Development and Validation of Time-Dependent Risk Prediction Models for the Incidence and Progression of Chronic Kidney Disease in Individuals with Type 2 Diabetes Mellitus

Background Chronic kidney disease (CKD) is a common and severe complication of type 2 diabetes mellitus (T2DM), contributing substantially to global disease burden. However, current prognostic models for CKD progression in Asian populations remain underdeveloped. Methods We developed and validated time-dependent CKD progression models using 17-year electronic health records (EHR) from Hong Kong. Multiple machine learning (ML) and deep learning (DL) models were compared to identify the best-performing model based on area under the receiver operating characteristic curve (AUC). Survival analyses based on the Weibull Accelerated Failure Time (AFT) model were applied to estimate progression risk across different predicted risk strata. Findings The final model included 158,205 individuals from 2003 to 2019. Deep neural network (DNN) models consistently outperformed other ML models, achieving AUCs of 87.1%, 85.3%, and 84.7% for 2-, 5-, and 10-year predictions, respectively. Key predictors included serum creatinine, sex, eye complications, systolic blood pressure, age, and two-year prescription history of angiotensin. The probability of survival during the follow-up period declined more rapidly in high-risk individuals predicted by the best model. External validation in the UK Biobank (n = 17,351) the China Health and Retirement Longitudinal Study (n = 4,174) cohorts yielded AUCs of 79.7% and 74.6%, respectively. Conclusion Deep learning-based prognostic models for CKD progression in individuals with T2DM show satisfactory performance in both internal and external cohorts of Asian populations as well as in the western populations. This model could serve as a powerful tool for CKD prevention and patient management, thereby enhancing risk management, supporting early intervention and clinical decision-making.