LAUSR

Objective To evaluate the sensitivity and specificity of a Comprehensive Artificial Intelligence Retinal Expert (CARE) system for detecting diabetic retinopathy (DR) in a Chinese community population. Methods This was a cross-sectional, diagnostic study. Participants with a previous diagnosis of diabetes from three Chinese community healthcare centers were enrolled in the study. Single-field color fundus photography was obtained and analyzed by the AI system and two ophthalmologists. Primary outcome measures included the sensitivity, specificity, positive predictive value, and negative predictive value with their 95% confidence intervals (CIs) of the AI system in detecting DR and diabetic macular edema (DME). Results In this study, 443 subjects (848 eyes) were enrolled, and 283 (63.88%) were men. The mean age was 52.09 (11.51) years (range 18–82 years); 266 eyes were diagnosed with any DR, 233 with more-than-mild diabetic retinopathy (mtmDR), 112 with vision-threatening diabetic retinopathy (vtDR), and 57 with DME. The image ability of the AI system was as high as 99.06%, whereas its sensitivity and specificity varied significantly in detecting DR with different severities. The sensitivity/specificity to detect any DR was 75.19% (95%CI 69.47–80.17)/93.99% (95%CI 91.65–95.71), mtmDR 78.97% (95%CI 73.06–83.90)/92.52% (95%CI 90.07–94.41), vtDR 33.93% (95%CI 25.41–43.56)/97.69% (95%CI 96.25–98.61), and DME 47.37% (95%CI 34.18–60.91)/93.99% (95%CI 91.65–95.71). Conclusions This multicenter cross-sectional diagnostic study noted the safety and reliability of the CARE system for DR (especially mtmDR) detection in Chinese community healthcare centers. The system may effectively solve the dilemma faced by Chinese community healthcare centers: due to the lack of ophthalmic expertise of primary physicians, DR diagnosis and referral are not timely.