Abstract
BACKGROUND:
Trauma programs continue to use the Trauma and Injury Severity Score (TRISS) to estimate expected mortality for quality review, although its coefficients reflect an older population and care era. We developed Contemporary Expected Mortality for Trauma (CEM-Trauma) for the registry-based benchmarking use case historically served by TRISS.
METHODS:
We hypothesized that a contemporary model using familiar injury, age, mechanism, and early physiology variables would more accurately predict in-hospital death than both historical and updated TRISS. We studied 5,399,314 TRISS-eligible Trauma Quality Improvement Program patients (2017–2024), trained models on 2017–2022, and tested them in 2023–2024. Comparators were historical TRISS, development-set intercept/slope recalibration, and development-set refitting of RTS, Injury Severity Score, and age coefficients separately by mechanism. We compared included and excluded patients, calculated 95% CIs, assessed prespecified subgroups, and tested age-by-injury, age-by-physiology, and mechanism-by-physiology interactions.
RESULTS:
The temporal test set included 1,581,953 patients and 40,674 deaths (2.6%). Full CEM-Trauma improved AUC versus historical TRISS and coefficient-refit TRISS (0.909 [95% CI, 0.907–0.910] vs. 0.879 [0.877–0.881] and 0.879 [0.877–0.881]) and reduced Brier score (0.01891 [0.01873–0.01908] vs. 0.02086 [0.02068–0.02104] and 0.01997 [0.01979–0.02016]). Updated TRISS corrected most overall calibration bias (O/E: 0.986 after recalibration and 0.981 after coefficient refitting); CEM-Trauma O/E was 0.957. CEM-Trauma improved the area under the receiver operating characteristic curve (AUC) across all prespecified subgroups, including patients aged 65 years or older (0.851 vs. 0.804) and those with a Glasgow Coma Scale score of 8 or lower (0.826 vs. 0.780). Tested interactions did not materially improve held-out performance. Among excluded patients with recorded outcomes, mortality was 5.9%; however, 48.3% lacked mortality data.
CONCLUSIONS:
CEM-Trauma improved risk ranking and average prediction error beyond both historical and coefficient-refit TRISS, whereas contemporary TRISS updating largely corrected overall calibration. These findings distinguish the benefit of recalibration from that of new predictors and model form. External validation is required before benchmarking adoption; the model is not intended for bedside decisions.
LEVEL OF EVIDENCE:
Prognostic and Epidemiological; Level III.