Only Three of 1,357 FDA-Cleared AI Medical Devices Tested for Patient-Centered Outcomes

The study by Abulibdeh et al. examined pre-authorization evidence for every AI device cleared under the 510(k) pathway, which requires only substantial equivalence to prior devices rather than new outcome data. Only 34 devices appeared in registered trials, with results posted for 12 and published for 12. Most trials occurred in high-resource settings and excluded pregnant patients, adults over 75, and non-English speakers. This pattern aligns with prior analyses of imaging AI, where surrogate metrics like diagnostic accuracy substituted for clinical endpoints.

Structural incentives favor rapid iteration over outcome trials. Developers face high costs and regulatory uncertainty for rigorous studies, while payers reimburse based on clearance alone. The original coverage understates downstream effects: devices cleared without equity data can widen outcome gaps when deployed in diverse populations. Low- and middle-income countries that reference FDA decisions may adopt tools whose performance remains untested outside narrow cohorts.

The authors propose a three-phase authorization model requiring effectiveness data across subgroups and settings. Related work in JAMA Network Open on radiology AI similarly documented sparse outcome evidence, reinforcing that current clearance standards prioritize speed over demonstrated benefit. Next steps include congressional review of the 510(k) framework for software and potential FDA guidance mandating outcome-linked post-market studies.

Evidence quality is limited to one observational mapping of trial registries without independent verification of posted results; randomized trials with patient-centered endpoints remain absent for nearly all devices.