An Explainable PCA-XGBoost Model for Predicting Bloodstream Infection in Hemodialysis Patients

Abstract

Bloodstream infection (BSI) is a life-threatening complication in hemodialysis (HD) patients with catheter-based vascular access, carrying mortality rates of 15–50%, yet early detection remains challenging due to high-dimensional clinical data with significant multicollinearity. This study develops a BSI prediction model integrating Principal Component Analysis (PCA), XGBoost, Synthetic Minority Oversampling Technique (SMOTE), and dual Explainable AI (XAI) methods to improve predictive performance and clinical transparency. A dataset of 391 HD patients (18.9% BSI-positive) was preprocessed with encoding, standardization, and median imputation. PCA reduced 37 features to 29 components retaining 95.0% variance; SMOTE was applied inside each cross-validation fold to prevent leakage; and hyperparameters were optimized via RandomizedSearchCV. The proposed model achieved 83.5% accuracy, 33.3% recall, 43.5% F1-score, 85.5% AUC-ROC, and 0.643 PR-AUC, outperforming the baseline (81.0% accuracy, 0.0% recall, 0.190 PR-AUC). Bootstrap 95% confidence intervals and Brier score calibration are reported; results require cautious interpretation given the small positive test set (n=15). SHAP and LIME identified PC1 (hematological parameters) and PC2 (inflammatory markers) as dominant predictors. This study explores PCA, XGBoost, and dual XAI integration for BSI prediction in HD patients, an approach not extensively examined in this context. External multicenter prospective validation is required before clinical deployment.