Beyond Thresholds: Can Machine Learning Improve Trauma Field Triage?
R L Wolansky1, M A Kendall1, L Hiraldo1
1Department of Surgery, University of South Florida College of Medicine, Tampa, FL, USA.
The American Surgeon
|July 24, 2026
Summary
A machine learning model significantly improved trauma patient triage accuracy compared to existing guidelines. This advanced model reduces both under-triage and over-triage, enhancing emergency medical services (EMS) for better patient outcomes.
Area of Science:
- Emergency Medicine
- Data Science
- Trauma Surgery
Background:
- Accurate prehospital triage of trauma patients by Emergency Medical Services (EMS) is critical for optimal outcomes and resource allocation.
- The 2021 National Field Triage Guidelines (FTG) provide a framework for EMS triage, but their collective performance using national data remains unevaluated.
- There is a need to assess current triage performance and develop improved predictive models for identifying seriously injured trauma patients.
Purpose of the Study:
- To evaluate a surrogate for the National Field Triage Guidelines (FTG) using a national trauma database.
- To develop and validate a predictive machine learning model for identifying patients at high risk for serious injury during prehospital care.
- To compare the performance of the machine learning model against the FTG surrogate in terms of sensitivity, specificity, and triage accuracy.
Main Methods:
- Utilized the Trauma Quality Improvement Program National Trauma Databank (2017-2020) including over 1.2 million traumatically injured adult patients.
- Defined serious injury using criteria such as Injury Severity Score (ISS) ≥16, blood transfusion, or specific surgical interventions.
- Developed an XGBoost machine learning model incorporating prehospital vital signs, demographics, and FTG criteria, with feature importance analysis using SHapley Additive exPlanations.
Main Results:
- The study included 1,267,039 patients, with 32.7% experiencing serious injury and an 8.5% mortality rate.
- The FTG surrogate demonstrated moderate performance (AUC 0.609), while the XGBoost model showed superior performance (AUC 0.722).
- The XGBoost model significantly reduced both under-triage (35.6% vs 37.1%) and over-triage (32.4% vs 41.1%) rates compared to the FTG surrogate.
Conclusions:
- Machine learning models significantly outperform database-derived FTG surrogates in identifying seriously injured trauma patients.
- The developed XGBoost model demonstrates potential for improving prehospital trauma triage by reducing misclassification rates.
- The model's reliance on routinely collected prehospital data suggests feasibility for EMS implementation, pending prospective validation.