Background Severe trauma remains a leading cause of admission to the intensive care unit. The Trauma and Injury Severity Score (TRISS) is an established standard for predicting outcomes and benchmarking the quality of trauma care globally. However, the TRISS model has some limitations when used for benchmarking trauma care. Objective This study aimed to determine whether machine learning–derived trauma phenotypes can complement the TRISS via multivariable modeling to improve in-hospital death prediction. We also introduce “Trauma-Vis,” a freely accessible web-based platform, to facilitate the availability of this integrated assessment approach to clinicians. Methods In this retrospective cohort study using the nationwide Japan Trauma Data Bank (JTDB), which encompasses data from 303 hospitals in Japan, we divided the data chronologically into a derivation cohort (JTDB 2015-2018) and a temporal validation cohort (JTDB 2019-2022). An integrated model was developed using multivariable logistic regression, incorporating the logit-transformed TRISS-predicted mortality and the assigned trauma phenotypes. After applying the exclusion criteria, 87,882 patients with blunt trauma were analyzed in the derivation cohort and 80,964 in the validation cohort. Predictive performance was evaluated using the area under the receiver operating characteristic curve, Brier score, logarithmic loss, net reclassification improvement, integrated discrimination improvement, and decision curve analysis. Results In the derivation cohort, multivariable modeling demonstrated that trauma phenotype classification significantly recalibrated mortality risk; multiple phenotypes exhibited significant independent associations with in-hospital death after adjusting for baseline TRISS predictions (eg, for phenotype 8: odds ratio 2.38, 95% CI 2.11-2.68; P<.001). In the temporal validation cohort, the integrated multivariable model yielded higher performance metrics than the baseline TRISS model: the area under the receiver operating characteristic curve increased from 0.889 to 0.897 (DeLong test, P<.001), Brier score improved from 0.0454 to 0.0394, and logarithmic loss decreased from 0.1670 to 0.1458. The integrated model demonstrated a calibration intercept of –0.152 and a slope of 0.965 and provided a higher net benefit in the decision curve analysis across evaluated threshold probabilities. Conclusions Integrating machine learning–derived trauma phenotypes with the TRISS via multivariable modeling improved the accuracy and utility of in-hospital death prediction. The developed “Trauma-Vis” platform demonstrates the technical feasibility of providing real-time risk stratification at the bedside.
Tachino et al. (2026) studied this question.