Real-World Diagnostic Accuracy of Fully Autonomous, FDA-Approved Artificial Intelligence Systems for Diabetic Retinopathy Screening in the United States: A Systematic Review and Meta-Analysis Shreya Parimoo Stockdale High School, Bakersfield, California, United States

Abstract

Abstract

Background: Fully autonomous, FDA-cleared artificial intelligence (AI) systems for diabetic retinopathy (DR) screening—LumineticsCore (formerly IDx-DR) and EyeArt—demonstrated high sensitivity and specificity in pivotal trials. Real-world U.S. performance outside controlled trial conditions is less well characterized.

Objective: To systematically review and meta-analyze the real-world diagnostic accuracy of fully autonomous, FDA-cleared AI systems for DR screening in U.S. clinical settings, excluding pivotal trials and pre-clearance system versions.

Methods: Searches of PubMed/MEDLINE, Cochrane Library, and IEEE Xplore (Embase inaccessible). Eligible studies: U.S.-based, real-world evaluations of FDA-cleared LumineticsCore/IDx-DR or EyeArt reporting sensitivity/specificity against a human-grader reference standard. Single-reviewer screening. QUADAS-2. Random-effects (DerSimonian–Laird) meta-analysis on the logit scale; likelihood ratios; diagnostic odds ratio; GRADE.

Results: Four studies (total N ≈ 1,429) met inclusion criteria after excluding one large pre-clearance EyeArt study. Pooled sensitivity was 98.4% (95% CI: 82.8–99.9%; I² = 93.9%) and pooled specificity was 80.7% (95% CI: 66.9–89.6%; I² = 94.7%). LR+ was 5.09, LR− 0.02, and diagnostic odds ratio 253. Specificity ranged from 60.3% to 89.2%. GRADE certainty was rated Low for both outcomes.

Conclusions: In real-world U.S. deployments of FDA-cleared autonomous AI systems, sensitivity remains high but specificity is variable and often lower than pivotal-trial figures. Site-level validation remains essential. Certainty of evidence is Low.