JMIR Publications

ABSTRACT

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:

We aimed to assess whether the previously identified machine learning-derived trauma phenotypes can complement the TRISS in identifying high-risk subgroups (i.e., “unexpected deaths”) and to introduce “Trauma-Vis,” an open-source online platform (available at https://github.com/jotarotachino/Trauma-Vis), to facilitate the availability of this integrated assessment approach to front-line clinicians.

Methods:

In this retrospective cohort study using the Japan Trauma Data Bank (JTDB), which encompasses data from 303 hospitals (tertiary care and emergency centers) in Japan, we developed an integrated model with a derivation cohort (JTDB 2019–2022) and assessed the model’s performance using a temporally distinct validation cohort (JTDB 2015–2018). After applying exclusion criteria, 80,964 patients with blunt trauma were analyzed in the derivation cohort and 87,882 in the validation cohort.

Results:

The TRISS demonstrated good discrimination [area under the receiver operating characteristic (AUROC): 0.889, 95% confidence interval (CI): 0.884–0.894] but poor calibration, systematically underestimating survival in low-probability cases. Among eight trauma phenotypes identified, phenotype 8 exhibited a disproportionately high rate of unexpected death (31.6%). By adjusting the predicted survival probability to 0.5 for phenotype 8 patients with TRISS-predicted survival probability >0.5, the integrated model accurately reclassified 25.9% of previously misidentified deaths while maintaining discriminative performance (AUROC: 0.892, 95% CI: 0.887–0.897; DeLong’s test p=0.020). Validation tests confirmed the model’s utility, revealing a 23.1% improvement in identification of unexpected deaths, despite not change in overall discrimination (both AUROC: 0.912, 95% CI: 0.908–0.916; p=0.545).

Conclusions:

Integrating trauma phenotypes can improve the accuracy of the TRISS by identifying high-risk subgroups missed by conventional scoring. “Trauma-Vis” demonstrates the technical feasibility of translating this approach to the bedside. These findings, once confirmed via true external validation, can aid in adoption of the integrated approach into clinical practice.