The trauma risk adjustment model: a new model for evaluating trauma care

Abstract

Background data: The trauma injury severity score (TRISS) has been used for over 20 years for retrospective risk assessment in trauma populations. The TRISS has serious limitations, which may compromise the validity of trauma care evaluations.

Objective: To derive and validate a new mortality prediction model, the trauma risk adjustment model (TRAM), and to compare the performance of the TRAM to that of the TRISS in terms of predictive validity and risk adjustment.

Methods: The Quebec Trauma Registry (1998-2005), based on the mandatory participation of 59 designated provincial trauma centers, was used to derive the model. The American National Trauma Data Bank (2000-2005), based on the voluntary participation of any US hospitals treating trauma, was used for the validation phase. Adult patients with blunt trauma respecting at least one of the following criteria were included: hospital stay >2 days, intensive care unit admission, death, or hospital transfer. Hospital mortality was modeled with logistic generalized additive models using cubic smoothing splines to accommodate nonlinear relations to mortality. Predictive validity was assessed with model discrimination and calibration. Risk adjustment was assessed using comparisons of risk-adjusted mortality between hospitals.

Results: The TRAM generated an area under the receiving operator curve of 0.944 and a Hosmer-Lemeshow statistic of 42 in the derivation phase. In the validation phase, the TRAM demonstrated better model discrimination and calibration than the TRISS (area under the receiving operator curve = 0.942 and 0.928, P < 0.001; Hosmer-Lemeshow statistics = 127 and 256, respectively). Replacing the TRISS with the TRAM led to a mean change of 28% in hospital risk-adjusted odds ratios of mortality.

Conclusions: Our results suggest that adopting the TRAM could improve the validity of trauma care evaluations and trauma outcome research.