Prognosis following severe head injury: Development and validation of a model for prediction of death, disability, and functional recovery

Background: A prognostic model for head trauma patients is useful only if it predicts clinically relevant outcomes accurately on new subjects in various settings. Most existing models consider only dichotomous outcome and have not been tested externally. We developed and validated a rule for prediction of three functional outcome states after severe head injury, using information from day 1. Methods: The model was developed in a cohort of 304 adults who were admitted to a Dutch trauma center and had survived and remained comatose for > 24 hours following severe head injury. We used ordinal logistic regression analysis to predict the extended Glasgow Outcome Scale after 2:12 months, merged into three categories. We preselected five known predictors of outcome and used bootstrapping techniques to avoid statistical overfitting. The performance of the model was subsequently tested in a cohort of 122 patients from an unrelated hospital. Results: The model contained age (p < 0.0001), best motor response on day 1 (p = 0.002), pupil response after resuscitation (p = 0.005), computed tomography findings (p = 0.004), and presence of arterial hypotension (p = 0.37) as predictor variables. In the external validation cohort the model showed adequate agreement between observed and predicted outcome probabilities (calibration). The model had a good ability to discriminate patients with different outcomes (c-statistic 0.808). The predictive accuracy was 66% when the model was used to classify patients across the three outcome categories. Conclusions: We have developed a practical model for predicting the probability of death, survival with major disability, and functional recovery in patients who are comatose 24 hours after severe head injury. The model performed well in an external setting, indicating that measures to avoid statistical overfitting were successful.