PolicyEthicsAI01 sources

Three of 1,357 cleared AI devices were tested on patient outcomes

A line chart from the paper rising steeply from near zero in 1995 to about 1,357 FDA-approved AI and ML devices by December 2025, with a flat red line along the bottom marked 34 linked to registered trials.

Abulibdeh et al., PLOS Digital Health 2026 (CC BY 4.0)CC BY

Rawan Abulibdeh, Leo Anthony Celi and colleagues went through every AI/ML-enabled medical device the FDA had cleared or approved as of 5 December 2025 and looked for evidence that any of them helped a patient. The paper was published in PLOS Digital Health on 19 August 2026 (doi 10.1371/journal.pdig.0001597).

Of 1,357 devices, 34 — 2.5% — were linked to a registered prospective clinical trial. Twelve posted results. Twelve had a peer-reviewed publication. Three, 0.2%, were evaluated against a patient-centered outcome such as mortality, morbidity or readmission.

The authors put the reason plainly: the 510(k) pathway asks whether a device is substantially equivalent to one already on the market, not whether it changes what happens to the person being scanned. Equivalence is a claim about the device. It carries no claim about the patient.

The validation that does exist is also narrow. The studies clustered in well-resourced settings and excluded pregnant women, people over 75 and non-English speakers — groups that will nonetheless be run through the same software.

It lands the same week the FDA opened its docket on generative AI devices, on a framework whose evidentiary floor this paper is measuring.

Sources

  1. [1]1,357 AI medical devices cleared, 3 actually tested on patient outcomesPLOS Digital Health··Paper