Predicting stroke inpatient rehabilitation outcome using a classification tree approach

A classification tree, a nonparametric statistical analysis, was used to develop decision rules to predict a favorable inpatient stroke rehabilitation outcome. Descriptive and functional status data collected on admission from 225 patients were the predictor variables. Favorable outcome was defined as having met three criteria: discharged to community, survival greater than 3 months postdischarge, and no more than minimal physical assistance required in functional activities on discharge. The classification tree correctly classified 88% of the sample using only four of the predictor variables (level of independence in Toilet Management, Bladder Management, and Toilet Transfer, and adequacy of Financial Resources). The cross validation error rate was 18%. The advantages of the classification tree approach over parametric methods are that it is desirable for ordinal data, it readily identifies the interactions among predictor variables, the results are easily communicated, and it provides additional insights into the factors that predict outcome.