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

Comprehensive Summary

Ling et al. developed and retrospectively evaluated a multi-model machine learning framework to predict in-hospital mortality by combining clinical data (vital signs, routine laboratory and blood-gas metrics) from a patient's first 30 minutes in the emergency department (ED). Multiple machine learning and ensemble models such as stacking and voting were implemented. Data from 408 critically injured trauma patients at a single center was analyzed. Area Under the Receiver Operating Characteristic Curve (AUROC) values for single-model test sets ranged from 0.743 to 0.927, with respective Area Under the Precision-Recall Curve (AUPRC) values from 0.438 to 0.904. The stacking ensemble reached an AUROC of 0.9462 and AUPRC of 0.8679, while the voting ensemble achieved an AUROC of 0.9506 and AUPRC of 0.8715. The stacking model found that Injury Severity Score (ISS) (mean AUROC decrease 0.0360), base excess (BE) (0.0258), Glasgow Coma Scale (GCS) (0.0247), and pH (0.0153) were the most critical predictors of in-hospital mortality.

Outcomes and Implications

The early classification of trauma patients at risk for in-hospital death is critical to direct time-sensitive emergency procedures and treatments. This study found that an ensemble machine learning model combining early vital signs and laboratory data achieved high performance in predicting mortality in trauma patients. Specifically, ISS, BE, GCS, and acid–base variables emerged as the strongest predictors. The study was limited by its single-center, retrospective design, the short 30-minute data window, and failure to test whether the model’s predictions remained fair and unbiased across different demographic groups. Nonetheless, the findings suggest that the integration of an ensemble machine learning model in severe trauma may help guide clinicians in providing timely, life-saving treatment.