Of the 1,357 AI-enabled medical devices authorised by the FDA, only three have been assessed for their impact on patient health

A survey published in PLOS Digital Health shows that the vast majority of medical AI systems authorised in the United States have never been tested for survival, complications or quality of life.

A team led by Rawan Abulibdeh (University of Toronto) and researchers from MIT Critical Data reviewed, in PLOS Digital Health, all AI-based medical devices authorized by the FDA until December 5, 2025. Of the 1,357 products, of which 1,059 were devoted to imaging (mammography, scanners), only 34 had one trial recorded on ClinicalTrials.gov, and only three had been evaluated on criteria that count for patients: mortality, complications or quality of life.

Small, American and unrepresentative trials

Existing trials are 94% conducted by manufacturers themselves, 68% conducted only in the United States, and nearly three quarters involve less than 500 participants. Children are excluded from all X-ray and cardiology trials, pregnant women have a significant share of them, and only nine trials detail their results according to demographic characteristics. The authors recall that equivalency authorization only means that the device looks like a product already marketed, not that it improves management.

They propose a three-step course: prior validation on representative data, a clinical study of at least 500 patients at the time of authorisation, and a multi-centre post-marketing trial on more than 2,000 patients, with results disaggregated by population.