A scoping review of artificial intelligence tools for diabetic retinopathy in low- and middle-income countries
Reviewed by Elliott Taylor

The authors conducted a scoping review to summarise the literature relating to artificial intelligence (AI) tools for diabetic retinopathy (DR) in low- and middle-income countries (LMICs). Eighty-one studies were included, following a comprehensive literature search. The majority of studies were from India, China or Thailand. The primary objective of most studies was to assess AI tool performance for a particular task. Most AI tools (49 of 81; 60%) were developed to interpret retinal photographs, classifying them into DR grades. Twenty of 81 (25%) involved another form of image interpretation (e.g., identification of pathological features). Studies with a focus beyond assessing diagnostic accuracy were sparse: one study assessed AI augmented human DR grading; three studies focussed on implementation; and two studies assessed cost-effectiveness. Nineteen externally validated studies reported AI tool sensitivity and specificity for the detection of referable DR, ranging from 83-100% and 69-98%, respectively. However, the authors noted that most studies excluded ‘ungradable’ images. Of those that did not exclude ‘ungradable’ images, AI tools considered a higher proportion ‘ungradable’ compared to human graders. This suggests that AI tool performance may diminish if evaluated in clinical pathways where ‘ungradable’ images are not excluded. The authors highlight that ‘ungradable’ images usually trigger a referral, thus more false positive cases could be referred to specialist services if AI tools are implemented in screening pathways. This could increase pressure on specialist services, many of which are under-resourced in LMICs. There is limited evidence on the feasibility and effectiveness of AI tools when implemented into DR screening pathways in LMICs. The authors claim that the potential benefit of clinical AI tools is possibly greatest in poorer regions of the world where there are fewer clinicians, and call for contextually relevant research around the development and implementation of clinical AI tools in LMICs.

Artificial intelligence for diabetic retinopathy in low-income and middle income countries: a scoping review.
Cleland CR, Rwiza J, Evans JR, et al.     
BMJ OPEN DIABETES RESEARCH & CARE
2023;11:e003424.