[Development and validation of an explainable machine learning-based model for predicting mortality risk in polytrauma patients].

Polytrauma is commonly defined as multisystem trauma involving at least two body regions with an Abbreviated Injury Scale (AIS) score ≥ 3, characterized by complex pathophysiological interactions and extreme clinical heterogeneity. This study retrospectively analyzed data from 171 274 polytrauma patients in the National Trauma Data Bank (NTDB) from 2018 to 2021. Key features were selected using the eXtreme Gradient Boosting (XGBoost) algorithm. Prediction models were then developed using XGBoost, logistic regression, random forest, naive Bayes, and support vector machine, respectively. Model performance was evaluated by receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Subgroup standardized mortality ratio (SMR) validation and global and local interpretability analysis based on SHapley Additive exPlanations (SHAP) theory were also conducted. The results demonstrated that the XGBoost model incorporating Glasgow Coma Scale (GCS), Injury Severity Score (ISS), massive transfusion within 4 hours, respiratory support, unplanned intubation, prehospital care-limiting directives, unplanned intensive care unit (ICU) admission, sepsis, ventilator-associated pneumonia, and liver cirrhosis achieved optimal performance, with an area under the ROC curve (AUC) of 0.90. Its calibration curve closely aligned with the ideal diagonal, and decision curve analysis indicated the highest clinical net benefit across a wide range of threshold probabilities. Subgroup SMRs ranged from 1.07 to 1.18, all close to 1.0. In conclusion, the developed in-hospital mortality risk prediction model for polytrauma patients exhibits favorable discrimination and clinical net benefit, and can assist clinicians in identifying high-risk populations while providing evidence to support individualized treatment decisions.