A clinically interpretable machine learning model for early detection of diabetic retinopathy in multiple community health centers

BackgroundEarly detection of diabetic retinopathy (DR) remains challenging in primary care, where access to ophthalmic screening is limited. We developed and validated a prediction model using routinely collected health data to identify diabetic patients at increased risk of DR.MethodsThis retrospective study included 1,475 diabetic patients from three community health centers in China. The cohort was split into a development set (n = 1,177) and a held-out test set (n = 298). We developed three machine learning models using 5-fold cross-validation: penalized logistic regression (GLMNET), extreme gradient boosting (XGBoost), and random forest (Ranger). Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), Brier score, calibration, and decision curve analysis. Feature importance was assessed using SHapley Additive exPlanations (SHAP).ResultsDR prevalence was 13.5%. In the test set, GLMNET achieved an AUROC of 0.770 (95% CI 0.671–0.856) and an AUPRC of 0.452 (95% CI 0.325–0.620). Its Brier score was 0.095, with a calibration intercept of 0.206 and a calibration slope of 0.953. XGBoost showed comparable discrimination, whereas Ranger performed less favorably. Decision curve analysis suggested possible net benefit across threshold probabilities from 10% to 40%. SHAP analyses identified urine glucose as the most influential predictor.ConclusionsThis model showed moderate discrimination and acceptable but imperfect calibration in a three-center community-based cohort. Its use of routinely collected variables and transparent model structure suggests potential value for risk stratification in primary care, but external validation and prospective implementation studies are required before routine clinical use.