Background: Traumatic brain injury (TBI) is a major cause of lost disability-adjusted life years, and a valid model allowing prediction of outcome would be welcome. For a clinical prediction model to be valid, generalization to other populations must be possible. The aim of this study was to externally validate a model for in-hospital mortality in patients with TBI, which was recently development at the University of Southern California (USC).
Methods: The validation cohort was derived from a hospital-based, prospectively collected trauma registry in Oslo, Norway. We included patients admitted with a head injury without hypotension, severe thoracic, or abdominal injury (n = 3,136). We calculated the probability of death according to the USC model. The performance of the model was evaluated using measures of calibration and discrimination in the total sample and subgroups according to initial Glasgow Coma Scale (GCS) score.
Results: The USC model provided excellent discrimination (area under the receiver operating characteristic curve, AUC = 0.93), but unsatisfactory calibration (p < 0.001) for the total sample (GCS 3-15). In the GCS 4-8 subgroup we found good discrimination (AUC = 0.89) but poor calibration (Hosmer-Lemeshow test, p < 0.001).
Conclusion: The findings question the external validity of the USC model, suggesting that it should not be implemented as a tool for short-term mortality prediction in our TBI population.