Interpretable machine learning identifies immune-inflammatory and immunothrombotic biomarkers for myocardial injury and mortality risk stratification in severe pneumonia with diverse infectious etiologies.
Wang Yicheng Y, Li Feng F, Yan Tianqiang T, Wu Yi Y et al.
Severe pneumonia is frequently associated with dysregulated immune-inflammatory responses and immunothrombotic activation, contributing to myocardial injury and adverse clinical outcomes. Early identification of patients at high risk for myocardial injury and mortality across varied infectious etiologies remains challenging. We aimed to characterize inflammatory and coagulation-related signatures associated with myocardial injury and construct interpretable machine learning models for risk stratification in patients with severe pneumonia. This retrospective cohort study enrolled 287 adult patients with severe pneumonia admitted to the intensive care unit from 2018 to 2024. All patients were stratified into groups with bacterial infection, COVID-19 and bacterial co-infection, and influenza and bacterial co-infection. Clinical variables collected within 48 h after admission were analyzed using multivariable regression, competing-risk models, and interpretable machine learning. SHapley Additive exPlanations (SHAP) were used to identify key predictive features. Myocardial injury was highly prevalent across all infectious subgroups. Patients with bacterial infection exhibited an exacerbated inflammatory and coagulation burden, characterized by elevated leukocyte counts, D-dimer levels, and prolonged prothrombin time. Multivariate analysis confirmed that D-dimer, prothrombin time, and creatinine were independently associated with myocardial injury. Machine learning analyses identified coagulation and inflammatory markers as major contributors to myocardial injury risk. For mortality prediction, the XGBoost model yielded optimal predictive performance with an AUC of 0.85. SHAP analysis revealed that vasopressor administration, prothrombin time, advanced cardiovascular support, age, and hypoxemia were the top prognostic determinants for mortality. Although infectious etiology was not independently associated with mortality, patients with myocardial injury and bacterial infection exhibited the poorest survival outcomes. Inflammatory and immunothrombotic signatures are closely associated with myocardial injury and adverse outcomes in severe pneumonia. Interpretable machine learning models exhibited promising discriminative performance for both myocardial injury and all-cause mortality in our cohort. Although the initial training dataset was derived from a single center with a relatively limited sample size, we performed temporal validation in this study, which preliminarily suggested the potential clinical applicability of these models for the early identification of high-risk patients.