Enhanced Sepsis Prediction Using Ensemble Learning with SMOTE-Based Data Balancing and Stratified Validation

Sepsis, a critical condition triggered by an abnormal immune response to infection, requires rapid and accurate identification to reduce the risk of mortality. In clinical settings, datasets often suffer from severe class imbalance, with sepsis cases significantly underrepresented, which complicates early prediction efforts. This study explores and compares the effectiveness of traditional and ensemble-based machine learning algorithms for sepsis detection. Initially, models such as Random Forest (RF), Support Vector Machine (SVM), XGBoost, and K-Nearest Neighbors (KNN) were trained on imbalanced datasets using median imputation. To improve model reliability and address class imbalance, advanced ensemble techniques, such as soft-voting and stacking, were incorporated, along with cost-sensitive learning and stratified k-fold validation. The Synthetic Minority Oversampling Technique (SMOTE) was later applied to balance the dataset, and the models were reassessed. Evaluation based on metrics such as ROC-AUC, PR-AUC, sensitivity, specificity, balanced accuracy, and Brier score revealed that the stacking ensemble, applied to SMOTE-processed data, delivered superior performance with a ROC-AUC of 0.9979. These results show how ensemble approaches and data balancing strategies can improve the precision and dependability of sepsis prediction models used in clinical decision-making.