Virtual Booth

Risk stratification may improve the prioritisation of diabetic retinopathy screening, but prognostic models must account for variable follow-up and demonstrate performance across changing populations. This study developed, compared and validated machine-learning survival models for predicting five-year risk of registry-recorded diabetic retinopathy among adults with type 2 diabetes mellitus in Johor, Malaysia. A retrospective cohort study analysed 87,005 patients from the National Diabetes Registry, comprising development (2015–2019; 47,953), internal-validation (2020–2022; 18,832) and temporal-validation (2023–2025; 20,220) cohorts. Sixteen routinely collected predictors were evaluated using discrete-time proportional hazards, elastic-net discrete-time survival, a random survival forest-like interval-hazard ensemble, extra survival trees and gradient-boosted survival. Three training-to-testing allocations were compared; the selected 80:20 allocation retained 38,362 patients for training and 9,591 for testing. The selected model was applied without retraining to both validation cohorts. Performance was assessed using Harrell’s concordance index, five-year time-dependent area under the receiver operating characteristic curve, Brier score and calibration measures. Shapley additive explanations assessed predictor contributions. Diabetic retinopathy was recorded during available follow-up in 6,402 (13.35%), 2,793 (14.83%) and 2,973 (14.70%) patients across the respective cohorts. Gradient-boosted survival achieved the highest concordance index within the selected split: 0.640 (95% confidence interval 0.616–0.663), compared with 0.636 for the forest-like model. Its five-year area under the curve was 0.659 (95% confidence interval 0.634–0.683), with a Brier score of 0.069. Concordance declined to 0.616 and 0.608 in the validation cohorts. Temporal calibration deteriorated, with an intercept of −0.630 and slope of 0.740. Nephropathy, diabetes duration and age contributed most strongly to predictions. Routinely collected registry variables supported modest risk discrimination, but temporal deterioration indicates that recalibration, geographically independent validation and prospective clinical-utility evaluation are necessary before implementation in screening pathways.