Development and validation of a short-term mortality risk prediction model for severe pneumonia

Background　Patients with severe pneumonia are in a critical condition and have a high mortality rate; early intervention in this population can significantly improve their prognosis. Objective　To identify risk factors for mortality in patients with pneumonia in intensive care units, develop a predictive model for short-term mortality (death within 7 days), and apply this model in clinical practice. Methods　Clinical data about 4 247 critically ill pneumonia patients were collected from the MIMIC-Ⅳ database (2008 — 2022). Data were randomly divided into a training set (n=3 185) and a validation set (n=1 062) in a 7.5:2.5 ratio. Five machine learning models were constructed using features selected via LASSO regression: logistic regression, Elastic Net (Enet), Random Forest (RF), extreme gradient boosting (XGBoost), and light gradient boosting machine (lightGBM). The optimal model was comprehensively evaluated using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA), sensitivity, specificity, and negative predictive value (NPV), with comparative analysis against the SOFA score. Model interpretability was enhanced through a nomogram for individualized risk visualization, and an online web calculator was developed to improve clinical utility. Results　The final model incorporated nine key features: pH, age, blood urea nitrogen, oxygen saturation, lactate, partial thromboplastin time (PTT), anticoagulant therapy, type of infection, and anti-pathogen treatment. Among these, the Logistic model demonstrated the best overall performance in the validation set (AUC=0.82, 95% CI: 0.77 - 0.89; NPV= 0.99, 95% CI: 0.97 - 0.99), with the calibration curve and DCA curve confirming the model's good stability and clinical applicability. The Logistic model had a Brier score of 0.042, a calibration slope of 0.944, and an intercept of 0.027; the Hosmer-Lemeshow goodness-of-fit test indicated a good fit (P=0.768), and its predictive performance was significantly superior to that of the SOFA score (DeLong test, P＜0.05). The study further deployed the final model as an online web-based calculator for the rapid assessment of short-term mortality risk in patients with severe pneumonia. Conclusion　This study has constructed and validated a machine learning-based clinical predictive model for short-term mortality risk in severe pneumonia. The model demonstrates high predictive efficacy and clinical utility, supporting early clinical decision-making and risk assessment.