Presence of aphasia in patients with stroke poses a problem in the use of the full form (eye-motor-verbal) Glasgow Coma Scale (GCS). Stroke investigators and clinicians have used three different strategies to deal with the untestable verbal subscale, i.e. eliminating the verbal subscale; pseudoscoring with ‘one’, and median value substitution; but the predictive accuracy of the strategies has not been compared. To compare the predictive accuracy of the three strategies for acute mortality in stroke, we prospectively applied the GCS to 275 consecutive patients with acute stroke and recorded their survival status before discharge from hospital. 95 (33.8%) patients died. 32 (12%) patients had untestable verbal score. Receiver-Operator-Characteristic curves for predicting mortality were constructed with the GCS sum score and with the multivariate logistic models, and areas under the curves were measured to compare the predictive accuracy. They were all found to be similar (0.87–0.88 sq unit). Specifically, the GCS with eye and motor subscale had 87% accuracy compared to 88% for the model with eye, motor and verbal scale. We conclude that the short-form (eye-motor) GCS is as good a predictor of early mortality (within 2 weeks) as the full form (eye-motor-verbal) GCS in patients with stroke.