Accuracy of artificial intelligence versus eyecare practitioners in detecting diabetic retinopathy using retinal fundus images: a systematic literature review

Abstract

Diabetic retinopathy (DR) prevalence rises among low- and middle-income countries where access to appropriate eye care services is lacking. Detecting DR requires retinal fundus imaging, but interpretation conventionally depends on eye-care practitioners, who are affected by shortcomings like inter-observer variability. Artificial Intelligence (AI), a promising technology for automated image examination and interpretation, enables wider, faster screening coverage and consistency. The review compares the diagnostic accuracy of AI and practitioners in detecting DR using retinal images. We conducted a systematic literature review using PRISMA 2020, searching major electronic databases with search terms: diabetic retinopathy, artificial intelligence, eye-care practitioners, and diagnostic accuracy.  Eligible studies were published in English, peer-reviewed, and met the inclusion and exclusion criteria. Data extraction, organization, and accuracy were enabled through an Excel data form. The review included 15 of 540 identified studies. AI diagnostic tools like Google’s ARDA show equal or higher accuracy than humans in detecting DR, in some instances, in categorizing images. AI brings hope and assurance in prompt identification, management, and referral. Its performance is limited by the tendency to flag images as false positives; however, this depends on image quality, disease severity, datasets used, and overall grading criteria. Human strength remains in clinical judgement, contextualised interpretation of ungradable or low-quality images, and patient-specific management approach. AI implementation to complement care should be steered by local validation through comparison with eye care givers and approval through standardized protocols, quality control, ethical considerations, and data security.