Machine learning models for early mortality prediction in trauma patients using public data: a nationwide
Seung Min Baik1, Jae Gil Lee1, Hongjin Shim1
1Division of Critical Care Medicine, Department of Surgery, Ewha Womans University Mokdong Hospital, Ewha Womans University College of Medicine, Seoul, Republic of Korea.
World Journal of Emergency Surgery : WJES
|May 29, 2026
Summary
Machine learning models accurately predict trauma patient mortality using nationwide data. The XGBoost model showed superior performance, identifying key risk factors for early intervention and improved patient outcomes.
Area of Science:
- Medical Informatics
- Public Health
- Emergency Medicine
Background:
- Trauma is a significant global cause of death, especially in young individuals.
- Early identification of high-risk trauma patients is crucial for effective treatment and better outcomes.
- Limited research exists on applying artificial intelligence and machine learning to predict trauma mortality.
Purpose of the Study:
- To develop and validate machine learning models for predicting mortality in trauma patients.
- To assess the performance and interpretability of various machine learning algorithms using a large, national dataset.
- To identify key predictors of mortality in trauma patients for clinical application.
Main Methods:
- Utilized a large dataset (207,012 cases) from the National Community-Based Critical Injury Survey (South Korea, 2016-2020).
- Trained and evaluated six machine learning algorithms: logistic regression, k-nearest neighbor, decision tree, random forest (RF), extreme gradient boosting (XGB), and multi-layer perceptron.
- Assessed model performance using AUROC and AUPRC; employed SHAP scores for feature importance interpretation.
Main Results:
- The XGBoost model achieved the highest performance (AUROC 0.985, AUPRC 0.957), closely followed by the Random Forest model (AUROC 0.984, AUPRC 0.956).
- Model performance demonstrated temporal robustness, remaining stable during the COVID-19 pandemic.
- Key predictors identified included out-of-hospital cardiac arrest, injury severity score, age, and time to transfusion.
Conclusions:
- Developed a high-performing, interpretable machine learning framework for early trauma mortality risk stratification using nationwide data.
- The model's strong predictive power and temporal stability suggest its utility as a system-level tool.
- Further validation and prospective studies are necessary before clinical integration.