Validation of the IMPACT outcome prediction score using the Nottingham Head Injury Register dataset

Comparison of traumatic brain injury (TBI) outcomes is severely limited by the absence of a universally accepted and validated outcome prediction score. The IMPACT group recently reported models predicting mortality and unfavorable outcome after TBI, based on the outcomes of patients with moderate and severe head injury reported in two large clinical trials.

We have used prospectively collected data from 1,276 adult patients from the Nottingham Head Injury Register admitted to a single UK neurosurgical unit during a 10-year period to validate the IMPACT score models. The two models were validated for discrimination, calibration, and accuracy, using multiple imputation to adjust for missing data.

One thousand sixty-one patients (83%) had a complete set of data. For the multiply imputed analysis, the IMPACT prognostic models showed satisfactory discrimination (area under the receiver operator curve for mortality, 0.835; 95% confidence interval, 0.811-0.858; unfavorable outcome, 0.828; 95% confidence interval, 0.805-0.851) and accuracy (Brier Accuracy Score for mortality, 0.403, p < 0.01; unfavorable outcome, 0.371, p < 0.01). Good calibration was evident for unfavorable outcome, but mortality risk was underestimated by the scoring system in our sample (Hosmer-Lemeshow test: mortality: p < 0.01; unfavorable outcome: p = 0.6). These results were not significantly changed when repeated using patients with complete data only.

The 2005 IMPACT model for unfavorable outcome performs well when used to predict outcome in adults with moderate and severe TBI presenting to a British neurosurgical center. However, the model for mortality fitted less well, slightly overestimating mortality in the higher-risk groups.