Machine Learning-Based Mortality Prediction in Hemodialysis Bloodstream Infection: A SHAP-Integrated Approach for Real-Time, Personalized Risk Stratification and Clinical Decision Support

Abstract: Background Bloodstream infection (BSI) represents a leading cause of morbidity and mortality in hemodialysis populations, with in-hospital mortality rates reaching 50%. Conventional clinical severity scoring systems (qSOFA, APACHE II) demonstrate inadequate predictive accuracy in dialysis cohorts, failing to identify 20-40% of high-risk patients at presentation. Machine learning with explainable artificial intelligence (SHAP) offers potential to improve mortality prediction while maintaining clinical interpretability. Methods This retrospective cohort study analyzed 148 hemodialysis patients with confirmed BSI from a tertiary dialysis center (2018-2022). Five machine learning algorithms were developed and validated using 5-fold stratified cross-validation: Random Forest, XGBoost, LightGBM, Logistic Regression, and Support Vector Machine. SHAP-derived feature importance analysis identified patient-specific and population-level mortality drivers. Performance was compared against qSOFA and APACHE II using AUROC as the primary metric. Results The cohort demonstrated 50% in-hospital mortality (n=74 deaths). Random Forest achieved superior discriminative ability with AUROC 0.94 (95% CI 0.89–0.98), representing 23-percentage-point improvement over qSOFA (AUROC 0.71, p