Early detection of trauma patients requiring intensive care in the emergency department: A next-generation risk score model

Background: Trauma remains a leading cause of death worldwide; therefore, it is important to identify patients who need intensive care unit (ICU) admission in the emergency department (ED). Current trauma scoring systems such as the Glasgow Coma Scale (GCS), Revised Trauma Score (RTS), and Injury Severity Score (ISS) are not very efficient at predicting ICU need. The application of machine learning (ML)-based predictive models is a novel approach to enhance the triage process. Objective: The primary objective of this study was to develop and validate a risk scoring model based on machine learning for early identification of trauma patients requiring ICU admission from the emergency department. The study also aimed to assess the predictive ability of the ML model compared to traditional scoring systems such as the GCS, RTS, and ISS. Methods: A retrospective, observational cohort study was conducted at Esenyurt Necmi Kadıoğlu State Hospital, collecting trauma patient data from January 1, 2024, to August 31, 2024. A total of 1,500 trauma patients aged ≥ 18 years with complete clinical, laboratory, and imaging data were included. Predictive variables consisted of demographics, trauma mechanism, vital signs, laboratory results, imaging findings, and existing trauma scores. The area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were used to train and evaluate ML algorithms (Logistic Regression, Random Forest, Support Vector Machines, XGBoost, and LightGBM). The model was compared to traditional scoring systems using the DeLong test. Results: Of the 1,500 patients, 50.73% (n = 761) required ICU admission. The developed ML model had an AUC of 0.999, with sensitivity of 99.22%, specificity of 99.69%, and accuracy of 99.56%, far outperforming traditional scoring systems. The strongest predictors of ICU admission were age, lactate level, RTS, systolic blood pressure, respiratory rate, and oxygen saturation. No significant difference in ICU admission rates was observed between blunt and penetrating trauma groups, indicating that trauma mechanism alone should not be used as a predictor. Conclusion: The machine learning-based risk scoring model demonstrated better predictive performance than traditional trauma scoring systems in identifying trauma patients requiring ICU admission. Integration of this model into ED workflows may improve triage and patient care. However, validation in multicenter prospective studies is needed before clinical implementation.