The Glasgow Coma Scale (GCS) has served as an assessment tool in head trauma and as a measure of physiologic derangement in outcome models (e.g., TRISS and Acute Physiology and Chronic Health Evaluation), but it has not been rigorously examined as a predictor of outcome.
Using a large trauma data set (National Trauma Data Bank, N = 204,181), we compared the predictive power (pseudo R2, receiver operating characteristic [ROC]) and calibration of the GCS to its components.
The GCS is actually a collection of 120 different combinations of its 3 predictors grouped into 12 different scores by simple addition (motor [m] + verbal [v] + eye [e] = GCS score). Problematically, different combinations summing to a single GCS score may actually have very different mortalities. For example, the GCS score of 4 can represent any of three mve combinations: 2/1/1 (survival = 0.52), 1/2/1 (survival = 0.73), or 1/1/2 (survival = 0.81). In addition, the relationship between GCS score and survival is not linear, and furthermore, a logistic model based on GCS score is poorly calibrated even after fractional polynomial transformation. The m component of the GCS, by contrast, is not only linearly related to survival, but preserves almost all the predictive power of the GCS (ROC(GCS) = 0.89, ROC(m) = 0.87; pseudo R2(GCS) = 0.42, pseudo R2(m) = 0.40) and has a better calibrated logistic model.
Because the motor component of the GCS contains virtually all the information of the GCS itself, can be measured in intubated patients, and is much better behaved statistically than the GCS, we believe that the motor component of the GCS should replace the GCS in outcome prediction models. Because the m component is nonlinear in the log odds of survival, however, it should be mathematically transformed before its inclusion in broader outcome prediction models.