Development and validation of a machine learning-based prediction model for in-ICU mortality in severe pneumonia: A dual-center retrospective study.

Abstract

Introduction: Severe pneumonia (SP) carries a high risk of death in the intensive care unit (ICU). There is a paucity of effective assessment tools for ICU mortality in clinical practice. Therefore, this dual-centre study collects common clinical characteristics, develops, and externally validates machine learning (ML)-based models for in-ICU mortality for SP, providing guidance for preventive strategies.

Methods: Retrospective data from adult SP patients at two hospitals (Yantaishan: training; Longkou: external validation; June 2023-Feb 2025) were analyzed. LASSO regression identified key predictors. Five ML models (Logistic Regression (LR) , Random Forest (RF), RBF-SVM, Linear SVM, (XGBoost) were built. The area under the ROC curve (AUC) was utilized to evaluate the overall model performance.Model performance (AUC, sensitivity, specificity at optimal threshold via Youden index), calibration, and clinical utility (decision curve) were evaluated on the external set.

Results: In total, 501 patients were ultimately included, among whom 222 (44 %) died in the ICU. LASSO regression identified age, use of vasopressors, recent chemotherapy, SpO₂ within 8 h of ICU admission, D-dimer, platelet count, NT-proBNP, and use of invasive mechanical ventilation as modeling variables. In the external validation set, model performance was as follows: LR (AUC = 0.76; threshold = 0.339; sensitivity = 0.761; specificity = 0.639); RF (AUC = 0.77; threshold = 0.574; sensitivity = 0.448; specificity = 0.876); RBF-SVM (AUC = 0.746; threshold = 0.404; sensitivity = 0.642;specificity = 0.701); SVM-Linear (AUC = 0.741; threshold = 0.475; sensitivity = 0.507; specificity = 0.814); XGBoost (AUC = 0.76; threshold = 0.459; sensitivity = 0.597; specificity = 0.742). Based on the optimal threshold probability, LR exhibited the best clinical accuracy. Therefore, a predictive nomogram was developed using LR.

Conclusion: ML models based on common interpretable clinical features demonstrate favorable predictive value for in-ICU mortality in SP patients, providing guidance for preventive strategies in clinical practice. However, the predictive performance requires further improvement. Therefore, future studies should incorporate additional efficient biomarkers to enhance model performance.