Detection of referable diabetic retinopathy using machine learning on routine clinical data

Early detection of referable diabetic retinopathy (RDR) is crucial to prevent vision loss. We developed and validated a machine learning (ML) model using clinical and laboratory variables to predict RDR without ophthalmic imaging.

We enrolled 562 adults with diabetes who underwent fundus examination at a single tertiary center from June 2015 to December 2023, retrospectively and prospectively. RDR was defined as moderate nonproliferative diabetic retinopathy or worse, or diabetic macular edema. Predictors included demographic factors, diabetes duration, glycemic control, blood pressure, lipid profiles, and kidney function markers. Patients were randomly divided into training (n = 175) and validation (n = 387) sets. Four ML models were trained, and performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Predictor importance was assessed using Shapley Additive Explanations (SHAP).

In the validation set, the random forest achieved the highest performance, with an AUROC of 0.932 (95% confidence interval, 0.90-0.96), sensitivity of 85.8%, specificity of 91.2%, and accuracy of 87.9%. SHAP ranked 15 predictors, with age showing the highest importance, followed by diabetes duration, fasting glucose, body mass index, diastolic blood pressure, height, smoking history, Cystatin C, systolic blood pressure, hemoglobin A1c, weight, estimated glomerular filtration rate, total cholesterol, insulin use, and sex.

A random forest model using routinely available clinical data identified RDR without fundus imaging. It may serve as a practical tool for early detection of RDR in resource-limited settings, enabling timely referral and supporting integration into clinical decision support systems.