AI medical device clearance skips outcomes data

Clinical takeaway: Treat FDA clearance as a floor, not evidence of patient benefit. Almost all authorized AI medical devices have been validated solely on accuracy or other surrogate measures. 

When an AI tool arrives in the workflow with a regulatory credential attached, most clinicians reasonably read that as a sign that regulators have evaluated it rigorously. But a minimal regulatory pathway built to compare new devices against products already on the market is now the main way AI is assessed before it reaches the clinic. 

FDA authorized almost 1,400 AI-enabled devices during the period examined, with more than 1,000 of those in radiology, and clinical decisions now routinely incorporate their output. Most entered with a 510(k) clearance, which requires only that a manufacturer show a new product is much like one already sold, rather than to demonstrate clinical benefit. A new regulatory evidence census traced every AI/ML-enabled device authorized by FDA to its registered trials and published outcome data. 

Of 1,357 authorized devices, 34 were linked to a registered prospective trial. Results reached ClinicalTrials.gov for 12 devices, with peer-reviewed publication for the same number. Only three were evaluated on outcomes patients experience directly, such as mortality, major morbidity, or readmission. That leaves 97.5% entering clinical use with no registered prospective trial behind them. Among the 34 trials that do exist, 82% used diagnostic accuracy or similar technical endpoints, and 94% were run by the device manufacturer. 

Nearly three-quarters of these trials enrolled fewer than 500 patients and 68% ran only in the US. Just 27% reported any subgroup analysis, and race or ethnicity was reported in three trials. Pediatric patients were excluded almost universally, pregnant patients from 42% of cardiovascular and 33% of radiology trials, and non-English speakers and patients with cognitive impairment were frequently omitted. 

The census covered every AI/ML-enabled device in the FDA device database and the American College of Radiology's AI Central catalogue through December 5, 2025, cross-referenced against ClinicalTrials.gov registrations linked from 510(k) summaries and against PubMed. Because it counts only what is publicly registered, it undercounts validation manufacturers ran but never disclosed. 

Health systems hold the realistic, near-term lever here. The authors point to institutional AI formularies that vet tools against local standards and track performance after deployment. Longer term, they propose that FDA adopt staged evidence requirements modeled on drug development. This would include representative retrospective validation before clearance, prospective workflow-embedded studies around it, and multicenter outcome trials after. 

"AI tools must be life-tested before they can be called life-saving," the authors conclude. "Regulatory clearance has far outpaced clinical validation, creating an ecosystem where innovation advances without accountability and patients bear the risks of unproven technologies." 

Source: Abulibdeh R, et al. (2026 Aug 19) PLOS Digit Health. 1,357 AI medical devices cleared, 3 actually tested on patient outcomes