Beyond Thresholds: Can Machine Learning Improve Trauma Field Triage?

Background Accurate triage of trauma patients by Emergency Medical Services (EMS) is essential for optimal outcomes and resource allocation. The 2021 National Field Triage Guidelines (FTG) assist EMS in prehospital triage; however, its collective performance has never been evaluated using a national database. We aimed to evaluate an FTG surrogate and develop a predictive model to identify patients at risk for serious injury. Methods The Trauma Quality Improvement Program National Trauma Databank (2017-2020) was queried for traumatically injured adults transported via ambulance. Serious injury was defined as ISS ≥16, blood transfusion, angiography with intervention, hemorrhage control surgery, intubation, intracranial pressure monitoring, admission to intensive care, or death. Feature engineering created variables for each FTG criterion. An XGBoost model using prehospital vitals, demographics, and triage criteria was developed with 60/20/20 train/validate/test split. SHapley Additive exPlanations identified feature importances. Results 1 267 039 patients were included; 414 146 (32.7%) had serious injury with a mortality rate of 8.5%. Seriously injured patients were younger and more often male ( P < .001). The FTG surrogate achieved 62.9% sensitivity, 58.9% specificity, and AUC 0.609. XGBoost demonstrated superior performance with sensitivity 64.4%, specificity 67.6%, AUC 0.722, while reducing both under-triage (35.6 vs 37.1%) and over-triage rates (32.4 vs 41.1%), P < .001. GCS Verbal, sex, prehospital blood pressure, pulse oximetry, and GCS Motor were most influential features. Conclusions Machine learning outperformed a database-derived FTG surrogate on a national trauma database, reducing both under and over-triage rates. The model’s reliance on routinely collected prehospital data supports its potential for EMS use, pending prospective validation.