Enhancing trauma triage in low-resource settings using machine learning: a performance comparison with the Kampala Trauma Score

Background: Traumatic injuries are a leading cause of morbidity and mortality globally, with a disproportionate impact on populations in low- and middle-income countries (LMICs). The Kampala Trauma Score (KTS) is frequently used for triage in these settings, though its predictive accuracy remains under debate. This study evaluates the effectiveness of machine learning (ML) models in predicting triage decisions and compares their performance to the KTS. Methods: Data from 4,109 trauma patients at Soroti Regional Referral Hospital, a rural hospital in Uganda, were used to train and evaluate four ML models: Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), and Support Vector Machine (SVM). The models were assessed in regard to accuracy, precision, recall, F1-score, and AUC-ROC (Area Under the Curve of the Receiver Operating Characteristic curve). Additionally, a multinomial logistic regression model using the KTS was developed as a benchmark for the ML models. Results: All four ML models outperformed the KTS model, with the RF and GB both achieving AUC-ROC values of 0.91, compared to 0.62 (95% CI: 0.61–0.63) for the KTS (p < 0.01). The RF model demonstrated the highest accuracy at 0.69 (95% CI: 0.68–0.70), while the KTS-based model showed an accuracy of 0.54 (95% CI: 0.52–0.55). Sex, hours to hospital, and age were identified as the most significant predictors in both ML models. Conclusion: ML models demonstrated superior predictive capabilities over the KTS in predicting triage decisions, even when utilising a limited set of injury information about the patients. These findings suggest a promising opportunity to advance trauma care in LMICs by integrating ML into triage decision-making. By leveraging basic demographic and clinical data, these models could provide a foundation for improved resource allocation and patient outcomes, addressing the unique challenges of resource-limited settings. However, further validation is essential to ensure their reliability and integration into clinical practice.