Predicting Sepsis Onset in ICU Patients Using Machine Learning and Feature Section: A Case Study of MIMIC-IV Data

This study aims to predict the onset of sepsis in ICU patients using multiple machine-learning models with data from the MIMIC-IV database. We employed seven prediction models-Logistic Regression, Gaussian Naive Bayes, Random Forest, Artificial Neural Network, SVM, XGBoost, and Gradient Boosting Decision Tree-using 81 features extracted from routine checks conducted within 12 hours before and 4 hours after ICU admission from 46,530 patients. The features were represented by key values such as maximum, minimum, and mean, referencing the official MIMIC derived daily data tables. XGBoost achieved the best performance with an AUC of 0.81 and an accuracy of 0.729. We then applied Recursive Feature Elimination (RFE) to identify the optimal feature subset for each model, finding 13 features commonly selected across all models. This overlap highlights their importance in predicting sepsis and suggests potential for model simplification without losing predictive power. Notably, 6 of these common features overlap with the top 20 SHAP features from the XGBoost model, validating their critical role. Training models with these 18 common features demonstrated that this feature selection process can lead to a simplified, yet effective, predictive model. Thus, we developed a simple, rapid, and practical tool for early detection and clinical intervention of sepsis.