevant alert. A practitioner can now cross-reference, in a single shift, tools that report a septic risk, predict rehospitalization, adjust a dosage and triage imaging examinations — without the net effect on mortality or quality of life having been established.