Enhancing Performance of the National Field Triage Guidelines Using Machine Learning: Development of a Prehospital Triage Model to Predict Severe Trauma
Qi Chen,Yuchen Qin,Zhichao Jin,Xinxin Zhao,Jia He,Cheng Wu,Bihan Tang
DOI: https://doi.org/10.2196/58740
IF: 7.076
2024-10-01
Journal of Medical Internet Research
Abstract:Background: 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 have proven to be relatively insensitive when identifying severe traumas. Objective: This study aimed to build a prehospital triage model to predict severe trauma and enhance the performance of the national field triage guidelines. Methods: This was a multisite prediction study, and the data were extracted from the National Trauma Data Bank between 2017 and 2019. All patients with injury, aged 16 years of age or older, and 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 672,309; 288,134; and 508,703 patients, respectively. As the national field triage guidelines recommended, age, 7 vital signs, and 8 injury patterns at the prehospital stage were included as candidate variables for model development. Outcomes were severe trauma with an Injured Severity Score ≥16 (primary) and critical resource use within 24 hours 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 benefit was assessed. Results: At a fixed specificity of 0.5, the model showed a sensitivity of 0.799 (95% CI 0.797-0.801), an undertriage rate of 0.080 (95% CI 0.079-0.081), and an overtriage rate of 0.743 (95% CI 0.742-0.743) for predicting severe trauma. The model showed a sensitivity of 0.774 (95% CI 0.772-0.776), an undertriage rate of 0.158 (95% CI 0.157-0.159), and an overtriage rate of 0.609 (95% CI 0.608-0.609) when predicting critical resource use, fixed at 0.5 specificity. The triage model's areas under the curve were 0.755 (95% CI 0.753-0.757) for severe trauma prediction and 0.736 (95% CI 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 2 validation sets. Conclusions: 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.
health care sciences & services,medical informatics