Machine learning models for early mortality prediction in trauma patients using public data: a nationwide retrospective study.

backgroundTrauma is a leading cause of morbidity and mortality worldwide, particularly in younger populations. Early identification of high-risk trauma patients is critical for timely interventions and improved outcomes. Although artificial intelligence and machine learning have demonstrated promise in healthcare, their application in trauma mortality prediction has been limited.

methodsThis study developed and validated machine learning models to predict mortality in trauma patients using a large public dataset from the National Community-Based Critical Injury Survey (South Korea, 2016-2020). Overall, 207,012 cases were analyzed. Six machine learning algorithms, including logistic regression, k-nearest neighbor, decision tree, random forest (RF), extreme gradient boosting (XGB), and multi-layer perceptron, were trained and evaluated. Their performance was assessed using the areas under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC), and other metrics. Shapley additive explanations (SHAP) scores were used to interpret feature importance.

resultsThe XGB model demonstrated the highest performance (AUROC 0.985; AUPRC 0.957), followed closely by the RF model (AUROC 0.984; AUPRC 0.956). Performance remained stable during the COVID-19 period, supporting the model's temporal robustness under systemic disruption. SHAP analysis identified clinically actionable features such as out-of-hospital cardiac arrest, injury severity score, age, and time to transfusion. Unlike many prior studies based on small or single-center datasets, our model was developed using a nationally representative cohort and prioritized interpretability, scalability, and generalizability.

conclusionsThis study presents a high-performing, interpretable machine learning framework for early mortality risk stratification in trauma patients using nationwide registry data. The strong discrimination and temporal robustness of the model support its value as a system-level prediction tool; however, further calibration analyses, external validation, and prospective implementation studies are required before integration into clinical workflows.