Background: Diabetic retinopathy (DR) is currently one of the main causes of preventable vision loss. This is particularly a problem in rural areas, where eye care is hard to come by, there are not enough specialists and screening is not conducted enough. Traditional screening efforts face formidable logistical and financial barriers in rural regions, such that when patients present with disease it is often at an advanced stage. Objective: We aimed to evaluate the diagnostic performance, feasibility, and cost-effectiveness of a particular artificial intelligence (AI)–aided screening model of diabetic retinopathy in rural medical care settings. Methods: A prospective cross-sectional diagnosis accuracy trial conducted in rural primary health centres and outreach clinics. All diabetic participants had non-mydriatic fundus photography performed by trained technicians. A deep learning artificial intelligence (AI) algorithm analysed retinal photos to identify vision-threatening diabetic retinopathy (VTDR). Independent masked evaluation by retina specialists served as the reference standard. Results: We screened 240 individuals or 480 eyes. For detecting VTDR, the sensitivity and specificity of AI system were 89.4 and 87.2%, respectively. Its positive and negative predictive values were 61.8% and 96.3%, respectively, with an area under the ROC curve of 0.93 as well. When AI assisted screening was compared to standard practice, that approach screened more people about 35% more and cut the cost per person screened by roughly 40%. Conclusion: AI aided DR screening is a reliable, feasible and cost-effective method for early detection of VTDR in remote populations with great potential for reducing unnecessary blindness in underserved regions.
Singh et al. (Wed,) studied this question.