Background: Trauma is a leading cause of death worldwide, particularly in upper-middle-income countries such as Colombia, where socioeconomic factors exacerbate outcomes. Machine learning (ML) shows promise in improving trauma care, but most models overlook social determinants. This study aims to develop ML models to predict in-hospital mortality among trauma patients by integrating clinical data and socioeconomic variables from trauma event locations. Methods: A prospective cohort study was conducted at three trauma centers in Cali, Colombia, including 553 patients. Thirty variables encompassing demographic, clinical, and socioeconomic data were collected. We evaluated five ML algorithms – Logistic Regression, Support Vector Machines, Random Forest, XGBoost, and Artificial Neural Networks – using hyperparameter optimization and stratified cross-validation. Model performance was evaluated with the Sustainability, Accuracy, Fairness and Explainability (SAFE) principles for ML applications, accuracy was assessed using F1-score, sensitivity, specificity, area under the curve (AUC) and Rank Graduation Accuracy (RAG), sustainability was evaluated with Rank Graduation Robustness. For model interpretability, SHapley Additive exPlanations (SHAP) values determined feature contributions, highlighting key factors influencing in-hospital mortality. Results: The XGBoost model was the best model according to the SAFE principles, with a RAG of 0.89, AUC of 0.97, F1-score of 0.84, sensitivity of 0.81, and specificity of 0.97. Key predictors of in-hospital mortality included Glasgow Coma Scale, systolic blood pressure, age, and socioeconomic factors such as housing deficit and social security category. SHAP analysis revealed the nonlinear impact of these variables, particularly the strong influence of Glasgow Coma Scale and the mechanism of injury. Conclusion: The XGBoost model, integrating physiological and socioeconomic variables, showed strong predictive performance for trauma-related in-hospital mortality. SHAP analysis highlighted key factors such as GCS, systolic blood pressure, and housing deficit, underscoring the value of including socioeconomic data in risk stratification
Orozco et al. (Thu,) studied this question.