Validation of the CRASH model in the prediction of 18-month mortality and unfavorable outcome in severe traumatic brain injury requiring decompressive craniectomy

Object. The goal in this study was to assess the validity of the corticosteroid randomization after significant head injury (CRASH) collaborators prediction model in predicting mortality and unfavorable outcome at 18 months in patients with severe traumatic brain injury (TBI) requiring decompressive craniectomy. In addition, the authors aimed to assess whether this model was well calibrated in predicting outcome across a wide spectrum of severity of TBI requiring decompressive craniectomy. Methods. This prospective observational cohort study included all patients who underwent a decompressive craniectomy following severe TBI at the two major trauma hospitals in Western Australia between 2004 and 2012 and for whom 18-month follow-up data were available. Clinical and radiological data on initial presentation were entered into the Web-based model and the predicted outcome was compared with the observed outcome. In validating the CRASH model, the authors used area under the receiver operating characteristic curve to assess the ability of the CRASH model to differentiate between favorable and unfavorable outcomes. Results. The ability of the CRASH 6-month unfavorable prediction model to differentiate between unfavorable and favorable outcomes at 18 months after decompressive craniectomy was good (area under the receiver operating characteristic curve 0.85, 95% CI 0.80-0.90). However, the model's calibration was not perfect. The slope and the intercept of the calibration curve were 1.66 (SE 0.21) and -1.11 (SE 0.14), respectively, suggesting that the predicted risks of unfavorable outcomes were not sufficiently extreme or different across different risk strata and were systematically too high (or overly pessimistic), respectively. Conclusions. The CRASH collaborators prediction model can be used as a surrogate index of injury severity to stratify patients according to injury severity. However, clinical decisions should not be based solely on the predicted risks derived from the model, because the number of patients in each predicted risk stratum was still relatively small and hence the results were relatively imprecise. Notwithstanding these limitations, the model may add to a clinician's ability to have better-informed conversations with colleagues and patients' relatives about prognosis.