The evidence narrows at every step

The authors treated the project as a regulatory evidence census rather than a meta-analysis of treatment effects. They matched FDA and radiology-catalogue records through December 5, 2025 with ClinicalTrials.gov and PubMed to identify prospective trials, posted results, publications, and patient-centered outcomes.

Of 1,357 authorized devices, 34 were linked to registered prospective trials. Twelve posted results, 12 had peer-reviewed publications, and three evaluated outcomes such as death, illness, or readmission. Most identified studies were observational, and the review reports small, relatively homogeneous cohorts, limited subgroup analysis, and frequent exclusion of vulnerable populations.

Clearance answers a narrower question

Many AI devices enter the U.S. market through the 510(k) pathway, which focuses on substantial equivalence to a predicate device rather than requiring a new prospective demonstration of clinical effectiveness. That pathway can be appropriate for market authorization while leaving a separate evidence question unresolved.

A tool may classify an image accurately, improve a workflow, or meet regulatory requirements without proving that its use improves patient health. The review does not infer that every device without a linked outcome study is harmful or useless. It documents how rarely the public record follows authorization through to clinically meaningful outcomes.

Global deployment can export the gap

FDA authorization often functions as an international signal. Health systems in countries with fewer evaluation resources may adopt products without local validation, even when patient populations, clinical workflows, disease patterns, connectivity, and follow-up capacity differ from the environment in which a model was developed.

Evidence standards should therefore follow the device after authorization: prospective outcome studies, representative subgroup analysis, transparent post-market surveillance, and contextual validation where the system will actually be used. Readiness should mean demonstrated patient benefit, not permission alone.

Primary trail

Go to the source

Read the evidence behind this analysis. External links open in a new tab.

PLOS Digital Health — Evidence review of 1,357 AI medical devices U.S. Food and Drug Administration — Artificial intelligence-enabled medical devices