Enhancing Performance of The National Field Triage Guidelines Using Machine Learning: Development of a Prehospital Triage Model to Predict Serious Trauma (pTEST) (Preprint)

Abstract

Prehospital trauma triage is essential to get the right patient to the right hospital. However, the national field triage guidelines proposed by the American College of Surgeons proved relatively insensitive when identifying severe traumas.

To build a prehospital triage model to predict severe trauma and enhance the performance of the national field triage guidelines.

This is a multi-site prediction study, and the data were extracted from the National Trauma Data Bank between 2017 and 2019. All patients with injury, aged ≥ 16 years, transported by ambulance from the injury scene to any trauma center were potentially eligible. The data were divided into training, internal, and external validation sets of 672309, 288134, and 508703 patients. As national field triage guidelines recommended, age, seven vital signs, and eight injury patterns at the pre-hospital stage were included as candidate variables for model development. Outcomes are severe trauma with Injured Severity Score ≥16 (primary) and critical resource use within 24 h of emergency department arrival (secondary). The triage model was developed using an extreme gradient boosting model and Shapley additive explanation analysis. The model’s accuracy regarding discrimination, calibration, and clinical utility was assessed.

At a fixed specificity of 0.5, the model showed a sensitivity of 0.799(0.797–0.801), an undertriage rate of 0.080(0.079–0.081), and an overtriage rate of 0.743(0.742–0.743) for predicting severe trauma. The model showed a sensitivity of 0.774(0.772–0.776), an undertriage rate of 0.158(0.157–0.159), and an overtriage rate of 0.609(0.608–0.609) when predicting critical resource use, fixed at 50% specificity. The triage model’s areas under the curve were 0.755(0.753–0.757) for severe trauma prediction and 0.736(0.734–0.737) for critical resource use prediction. The triage model’s performance was better than those of the Glasgow Coma Score, Prehospital Index, revised trauma score, and the 2011 national field triage guidelines RED criteria. The model’s performance was consistent in the two validation sets.

The prehospital triage model is promising for predicting severe trauma and achieving an undertriage rate of < 10%. Moreover, machine learning enhances the performance of field triage guidelines.