A comparison of accuracy of different Trauma and Injury Severity Score (TRISS)-like models in different trauma populations

Abstract

Background: Because of a growing demand for an easy scoring, reliable and accurate model to evaluate quality of trauma care, this study compares different Trauma and Injury Severity Score (TRISS)-like models with regard to performance and power in different trauma populations. Methods: A total number of 10,777 trauma patients admitted to our level 1 trauma centre between 1997 and 2006 were included in the analysis. The probability of survival for each patient was calculated if required data were present using the respective formulas of the prediction models of the Major Trauma Outcome Study (MTOS), Trauma Audit & Research Network (TARN) and Base Excess Injury Severity Scale (BISS). Subsequently new coefficients were calculated by logistic regression based on this dataset. Finally, the existing BISS model was extended with the Glasgow Coma Scale and also tested. The discriminative power of all original and updated models was calculated for several subsets of patients using the area under the ROC-curve (AUC), a parameter for prediction accuracy ranging from 0.5 until 1.0. Results: Far most AUCs had a value of 0.8 or more. For the total population the TARN update 2007 model with new coefficients had the highest AUC (0.924). For the subset of patients in which all parameters for the various models were available the BISS model including GCS had the highest AUC (0.909). All models had a high discriminative power (AUC range 0.878–0.990) if patients were younger than 55 years. In older patients, patients with severe head injury or intubated patients the discriminative power of the prediction models dropped. Conclusions: Relative simple models, like MTOS, TARN and BISS predict mortality pretty reliably. Each model tested in our study had specific advantages and disadvantages. Discrimination power strongly depended on the case mix. The accuracy to predict the chance of survival decreases in severely injured and older patients head injury and comorbidity are likely to attribute to this phenomenon. Therefore adjustment for these factors in future models might be necessary.