To Register      SMDM Homepage

Tuesday, 19 October 2004

This presentation is part of: Poster Session - Clinical Strategies; Judgment and Decison Making

USING CLASSIFICATION AND REGRESSION TREE ANALYSIS TO IDENTIFY CLINICAL PREDICTORS OF UTILITIES IN PARKINSON’S DISEASE

Uwe Siebert, MD, MPH, MSc, Massachusetts General Hospital, Harvard Medical School, Institute for Technology Assessment and Department of Radiology, Boston, MA, Bernhard Bornschein, MD, MPH, Ludwig Maximilians University Munich, Germany, Program on HTA and Decision Sciences, Institute of Medical Informatics, Biostatistics, and Epidemiology, Munich, Germany, and Richard C. Dodel, MD, University of Bonn, Department of Neurology, Bonn, Germany.

Purpose: To identify the most relevant items of a commonly used clinical rating scale in Parkinson’s disease for the prediction of utilities.

Methods: We used data from a prospective clinical study of the German Competence Network for Parkinson Syndromes (n=122). We used 38 single items of the Unified Parkinson’s Disease Rating Scale (parts II-IV) as potential predictors and utilities derived from EuroQol (EQ-5D) as outcome. We performed a classification and regression tree analysis (CART) with the t-test as test statistic for group selection and adjusted p-values for non-dichotomous variables by the method of Miller & Siegmund and the square-root (n)-method as stopping rule. To determine the explained variance, we entered the identified groups as indicator variables into a linear regression model.

Results: The final CART model had 3 levels with 4 variables partitioning the sample into 5 subgroups. These groups were defined by the degree of posture (item 28), the level of rigidity (item 22), problems with arising from a chair (item 27), and unpredictable fluctuations (item 36) Figure 1 shows the mean utilities in the 5 subgroups with 95% confidence intervals. Explained variance (adjusted R-square) was 0.50.

Conclusions: We successfully applied CART analysis to identify a parsimonious tree that predicts mean utilities based on clinical rating scale values. Our prediction tree is a simple tool that can easily be applied in the routine health care of bedside decision making in Parkinson’s disease. However, these results need to be externally validated with independent data.

Figure 1: Utility subgroups in Parkinson's disease represented by a classification and regression tree (CART), based on 122 patients. Results at the "leaves" of the tree denote mean subgroup utilities with 95%-confidence intervals.


See more of Poster Session - Clinical Strategies; Judgment and Decison Making
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)