An artificial intelligence-based system can accurately identify patients with referable diabetic retinopathy in routine clinical practice, according to study results published in
Nature.
Researchers evaluated an automated retinal image analysis system implemented in an endocrinology clinic at Erasmus Hospital in Brussels, Belgium. The system analyzes non-mydriatic fundus photographs to detect referable diabetic retinopathy and diabetic macular edema, conditions that require ophthalmologic evaluation. The findings indicate that AI-assisted screening could reduce reliance on specialist graders while maintaining high diagnostic performance.
Upon evaluating the AI-based screening system in a real-world hospital setting, it was found to maintain "very high" diagnostic accuracy, with performance "exceeding" regulatory benchmarks for sensitivity and specificity, noted researchers.
"These findings suggest that AI could help reduce the burden on ophthalmology services by triaging large numbers of patients with diabetes and allowing specialists to focus on those who most urgently need care," said Lina Berrada, MD, in an interview with AACE Endocrine AI. "Ultimately, integrating AI into routine diabetes care could improve access to screening and help prevent avoidable blindness."
The study authors noted that increasing diabetes prevalence is placing pressure on ophthalmic services, writing that "diabetic retinopathy is the most common microvascular complication of diabetes and a leading cause of vision loss and preventable blindness in working-age adults worldwide."
Between January and April 2024, investigators screened 405 adult patients with diabetes attending routine consultations. After excluding individuals with missing or ungradable images, 353 patients were included in the primary analysis. The median age of participants was 56 years, and the median duration of diabetes was 13 years, noted Dr. Berrada of the ULB Center for Diabetes Research, at the Université Libre de Bruxelles, in Brussels, and colleagues.
Using retinal images captured with a non-dilated camera, the AI system evaluated each patient for referable diabetic retinopathy. All images were independently graded by a retinal specialist using the Early Treatment Diabetic Retinopathy Study classification, which served as the reference standard.
Among patients identified by the specialist as having referable disease, three had vision-threatening diabetic retinopathy, and all were correctly flagged by the AI system.
The authors wrote that the results demonstrate strong agreement between automated and human grading, noting that the findings "support the robustness, generalizability and, operational feasibility of this AI system for diabetic retinopathy screening in routine clinical care."
Performance remained consistent across demographic and clinical subgroups, including age, sex, ethnicity, diabetes type, and body mass index. The model also confirmed known clinical risk factors. Multivariate analysis showed that higher HbA1c levels at diagnosis and longer duration of diabetes were significant predictors of referable diabetic retinopathy in both AI-based and human assessments.
Beyond identifying diabetic retinopathy, fundus photography also revealed other ocular findings requiring specialist evaluation. Overall, 23.5% of patients were referred for newly detected ocular abnormalities, although only 15.7% of those referrals were due to diabetic retinopathy.
Researchers noted that AI systems are intended to support, not replace, clinical expertise. Automated analysis may help triage large volumes of screening images and allow ophthalmologists to focus on patients with more advanced disease.
Further research is needed to assess long-term outcomes, referral adherence, and cost-effectiveness before widespread implementation.
The authors declared no competing interests.
AACE Endocrine AI
is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (AACE®). The ideas and opinions expressed in
AACE Endocrinology do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.