Application of artificial intelligence for community-based diabetic retinopathy detection and referral / 中华实验眼科杂志

ABSTRACT

Objective:To evaluate the value of applying an artificial intelligence (AI) system for diabetic retinopathy (DR) detection and referral in community.Methods:A diagnostic test study was conducted.Four hundred and twenty-one patients (812 eyes) diagnosed with diabetes in three Dongguan community healthcare centers from January 1, 2020 to December 31, 2021 were enrolled.There were 267 males, accounting for 63.42% and 154 females, accounting for 36.58%.The subjects were 18-82 years old, with an average age of (51.72±11.28) years.The disease course of the subjects was 0-30 years, with an average course of 3.00 (1.00, 7.00) years.At least one macula-centered 50-degree fundus image was taken for each eye to build a DR image database.All the images were independently analyzed by an AI-assisted diagnostic system for DR, trained and qualified community physicians and ophthalmologists to make diagnosis including with or without DR, referable diabetic retinopathy (RDR) and referral recommendation or not.With diagnoses from ophthalmologists as the standard, sensitivity and specificity of the AI system in detecting DR and RDR were evaluated.The consistency and effective referral rate of the AI system and community physicians in detecting DR, especially in detecting RDR were evaluzted.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of Dongguan Tungwah Hospital (No.2019DHLL046).Results:Of 812 eyes, 242 eyes were diagnosed with DR, including 23 with mild nonproliferative diabetic retinopathy (NPDR), 120 with moderate NPDR, 60 with severe NPDR and 39 with proliferative diabetic retinopathy (PDR). The other 570 eyes were diagnosed without DR.The sensitivity/specificity of AI system to detect DR and RDR was 87.60%/97.89% and 90.41%/96.29%, respectively.Compared with the ophthalmologists' diagnosis, the Cohen' s Kappa statistic of AI system to detect DR/RDR was 0.87/0.87, which was lower than 0.93/0.98 of community physicians.Among the referral-recommended cases by ophthalmologists, the effective referral rate of the AI system was 90.87% (199/219), which was higher than 89.50% (196/219) of community physicians, without statistically significant difference ( P=1.000). Conclusions:The AI system shows high sensitivity, specificity and consistency in DR detection, especially in RDR.The AI system is better in recognizing RDR than trained community physicians.