Scoping review on AI fundus photography for Type 2 Diabetes highlights limited external validation and calibration

This scoping review synthesizes evidence regarding artificial intelligence models utilizing fundus photography within the context of Type 2 Diabetes Mellitus. The authors mapped the current landscape of AI applications for retinal screening, noting that sample sizes and settings were not reported across the included literature. The scope encompasses model performance metrics, validation strategies, and equity considerations relevant to clinical deployment.

Key findings indicate that AI models frequently demonstrated promising discrimination capabilities. However, the review highlights significant gaps in robustness. External validation was limited across studies, and calibration was inconsistently assessed. Furthermore, subgroup analyses addressing fairness and device-related domain shift were rarely reported. Most studies emphasized discrimination metrics without comprehensive evaluation of clinical utility.

The authors acknowledge substantial heterogeneity in cohort size, outcome definitions, imaging modalities, and validation strategies as a major limitation. This variability complicates direct comparisons and generalizability. The certainty of evidence is significantly reduced by these inconsistencies and the lack of standardized reporting frameworks.

Practice relevance suggests retinal AI shows potential for scalable systemic risk surveillance in Type 2 Diabetes. However, rigorous external validation, standardized reporting, and prospective implementation studies are required to enable safe and equitable clinical translation. Clinicians should interpret current findings cautiously until these methodological gaps are addressed in future research.