Multimodal artificial intelligence in diabetes care: A scoping review

Objective Diabetes is a complex multisystem metabolic disorder characterized by heterogeneous and multidimensional data. Multimodal artificial intelligence (AI) offers opportunities to integrate diverse information for precision health. This scoping review aimed to systematically characterize the current applications, methodological approaches, and performance of multimodal AI in diabetes. Methods Following the PRISMA-ScR framework, we systematically searched PubMed, Web of Science, Scopus, the Cochrane Library, CNKI, and Wanfang Data for eligible studies investigating multimodal AI applications in diabetes. Data on clinical applications, data modalities, AI architectures, fusion strategies, explainability, and model performance were extracted and synthesized descriptively. Results Thirty-six studies were included. Multimodal AI applications spanned five core domains, with intermediate fusion being the most commonly adopted strategy for integrating heterogeneous data. Comparative evidence generally favored multimodal over unimodal approaches, with average relative performance gains of 11.2% in discriminative tasks and 27.4% in dynamic glycemic prediction. However, performance gains varied across tasks, and the current evidence base was characterized by limited external validation and substantial methodological heterogeneity. Conclusions Multimodal AI shows promising potential across multiple domains of diabetes care and generally provides incremental predictive value over unimodal approaches. Future research should prioritize robust external validation, standardized multimodal benchmarks, and clinically feasible integration strategies to facilitate the translation of multimodal AI from task-specific prediction toward comprehensive and personalized diabetes management.