Authors
Yang Xu, Jeremy Veillard, Jude Dzevela Kong
Published in
Journal of medical Internet research. Volume 28. Pages e95157. Sep 21, 2026. Epub Sep 21, 2026.
Abstract
AI systems have achieved strong performance on discrete diagnostic, predictive, and decision support tasks; yet, translation into sustained clinical benefit remains uneven. Existing reviews often examine single application areas, obscuring barriers that operate across the health care AI ecosystem.
This review aimed to map the published evidence on AI for health care quality and patient safety across application domains, characterize evidence maturity, and identify cross-domain barriers that limit translation into clinical value.
We conducted a scoping review following Joanna Briggs Institute methodology and reported it according to PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews), with search documentation guided by PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension). Five databases (MEDLINE via PubMed, Scopus, Web of Science Core Collection, IEEE Xplore, and CINAHL Plus with Full Text) were searched for English-language records published from January 1, 2017, to April 30, 2026. Eligibility followed the population-concept-context framework. Application domains were coded nonmutually. World Health Organization (WHO) quality of care dimensions served as a deductive scaffold, and recurring constraints were grouped into cross-domain barriers. Methodological characteristics were charted descriptively, informed by validation approach, implementation context, and established AI reporting frameworks.
The search identified 43,394 records, of which 275 were retained in the core charting corpus after staged screening. The evidence covered 4 nonmutually coded domains: diagnostic AI (n=142, 51.6%), predictive analytics (n=192, 69.8%), clinical decision support or implementation-related applications (n=233, 84.7%), and economic or value assessment (n=53, 19.3%). It was broad but uneven in maturity, with the same gaps across all 4 domains. Prospective and external validation were uncommon, and retrospective results often lacked evidence of improved care. Workflow or process measures were recorded in only 13 (4.7%) records, and explicit equity or subgroup analyses were recorded in 20 (7.3%) records, indicating limited measurement of real-world impact. Economic and governance evidence was also thin, with formal economic evaluations found in 4 (1.5%) records and little attention to monitoring models after deployment. These gaps indicate limited evidence on how AI performs in clinical practice.
We identified 5 recurring barriers to translating health care AI into clinical value: limited prospective external validation, workflow-mediated effectiveness, infrastructure fragmentation, equity and generalizability deficits, and economic and governance uncertainty. Unlike domain-specific reviews, this review maps evidence across the 4 domains, locates the barriers within the WHO quality of care dimensions, and distinguishes descriptive findings from interpretive synthesis. The binding constraint on clinical value is no longer primarily technical but structural, recurring across domains. Health systems and regulators should therefore treat prospective external validation, workflow-integrated deployment, infrastructure readiness, equity monitoring, and lifecycle governance as prerequisites for responsible adoption.
PMID:
42766289
Bibliographic data and abstract were imported from PubMed on 22 Sep 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 13
- Comments 0