Predicting the outcome of acute stroke: prospective evaluation of five multivariate models and comparison with simple methods.

Five multivariate models designed to predict the outcome of stroke were tested prospectively on 102 consecutive stroke patients admitted to a district general hospital. The results were compared with predictions made using two simple clinical variables (the conscious level on admission and the state of urinary continence at four weeks). Of the three models (developed in Belfast, Guy's Hospital and Uppsala) intended for use in the acute stages of stroke the last two were slightly more accurate in their prediction of death (75%) than was the admission conscious level alone (65%), whereas the Belfast model had an accuracy of only 50% in this situation. At a later stage, the state of urinary continence predicted good and poor outcomes with similar accuracy to that of a multivariate model from Edinburgh. A model developed in Bristol performed poorly. When tested prospectively, these multivariate models proved considerably less accurate than when they were first described. Only the Uppsala model showed any advantage over simple clinical methods. This might be of value in defining prognostic strata for clinical studies, but not in the management of individual patients. Simple clinical variables thus offer as much to clinicians as complex multivariate models.