To map multimodal AI applications and methods in diabetes care, assess evidence comparing multimodal models with unimodal baselines, and summarize study limitations and implementation barriers.
Approach:
Design and reporting: Protocol registered with the Open Science Framework; review conducted according to PRISMA-ScR.
Search: Searched PubMed, Web of Science, Scopus, The Cochrane Library, CNKI, and Wanfang Data from database inception through November 4, 2025. English- and Chinese-language publications and grey literature were eligible.
Selection and extraction: Studies were screened using prespecified Population, Concept, and Context criteria. Two reviewers independently screened records and extracted data; screening agreement was high (Cohen’s κ = 0.94).
Quality assessment: Included studies were assessed with PROBAST+AI across participant/data, predictor, outcome, and analysis domains.
Key Findings:
The supplied context describes the review rationale and methods but does not report included-study counts, application findings, performance comparisons, or quality-assessment results.
Interpretation:
The review was designed to organize a fragmented literature on multimodal AI in diabetes, including its clinical applications, technical methods, comparative performance, interpretability, and barriers.
Limitations:
The provided article context ends during the quality-assessment methods and contains no results or discussion.
Eligibility was restricted to English- and Chinese-language publications.
