Prevalence of glaucomatous-appearing discs in patients undergoing artificial intelligence screening for diabetic retinopathy

Abstract

PurposeTo determine the prevalence of optic disc changes suspicious for glaucoma (OCSG) which may be overlooked by artificial intelligence-based diabetic retinopathy (DR) screening systems in place of in-person dilated screening exams. MethodsThis was a retrospective cross-sectional study. High-resolution images were retrieved for 677 consecutive automated DR screening exams performed in adult patients. These were reviewed by two blinded readers for evidence of OCSG, which included increased vertical cup-to-disc ratio (VCDR), beta-zone peripapillary atrophy, optic disc hemorrhage, vessel bayoneting, rim thinning, and/or central vessel nasalization. The prevalence of OCSG was reported and compared between subjects designated as referable versus non-referable for DR. ResultsIn total, 823 eyes of 413 subjects were included. The total OCSG prevalence was 4.8% to 20.6%, with no statistically significant difference between patients designated referable versus non-referable on artificial intelligence-based DR screening. The combined classification reliability of the image readers was moderate to substantial compared to a glaucoma specialist. ConclusionsOCSG were detected in 4.8% to 20.6% of adult diabetic patients who underwent AI-based DR screening in our cohort. This represents a significant prevalence of optic disc findings which, if detected on manual screening, would have prompted referral for a glaucoma workup. Such findings illustrate the need for artificial intelligence-based screening platforms which cover multiple high-prevalence ophthalmic conditions. Replacing manual dilated screening exams with systems optimized for a single disease may increase the risk of missing comorbid ocular disease.