Background
The accurate prediction of functional outcome requires the development of multivariate models. To enhance their contribution to such models, the predictive power of each component must be optimized.
Objectives
To improve the predictive power of coma scales as the first step in building more sophisticated multivariate models to predict specific levels of functional outcome.
Design
Prospective descriptive study.
Setting
Neurology and neurosurgery intensive care unit (NNICU) in a tertiary care academic center.
Patients
Eighty-four patients with acute traumatic brain injury, intracerebral hemorrhage, subarachnoid hemorrhage, or ischemic stroke.
Interventions
None.
Main Outcome Measures
The Glasgow Coma Scale (GCS) and Innsbruck Coma Scale (ICS) were administered within 24 hours of admission to the NNICU and then at 48-hour intervals until discharge of the patient from the NNICU. The assessments were performed by 3 occupational therapy graduate students working under the supervision of the medical director of the NNICU. The functional outcome at 3 months after discharge from the hospital was assessed by telephone by the same nurse using the following categories: (1) dead, (2) receiving nursing home or custodial care, (3) home with help, or (4) independent. Cronbach's 
estimates of reliability for each scale were computed using all scores obtained during the study. The analyses indicated that the verbal response item of the GCS and the oral automatisms item of the ICS were less reliable in this patient population. The scales were modified by deleting those items, and predictive validity for the original and modified scales was computed using a discriminant function of the admission scores.
Results
Before modification, both scales were best at predicting independence (GCS and ICS, 71% correct) and mortality (GCS, 60% correct; ICS, 56% correct). The modifications produced a modest improvement in the ability of both scales to better predict levels of outcome (modified GCS: home with help, 33% correct, independent, 71% correct; modified ICS: home with help, 0% correct, independent, 74% correct).
Conclusions
By deleting items with low reliability from the ICS and the GCS we achieved improved reliability and predictive validity. The improvement in predictive power, however, was inadequate to accurately predict functional outcome. Combining clinical scales with other demographic, physiological, functional, and radiographic data will be needed to achieve useful predictions of functional outcome.
