Time-dependent LSTM for Survival Prediction and Patient Subtyping in Kidney Disease Trajectory

Chronic kidney disease (CKD) affects over 10% of the global population and is projected to become the fifth leading cause of years of life lost (YLL) by 2040. Accurate prediction of CKD progression to end-stage kidney failure (ESKF) is critical for timely interventions that can slow or halt disease progression. However, current models often fail to address the complexities of time-varying biomarkers like estimated glomerular filtration rate (eGFR) and the irregular nature of longitudinal health data, resulting in suboptimal predictions. In this study, we develop a Time-dependent Long Short-Term Memory (TdLSTM) network to analyze longitudinal eGFR data and predict time-to-ESKF. Our model is specifically designed to handle irregular time intervals and temporal dynamics, capturing nuanced patterns of CKD progression. We conducted experiments on two independent CKD cohorts, MASTERPLAN and NephroTest, using patient data including age, gender, eGFR, UACR, and diagnosis. The TdLSTM model outperformed traditional and state-of-the-art predictive models, demonstrating superior accuracy in estimating time-to-ESKF and identifying subtypes of CKD progression through unsupervised clustering. By leveraging the temporal dynamics of biomarkers, our approach offers a robust tool for personalized survival prediction and risk stratification. These findings highlight the potential of deep learning in improving CKD management and identifying high-risk patients in time for effective intervention.