Machine learning for the early prediction of 28-day mortality in patients with sepsis in the ICU based on MIMIC-Ⅳ database

Sepsis triggers a life-threatening organ dysfunction due to infection that contributes to high mortality. We aim to develop a machine-learning model based on Medical Information Mart for Intensive Care Ⅳ (MIMIC-Ⅳ) 2.2 database for predicting 28-day mortality in patients with sepsis in the intensive care unit (ICU). The clinical data of 1448 sepsis patients were extracted from the MIMIC-Ⅳ(2.2) database. The least absolute shrinkage and selection operator (LASSO) and multivariate logistic regression were employed to analyze potential predictors of mortality in sepsis patients. After comparing eight different algorithms, including Extreme Gradient Boosting (XGBoost), logistic regression, Light Gradient Boosting Machine (LightGBM), random forest (RF), Gaussian Naïve Bayes (GNB), multilayer perceptron (MLP), support vector machine (SVM), and decision tree (DT), the optimal model was identified based on its predictive performance. The Shapley Additive exPlanations (SHAP) interpretation was applied for personalized risk assessment. SOFA score, lactic acid, age, RDW and albumin were the predictors of mortality in patients with sepsis. Among the eight algorithms, the logistic classification model outperformed the others significantly. After remodeling with the best algorithm, the AUC in the training set was 0.754, and the algorithm performed well in the testing set (AUC = 0.746, sensitivity = 0.714, and specificity = 0.721, F1 score = 0.486). A machine learning model based on routinely available clinical data from MIMIC-IV demonstrated moderate discriminatory ability for predicting 28-day mortality in septic patients, with SHAP analysis providing insights into model interpretability. The study constructed a predictive model based on the multiple machine learning model, and the logistic regression model showed a better performance.