Development of Machine Learning Models to Predict 28-Day Mortality in Patients with Sepsis-Associated Liver Injury

Background: Sepsis-associated liver injury (SALI) is a serious complication of sepsis that increases the risk of organ dysfunction and mortality; however, early identification of high-risk patients remains difficult due to nonspecific clinical features and complex pathophysiology. This study aimed to develop machine learning (ML) models to predict 28-day mortality in SALI patients within the first 24 h of intensive care unit (ICU) admission. Methods: A total of 1157 patients were included, comprising 826 from the MIMIC-IV (v2.2) database, 225 from MIMIC-III (v1.4), and 106 from eICU (v2.0). Data from MIMIC-IV were split into training and internal validation sets (7:3), while MIMIC-III and eICU served as external validation cohorts. Thirty clinically relevant features were selected. Eight ML models were evaluated using AUROC, accuracy, precision, recall, F1-score, and specificity. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) enhanced interpretability. Results: XGBoost model achieved the best performance, with an AUROC of 0.8556 (95% CI: 0.807–0.898), accuracy of 0.7702, recall of 0.8469, and specificity of 0.7200. SHAP identified lactate, blood urea nitrogen, heart rate, hemoglobin, and diastolic blood pressure as key predictors, while LIME provided patient-level interpretability. Conclusions: The XGBoost-based model may facilitate early mortality risk stratification and support clinical decision-making for SALI patients in ICU settings.