Abstract
PURPOSE: Accurate pre-hospital trauma triage is essential for optimizing survival and functional outcomes within inclusive trauma systems. A prediction model incorporated into the TraumaTriage App (TTApp) has been shown to improve patient allocation in the Netherlands, but was developed based on a relatively small cohort. This study aims to redevelop and improve the performance and applicability of the TTApp’s prediction model before nationwide implementation. METHODS: In this prospective multicenter cohort study, data from all trauma patients transported within two emergency medical service (EMS) regions (i.e., Brabant and Utrecht), between February 2015 and October 2019, were used to develop and externally validate a prediction model to identify severely injured adults (Injury Severity Score [ISS] ≥ 16) using routinely available pre-hospital data. A Gradient Boosting Decision Tree (GBDT) algorithm was applied, and model performance was evaluated in terms of discrimination and calibration. RESULTS: The development cohort included 51,001 patients (median age 63.8 years, median ISS 9), and the external validation cohort included 29,737 patients (median age 62.1 years, median ISS 9). In external validation, the GBDT model showed excellent discrimination (c-statistic 0.850; 95% CI, 0.837–0.863) and good calibration (calibration-in-the-large 0.009; slope 0.952). Sensitivity was 91.3% and 85.2% at specificity thresholds of 50% and 65%, respectively. CONCLUSIONS: This externally validated prediction model effectively identifies severely injured patients in the pre-hospital setting and represents a first iteration towards enhancing the TTApp. Designed for use in urgent situations, the model allows predictions with incomplete data and shows potential to reduce undertriage rates toward 10% while maintaining acceptable overtriage levels. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00068-026-03175-8.