The implementation of artificial intelligence driven diabetic retinopathy screening in a district hospital in Myanmar: a pilot study

Abstract

Purpose

Diabetic retinopathy (DR) causes irreversible blindness. Early detection and timely treatment can prevent blindness. However, manpower and access to DR screening is challenging in remote areas. This study aims to validate and assess the acceptability of artificial intelligence assisted diabetic retinopathy screening (AI-DRS) versus standard care (SC) with slit lamp examination.

Design

This was a cross sectional, comparative cohort study.

Methods

All patients underwent a mydriatic two-field fundus photography, automatically analyzed by the AI and a dilated clinical fundus examination by a consultant ophthalmologist, blinded to the AI-DRS results. The primary outcome measure was the agreement between AI-DRS (Singapore Eye Lesion Analyzer (SELENA+) performed with a portable fundus camera) and SC, measured by sensitivity, specificity and kappa indices. Questionnaires were administered to test acceptability of AI-DRS.

Results

414 eyes of 207 patients (mean age 60.13, SD=9.39), male (20.3 %) and female (79.7 %) were screened. AI-DRS versus SC had a sensitivity of 85 % and specificity of 100 % with high agreement (kappa value 0.915 (p < 0.001)). Comparing AI-DRS versus SC, 350 versus 362 eyes had no DR/mild DR, 6 versus 7 eyes had moderate/severe non-proliferative DR(NPDR), 6 versus 7 had proliferative DR(PDR). AI-DRS had more ungradable eyes versus SC (52 versus 39 eyes). Questionnaires showed 98.6 % of patients found AI-DRS acceptable.

Conclusion

The AI-DRS is a valid method of DR screening in suburban Myanmar populations with high rates of acceptability, may improve access and coverage of DR screening services for Myanmar and in other similar settings.