Predicting outcomes after traumatic brain injury: the development and validation of prognostic models based on admission characteristics.

Background: Early estimation of prognosis for the patient with traumatic brain injury is an important factor in making treatment decisions, resource allocation, classify patients, or communicating with family. We aimed to develop and validate practical prognostic models for mortality at 30 days and for 6 months unfavorable outcome after moderate and severe traumatic brain injury.

Methods: Retrospectively collected data from our department were used to develop prognostic models for outcome. We developed four prognostic models based on admission predictors with logistic regression analysis. The performance of models was assessed with respect to discrimination and calibration. Discriminative ability was evaluated with C statistic, equal to the area under the receiver operating characteristic curve. Calibrative ability was assessed with the Hosmer-Lemeshow test (H-L test). The internal validity of models was evaluated with the bootstrap re-sampling technique. We validated three of the models in an external series of 203 patients that collected from another research center. Discrimination and calibration were further assessed to indicate the performance of the models in external patients.

Results: Logistic regression showed that age, pupillary reactivity, motor Glasgow Coma Score, computed tomography characters, glucose, hemoglobin, D-dimer, serum calcium, and intracranial pressure were independent prognostic factors of outcome. The models discriminated well in the development patients (C statistic 0.709-0.939). We extensively validate three of the models. Internal validation showed no overoptimism in any of the models' predictive C statistics. External validity was much better (C statistic 0.844-0.902). Calibration was also good (H-L tests, p > 0.05). Computer-based calculator that based on prognostic models was developed for clinical use.

Conclusion: Our validated prognostic models have good performance and are generalizable to be used to predict outcome of new patients. We recommend the use of prognostic models to complement clinical decision making.