Agreeability of Tele-ophthalmology and Artificial Intelligence-Based Diagnosis of Diabetic Retinopathy

ABSTRACT

Purpose: The COVID-19 pandemic exposed the need for increased mobilization of teleophthalmology resources. Artificial intelligence (AI) may serve as a tool to assist physicians in triaging highest need patients if the AI's assessment of disease is comparable to the physician's assessment. This study assesses the ability of AI software to diagnose diabetic retinopathy (DR) as compared to Tele-ophthalmology and in-person examination by a retina specialist. Methods: Records of forty patients (average age 55.1±10.9 years) presenting to an urban retina clinic were reviewed retrospectively for factors including demographics, retinal photos taken by Canon CR-2 Plus AF Retinal Imaging camera (Tokyo, Japan), and diagnosis of DR based on the International Clinical Diabetic Retinopathy (ICDR) classification scale during an in-person clinic visit in which a fundus exam was performed. Retinal photos were graded by AI software, EyeArt (EyeNuk, CA), as Normal, Mild DR, or More than Mild DR. Retinal images were also graded remotely by a retina specialist using the ICDR classification scale via TeamViewer software (Tele). Agreement between Tele, AI, and inperson DR diagnosis was assessed using Cohen's Kappa (κ) coefficient using IBM® SPSS® Statistics software. Results: Among 80 eyes, 33 were diagnosed in-person with no DR, 5 with mild nonproliferative DR (NPDR), 9 with moderate NPDR, 3 with severe NPDR, 7 with proliferative diabetic retinopathy (PDR), and 23 with regressed PDR. Eleven and 26 eyes could not be graded by Tele or AI, respectively. κ±SE for in-Person diagnosis vs Tele was 0.859±0.058 (p<.001), in-person vs AI was 0.751±0.082 (p<.001), and Tele vs AI was 0.883±0.063 (p<.001). Conclusions: AI is a reliable tool for screening patients for DR and referring them for physician evaluation since AI had a substantial rate of agreement with the in-person diagnosis and near perfect agreement with Tele. Tele grading was in near perfect agreement with the in-person diagnosis, showing that Tele is a reliable option for a physician to remotely screen patients that may be ungradable by AI. However, improvements are needed due to the high number of images that are ungradable via Tele and AI. Further studies should assess ways to reduce the number of ungradable images via Tele and AI and create a trend analysis for multiple visits for a given patient.