Predicting long-term neurological outcomes after severe traumatic brain injury requiring decompressive craniectomy: A comparison of the CRASH and IMPACT prognostic models

Abstract

BackgroundPredicting long-term neurological outcomes after severe traumatic brain (TBI) is important, but which prognostic model in the context of decompressive craniectomy has the best performance remains uncertain. MethodsThis prospective observational cohort study included all patients who had severe TBI requiring decompressive craniectomy between 2004 and 2014, in the two neurosurgical centres in Perth, Western Australia. Severe disability, vegetative state, or death were defined as unfavourable neurological outcomes. Area under the receiver-operating-characteristic curve (AUROC) and slope and intercept of the calibration curve were used to assess discrimination and calibration of the CRASH (Corticosteroid-Randomisation-After-Significant-Head injury) and IMPACT (International-Mission-For-Prognosis-And-Clinical-Trial) models, respectively. ResultsOf the 319 patients included in the study, 119 (37%) had unfavourable neurological outcomes at 18-month after decompressive craniectomy for severe TBI. Both CRASH (AUROC 0.86, 95% confidence interval 0.81–0.90) and IMPACT full-model (AUROC 0.85, 95% CI 0.80–0.89) were similar in discriminating between favourable and unfavourable neurological outcome at 18-month after surgery (p=0.690 for the difference in AUROC derived from the two models). Although both models tended to over-predict the risks of long-term unfavourable outcome, the IMPACT model had a slightly better calibration than the CRASH model (intercept of the calibration curve=−4.1 vs. −5.7, and log likelihoods −159 vs. −360, respectively), especially when the predicted risks of unfavourable outcome were <80%. ConclusionsBoth CRASH and IMPACT prognostic models were good in discriminating between favourable and unfavourable long-term neurological outcome for patients with severe TBI requiring decompressive craniectomy, but the calibration of the IMPACT full-model was better than the CRASH model.