Factors associated with mortality in trauma: re-evaluation of the TRISS method using the National Trauma Data Bank

BACKGROUND: TRISS remains a standard method for predicting survival and correcting for severity in outcome analysis. The National Trauma Data Bank (NTDB) is emerging as a major source of trauma data that will be used for both primary research and outcome benchmarking. We used NTDB data, to determine whether TRISS is still an accurate predictor of survival coefficients and to determine whether the ability of TRISS to predict survival could be improved by updating the coefficients or by building predictive models that include information on co-morbidities. METHODS: To compare the utility of different methods of TRISS calculation we identified the records of 72,517 trauma patients (62,103 blunt trauma and 10,414 penetrating trauma) who had complete information for all of the covariates to be considered in the analysis. Multiple logistic regression was used to recalculate the TRISS coefficients in models using both the original TRISS covariates and in models which also included variables for co-morbidities that could potentially affect survival. Model discrimination was evaluated by calculating the area under the receiver operating characteristic curves (AUC), and model calibration was evaluated with the Hosmer-Lemeshow Goodness-of-Fit Statistic (H-L). RESULTS: For penetrating trauma the original TRISS equation had good discriminative ability (AUC=0.98), but was poorly calibrated (H-L=267.04). When logistic regression was used to generate revised coefficients, discrimination was unchanged, but calibration improved (H-L=38.66). The only co-morbid factor significantly associated with survival after penetrating trauma was acute alcohol consumption, which was associated with increased survival (p < 0.0001). However, its inclusion in a logistic model did not improve discrimination, but improved calibration somewhat (AUC =0.98; H-L=19.95). The original TRISS equation was a less accurate predictor of survival after blunt trauma (AUC = 0.84; H-L= 10,720.7). When logistic regression was used to generate revised coefficients for the original TRISS covariates, predictions after blunt trauma improved (AUC = 0.94; H-L=25.45). With blunt trauma, acute alcohol consumption and prior hypertension were associated with increased survival, and male gender, congestive failure, cirrhosis, and prior myocardial infarction were associated with decreased survival. However, inclusion of these covariates in a logistic model did not improve predictions of survival (AUC = 0.94; H-L= 34.83). CONCLUSIONS: In the NTDB the traditional TRISS had limited ability to predict survival after trauma. Accuracy of prediction was improved by recalculating the TRISS coefficients, but further improvements were not seen with models that included information about co-morbidities.