A Deep Learning Model for Patient Subtyping and Survival Analysis in Kidney Disease Trajectory

Chronic kidney disease (CKD) is a global public health problem, affecting over 10% of the general population worldwide. CKD patients face an increased risk of progressing to end-stage kidney failure (ESKF), which is associated with higher hospitalization rates and cardiovascular mortality. The estimated glomerular filtration rate (eGFR) is a key indicator of kidney function and a critical predictor of CKD progression. However, existing predictive models often fail to incorporate the dynamic changes in eGFR over time, leading to less accurate forecasts of disease progression. In this study, we develop a novel deep learning model, the time-dependent LSTM (TdLSTM), to improve the prediction of eGFR and time-to-ESKF by effectively handling time-varying variables and irregular measurement time intervals in longitudinal study. Our method captures the evolving patterns of CKD progression, enabling more accurate predictions of patient outcomes. The specifically designed survival time estimate approach yields accurate predictions of the time of reaching the end stage of the disease. We trained and validated the TdLSTM model using data from two cohorts of CKD patients, demonstrating its superiority over existing models in predicting CKD progression and identifying distinct patient subtypes with varying progression rates. The proposed TdLSTM model enhances our ability to predict CKD outcomes by integrating temporal information into the predictive process, ultimately generating better prediction accuracy on the redefined time-varying evaluation metrics such as AUC and concordance index.