Mortality Prediction and Survival Estimation in Dialysis Patients Using Logistic and Cox Regression with Machine Learning Feature Selection

Abstract

Mortality remains high among patients undergoing maintenance dialysis for end-stage renal disease (ESRD). Identification of key mortality predictors is paramount for improving prognosis and guiding care. Recent advances in machine learning (ML) offer potential to enhance risk stratification beyond traditional statistical models. This study compares feature selection methods—LASSO, Random Forest, and Gradient Boosting—in predicting mortality risk among dialysis patients, integrating logistic regression and Cox proportional hazards modelling. Retrospective data from 224 ESRD patients on maintenance haemodialysis were analysed. Thirty-three clinical and demographic variables were evaluated. Feature subsets were generated using ML algorithms and used for building predictive models. Model performance was assessed via discrimination (AUC), accuracy, sensitivity, specificity, and survival prediction concordance index (C-index). LASSO-selected features yielded an AUC of 0.82 and C-index of 0.81, demonstrating strong discriminatory ability. Random Forest showed highest AUC (0.85) but lower sensitivity. Gradient Boosting offered balanced sensitivity and specificity with an AUC of 0.81. The parsimonious common-feature model (dialysis session frequency, diabetes) achieved the best survival discrimination (C-index 0.83). Full models with all variables demonstrated moderate performance, highlighting potential overfitting. Key mortality predictors included dialysis adequacy, diabetes status, respiratory comorbidities, and hemodynamic parameters. Machine learning–aided feature selection enhances mortality risk prediction in dialysis patients. Parsimonious models focusing on consistent predictors may optimize clinical applicability. These findings support integrating ML and traditional regression approaches to refine prognostic tools and inform personalized care strategies in ESRD.