Study Finds 3 of 1,357 FDA-Cleared AI Devices Were Tested on Patient Outcomes

Only three of the 1,357 artificial intelligence-enabled medical devices cleared or approved by the U.S. Food and Drug Administration through December 5, 2025, have publicly documented evaluations of patient-centered outcomes such as mortality, morbidity or readmissions, according to an analysis published August 19 in PLOS Digital Health.

The paper, by Rawan Abulibdeh and colleagues at institutions including the University of Toronto, MIT, Johns Hopkins University and Uganda's Mbarara University of Science and Technology, is titled "1,357 AI medical devices cleared, 3 actually tested on patient outcomes." The authors searched the FDA device database and the American College of Radiology Data Science Institute catalogue, then linked each device to ClinicalTrials.gov registrations and PubMed publications.

A thin trail of trials

Of the 1,357 devices, 34 (2.5%) were linked to registered prospective trials, 12 (0.9%) had posted trial results, 12 (0.9%) had peer-reviewed publications and three (0.2%) evaluated patient-centered outcomes, the researchers reported. Most of the studies they found (62%) used observational designs.

Radiology accounted for 78% of cleared devices (1,059 of 1,357) but had prospective trials for fewer than 1% of its tools, according to the paper. Cardiovascular and neurology devices had somewhat higher shares of registered trials, at 9.5% and 9.7%. Anesthesiology had 22 cleared devices and no registered prospective trials.

Among the 34 registered trials, nearly three-quarters enrolled fewer than 500 participants and one-quarter fewer than 100. Sixty-eight percent were run only in the U.S., and nine (27%) reported any subgroup analysis. Many excluded groups such as pregnant patients, children, older adults and non-English speakers, the authors wrote.

They attributed the gap to structural barriers including misaligned financial incentives, reliance on predicate-based regulatory pathways and the logistics of multi-center trials. The 510(k) pathway asks a company to show that a new device is substantially equivalent to one already on the market, not to prospectively validate its clinical effectiveness, the researchers noted, as reported by Healio.

Proposals for regulators and payers

The authors set out three proposals: prospective trials measuring patient outcomes rather than algorithmic accuracy, with the FDA mandating pre-registration of such trials for all Class II and III AI devices before clearance; validation in the populations where tools will be used, with the World Health Organization requiring local validation before international deployment; and transparency, with CMS and insurers tying reimbursement to demonstrated clinical benefit rather than regulatory status.

The paper acknowledges limits. Because it relies on public registries, it may undercount validation studies that manufacturers ran internally and did not disclose, and not every FDA 510(k) summary was available in full text, the authors wrote.

The study appeared as the FDA seeks feedback on a potential approach to regulating AI-enabled devices, Healio reported. In a statement to Healio, which reported the findings on August 21, the agency said it intends to work collaboratively on "efficient, scientifically sound and least burdensome approaches" to evaluating generative AI-enabled medical devices across the total product life cycle. Comments on the discussion paper are due October 19.