Comparative Evaluation of Machine Learning Classifiersfor Early Sepsis Detection in Intensive Care Units

Abstract

Sepsis remains a leading cause of ICU mortality worldwide, with each hour of delayed treatment increasing mortality risk by approximately 8%. Traditional scoring systems like SOFA and SAPS-II rely on manual calculation and fail to capture complex, non-linear interactions in high-dimensional physiological data. This study presents a comparative evaluation of four machine learning classifiers—Logistic Regression, Support Vector Machine (RBF kernel), Random Forest, and XGBoost—for early sepsis prediction four to six hours prior to onset. Using the MIMIC-III database, 38 clinical features were extracted from over 25,000 ICU admissions and evaluated via stratified five-fold cross-validation. Class imbalance was addressed using SMOTE, and missing values were handled through MICE imputation. XGBoost achieved the highest AUC-ROC(0.9242) and specificity(94.8%), while Logistic Regression led on sensitivity(75.0%). SHAP analysis identified lactate, systolic blood pressure, heart rate, and BUN as the most influential predictors, supporting the feasibility of ensemble-based clinical decision support in critical care settings.