Diagnostic Performance of Deep Learning Models in Detecting Diabetic Retinopathy: A Systematic Review and Meta-Analysis of Clinical Studies

Diabetic retinopathy (DR ) remains a leading cause of vision loss among working-age adults worldwide, particularly in settings with limited access to ophthalmic care. Deep learning (DL) algorithms have shown promise in automating DR detection from retinal fundus images, but a comprehensive evaluation of their diagnostic performance across diverse clinical studies is needed. This systematic review and meta-analysis aimed to assess the pooled diagnostic accuracy of DL models for DR detection using fundus photographs. A literature search was conducted in PubMed, Scopus, Web of Science, and IEEE Xplore for peer-reviewed studies published between 2010 and 2024. Eligible studies included clinical validations of DL algorithms reporting sensitivity, specificity, and area under the curve (AUC) against a reference standard. Data extraction and analysis followed PRISMA 2020 guidelines. A bivariate random-effects model was used to pool diagnostic metrics. 14 studies involving over 500,000 retinal images were included. The pooled sensitivity was 92.1% (95% CI: 89.3–94.4%), specificity 90.4% (95% CI: 87.2–93.1%), and AUC 0.961 , with a diagnostic odds ratio of 106.3 . Subgroup and meta-regression analyses indicated consistent performance across model architectures, validation strategies, imaging modalities, and geographic settings. These findings suggest that DL systems can reliably detect referable and vision-threatening DR with performance comparable to expert human graders. Given their scalability and diagnostic strength, DL-based tools are clinically ready to augment or replace manual screening in real-world workflows, particularly in resource-constrained environments. Wider deployment, coupled with real-time validation and cost-effectiveness studies, could help transform global DR screening programs.