Early mortality prediction after severe trauma using ensemble machine learning: a single-center retrospective study

Early identification of trauma patients at risk of in-hospital death is essential for guiding time-critical resuscitation and operative strategies. We developed and evaluated a multi-model machine learning framework that integrates vital signs, and routine laboratory and blood-gas metrics indices obtained within 30 min of emergency department arrival to predict in-hospital mortality. This single-center retrospective study included 408 critically injured trauma patients treated at the Emergency Department of the Affiliated Kunshan Hospital of Jiangsu University (January 2020-December 2021). We implemented multiple machine learning models [logistic regression, Random Forest, Gradient Boosting, XGBoost, LightGBM, and multilayer perceptron (MLP)], and constructed stacking and soft-voting ensembles. On the test set, single-model AUROC ranged from 0.743 to 0.927, with corresponding AUPRC spanning 0.438 to 0.904. The stacking ensemble achieved AUROC 0.9462 and AUPRC 0.8679; the voting ensemble achieved AUROC 0.9506 and AUPRC 0.8715. Permutation importance in the stacking model prioritized Injury Severity Score (ISS) (mean AUROC decrease 0.0360), base excess (BE) (0.0258), Glasgow Coma Scale (GCS) (0.0247), and pH (0.0153). In conclusion, an ensemble machine learning approach integrating early vital signs and laboratory data provides excellent discrimination and calibration for in-hospital mortality after severe trauma, with dominant contributions from ISS, GCS, and acid-base variables. These findings support the feasibility of interpretable ensemble learning for early mortality risk stratification in severe trauma.