Early prediction of sepsis in the ICU: a comparative analysis of multiple machine-learning algorithms using the MIMIC-III database

Sepsis is a high-burden, highly heterogeneous clinical challenge that affects up to 30% of ICU patients. Reliable early prediction is essential for timely intervention and improved outcomes. We aimed to develop and validate a machine-learning model for predicting sepsis onset beyond the first 24 h of ICU admission. Data from septic patients were extracted from the Medical Information Mart for Intensive Care III (MIMIC-III) database. Feature selection was performed with the Boruta algorithm. Nine algorithms-XGBoost-DART, Gaussian Naïve Bayes, LightGBM-DART, Random Forest, AdaBoost, Multi-Layer Perceptron (MLP), Support Vector Machine (SVM-RBF), k-Nearest Neighbors (KNN), and Ridge Regression-were trained and comprehensively evaluated with respect to discrimination, calibration, and clinical utility. Class imbalance was addressed using SMOTE on the training set and cost-sensitive learning for applicable algorithms. Among 1,634 ICU patients included (after excluding those meeting Sepsis-3 criteria within the first 24 h), 349 (21.4%) developed sepsis after the 24-hour observation window. AUROCs ranged from 0.794 to 0.881 across the nine models. AUROCs ranged from 0.810 to 0.895 across the nine models. XGBoost-DART achieved the highest AUROC (0.881, 95% CI: 0.854-0.908) along with the best accuracy (0.847), F1-score (0.762), and specificity (0.897). Decision-curve analysis demonstrated that XGBoost-DART delivered the greatest net benefit over the widest range of threshold probabilities, underscoring its strong clinical utility. In summary, machine-learning models provide a reliable tool for early sepsis prediction in the ICU. The XGBoost-DART model, with its outstanding performance, empowers clinicians to identify high-risk patients and initiate timely interventions to reduce mortality.