A comparative analysis of trauma-related mortality in South Korea using classification models.

Abstract

Reducing mortality among severe trauma patients requires the establishment of an effective emergency transportation system and the rapid transfer of patients to appropriate medical facilities. Machine learning offers significant potential to enhance the efficiency and quality of these emergency medical services. A retrospective secondary analysis was conducted using region-specific trauma survey data. The analysis focused on socio-economic characteristics, mechanisms of injury, injury severity, and variables indicating the effectiveness of the emergency medical system in optimizing machine learning algorithms for predicting severe patient transportation decisions. Among the 8,769 patients with severe trauma, 7.2 % died in the hospital, with an average age of 50.06 years. The average injury severity score was 8.44, and the average time from accident reporting to arrival at the emergency medical facility was 55.39 min. The trend showed that as the level of the emergency medical institution increased, the patient transport time increased, while the mortality rate decreased. Additionally, XGBoost showed the best performance in mortality classification using a dataset sampled with SMOTE-ENN. Although the difference was minimal, undersampling slightly outperformed oversampling in the classification of emergency patients. The treatment of emergency patients is influenced not only by transport time but also by the resources and staff levels of specialized emergency medical centers, which in turn affect survival rates. Furthermore, given the superior performance of composite sampling methods in analyzing imbalanced datasets, the importance of considering such imbalanced datasets in the analysis is evident.