Staging of diabetic retinopathy using an artificial intelligence-based software platform

Abstract

Objective: The aim of our study was to evaluate the potential of artificial intelligence in real-world population screening to detect stages of diabetic retinopathy, identify patients with previously undiagnosed diabetes, and establish and optimise patient pathways. Materials and methods: This prospective, multicentre, open-label observational study evaluated the feasibility of automated analysis of colour fundus images to identify individuals with suspected undiagnosed diabetes mellitus by detecting signs of diabetic retinopathy in a real-world screening programme. It was conducted at primary healthcare facilities in collaboration with The Filatov Institute of Eye Diseases and Tissue Therapy of NAMS of Ukraine. Patients were not preselected based on diabetes status, allowing assessment of fundus examination as a tool for opportunistic detection of systemic metabolic disease. Results: A total of 12,922 patients were included in the population-based screening and divided into three groups: 2,940 patients with established diabetes mellitus (used to assess detection of diabetic retinopathy in a confirmed population), 7,411 individuals without diagnosed diabetes but with risk factors (the key group for evaluating opportunistic detection), and 2,571 controls without diabetes or significant risk factors. Fundus image analysis identified signs of diabetic retinopathy in 1,308 patients (10.1%), while no retinal pathology was detected in 9,005 individuals (69.7%). Disease staging showed: mild non-proliferative retinopathy in 835 patients (6.5%), moderate in 256 (2.0%), severe in 65 (0.5%), and proliferative retinopathy in 152 (1.2%). Conclusions: Artificial intelligence demonstrated high efficacy in population-based screening for diabetic retinopathy. The screening revealed a significant prevalence of diabetic retinopathy among the population. Among the 12,922 individuals screened, signs of diabetic retinopathy were detected in 10.1% of patients. These findings highlight the importance of early detection and large-scale screening programmes. AI algorithms enable effective identification of at-risk patients, including those without a confirmed diabetes diagnosis, creating new opportunities for early diagnosis and prevention of severe complications. Automated disease staging supports clinical decision-making, while AI-based screening improves access to care, reduces the risk of vision loss, and optimises ophthalmology services, especially in resource-limited settings.