Adoption of Predetermined Change Control Plans increased among radiology artificial intelligence–enabled devices after final US Food and Drug Administration guidance was issued, although publicly available regulatory summaries often lacked details on postmarket performance monitoring, according to a systematic scoping review published in Radiology: Artificial Intelligence.
Investigators evaluated FDA-cleared radiology artificial intelligence and machine learning (AI/ML) devices from 2015 to 2025 to characterize use of Predetermined Change Control Plans (PCCPs) and the transparency of public documentation. The systematic scoping review followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews guidance and was registered in PROSPERO. FDA 510(k), de novo, premarket approval, and breakthrough device databases were accessed April 1, 2026, and PCCP devices were manually verified against regulatory summaries and an independent registry.
Investigators assessed PCCP documentation using a prespecified 8-point rubric covering specificity, data and validation transparency, real-world evaluation, and postmarket monitoring. Two independent observers scored the records, with an intraclass correlation coefficient of 0.93. The score reflected completeness of public disclosure rather than independent assessment of device performance or field monitoring.
Among 1,394 FDA-listed AI/ML submissions, 1,080 (78%) were in radiology, including 1,068 (99%) cleared through the 510(k) pathway. Among 870 unique radiology devices, 130 (15%) underwent sequential 510(k) clearances for significant device modifications. The mean interval between sequential submissions decreased from 25 months prior to 2021 to 13 months following 2021.
Across FDA panels, 170 devices were identified as PCCP-cleared from 2020 to 2025. Of 37 radiology devices cleared with PCCPs, 34 (92%) were AI/ML devices, and 22 of those 34 (65%) were cleared in 2025. Manual correction was required for 9 of 34 devices because PCCP classifications differed between the public FDA database and individual regulatory summaries.
Public documentation had a mean score of 5 on the 8-point rubric. Clearly defined modification protocols were documented for 26 of 34 devices. Public summaries for only 3 devices described postmarket surveillance with continuous monitoring, and 1 referenced real-time updates to clinical sites regarding performance drift. Although mean documentation scores increased from 4 in 2024 to 5 in 2025, the analysis did not find evidence that documentation completeness differed according to whether submissions occurred before or after final PCCP guidance.
The investigators cautioned that the analysis was limited by publicly available data, with internal safety assessments and minor software changes inaccessible. Other limitations included the recency of final FDA guidance, database inconsistencies, challenges deduplicating devices because of corporate rebranding and incremental name changes across versions, and use of a transparency rubric that was not designed as a comprehensive quality evaluation.
The investigators reported no study funding. Full disclosures can be found in the published study.
