• Nino A Mushkudiani, Chantal W P M Hukkelhoven, Adrián V Hernández, Gordon D Murray, Sung C Choi, Andrew I R Maas, and Ewout W Steyerberg.
• Center for Medical Decision Making, Department of Public Health, Erasmus MC, P.O. Box 2040, 3000 CA Rotterdam, The Netherlands.
• J Clin Epidemiol. 2008 Apr 1;61(4):331-43.

ObjectivesTo describe the modeling techniques used for early prediction of outcome in traumatic brain injury (TBI) and to identify aspects for potential improvements.Study Design And SettingWe reviewed key methodological aspects of studies published between 1970 and 2005 that proposed a prognostic model for the Glasgow Outcome Scale of TBI based on admission data.ResultsWe included 31 papers. Twenty-four were single-center studies, and 22 reported on fewer than 500 patients. The median of the number of initially considered predictors was eight, and on average five of these were selected for the prognostic model, generally including age, Glasgow Coma Score (or only motor score), and pupillary reactivity. The most common statistical technique was logistic regression with stepwise selection of predictors. Model performance was often quantified by accuracy rate rather than by more appropriate measures such as the area under the receiver-operating characteristic curve. Model validity was addressed in 15 studies, but mostly used a simple split-sample approach, and external validation was performed in only four studies.ConclusionAlthough most models agree on the three most important predictors, many were developed on small sample sizes within single centers and hence lack generalizability. Modeling strategies have to be improved, and include external validation.

12 keywords...

hide…