Clinical setting-dependent diagnostic accuracy of artificial intelligence and store-and-forward diabetic retinopathy screening: a systematic review and meta-analysis

Abstract

Population-based diabetic retinopathy (DR) screening requires diagnostic strategies that optimize clinical utility by balancing missed disease against referral burden. We performed a Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy studies (PRISMA-DTA)-guided systematic review and meta-analysis comparing autonomous artificial intelligence (AI) screening with store-and-forward (SAF) or conventional image-based teleophthalmology pathways, using manual, expert, or reading-center grading as the reference standard, across any DR, referable DR (RDR), vision-threatening DR (VTDR), and diabetic macular edema (DME). Twenty-eight diagnostic accuracy studies were included. AI showed higher pooled sensitivity than SAF for any DR (86.9% vs 80.9%), RDR (96.2% vs 88.6%), VTDR (96.2% vs 84.2%), and DME (97.2% vs 87.4%). AI also showed higher pooled specificity for any DR, RDR, and VTDR, whereas DME specificity was similar between pathways. Translating operating characteristics into decision consequences demonstrated that pathway preference depends on prevalence, decision thresholds, and misclassification weighting: at 15% prevalence, AI yielded higher net benefit (140.7 vs 120.8 net true-positive decisions per 1000 screened at pₜ = 0.10). These findings support pathway-specific deployment strategies rather than direct superiority claims.