Clinical Specialty Expansion of AI-Enabled and Machine Learning-Enabled Medical Devices Authorized by the US Food and Drug Administration From 1995 to 2025: Longitudinal Content Analysis.
Authors
Affiliations (2)
Affiliations (2)
- Department of Digital Healthcare, College of Health and Medical Science, Daejeon University, #505 Moonmugwan 62, Daehak-ro, Dong-gu, Daejeon, Republic of Korea, +82-42-280-2062.
- Department of Healthcare Management, College of Health and Medical Science, Daejeon University, #505 Moonmugwan 62, Daehak-ro, Dong-gu, Daejeon, 34520, Republic of Korea, +82-42-280-2062.
Abstract
The US Food and Drug Administration (FDA) has authorized AI-enabled and machine learning (ML)-enabled medical devices since 1995 and maintains a public registry of these authorizations. Prior analyses report that radiology dominates this landscape, but whether that concentration has persisted, intensified, or begun to reverse across 3 decades, particularly since 2022, remains insufficiently characterized. This study aimed to (1) characterize the longitudinal growth of FDA-authorized AI/ML-enabled devices from 1995 to 2025, (2) quantify the temporal evolution of clinical specialty distribution across 4 eras, (3) identify emerging specialties, and (4) examine the association between manufacturer type and nonradiology authorization. All 1430 devices in the FDA AI-Enabled Medical Devices registry (downloaded on March 1, 2026) with final marketing-authorization decisions through December 31, 2025, were analyzed. Devices were stratified by clinical specialty (FDA advisory committee panel) and 4 eras: Era 1 (1995-2015), Era 2 (2016-2019), Era 3 (2020-2022), and Era 4 (2023-2025). Concentration was quantified using the Herfindahl-Hirschman Index (HHI) with bootstrap CIs; the Cochran-Armitage test assessed trends in specialty share, with Bonferroni correction. Multivariable logistic regression estimated the odds of nonradiology authorization by manufacturer type and era, with an era-by-manufacturer interaction term. Sensitivity analyses used cluster-robust standard errors, a continuous authorization year variable, and Firth penalized regression. Manufacturers were classified using FDA records, Crunchbase, PitchBook, and company websites. Annual authorizations rose from a mean of 2.0 (SD 2.0) in Era 1 to a mean of 264 (SD 58.2) in Era 4, with 331 authorizations in 2025 alone; the 510(k) pathway accounted for 96.2% (1376/1430). Radiology led in every era but followed a nonmonotonic trajectory, rising from 35.7% (15/42, Era 1) to a peak of 85.5% (347/406, Era 3) before declining to 77.5% (614/792, Era 4), the first significant decline on record (<i>P</i>=.001). The HHI fell from 0.738 (Era 3) to 0.612 (Era 4; bootstrap <i>P</i><.001), indicating measurable diversification. Specialty distribution was associated with era (<i>χ</i>²<sub>48</sub>=328.0; <i>P</i><.001; Cramér <i>V</i>=0.28). Compared with incumbents, start-ups (odds ratio [OR] 5.09, 95% CI 3.33-7.79) and technology companies (OR 50.62, 95% CI 12.90-198.64) had higher odds of nonradiology authorization; the nonsignificant era-by-manufacturer interaction (likelihood ratio test <i>χ</i>²<sub>8</sub>=11.16; <i>P</i>=.19) indicates a persistent rather than widening effect. The technology-company OR derives from only 13 devices across 5 firms; although directionally robust in sensitivity analyses, it is imprecise and warrants cautious interpretation. Radiology remained dominant, accounting for 77.5% (614/792) of Era 4 authorizations, but the specialty distribution showed measurable diversification during 2023 to 2025, associated with start-up and technology-company activity. Maturation of clinical data infrastructure beyond imaging is a plausible but unmeasured contributing condition, and authorization is not adoption. The findings bear on health-system readiness, workforce training, and specialty-specific regulatory frameworks.