Verifying clinical criteria for parkinsonian disorders with CART decision trees

Abstract

The possibility for an expert to verify and evaluate a decision tree is the major advantage of using this machine learning method, especially for medical diagnostics. In this paper we explore the use of a machine learning method based on decision trees using CART for verifying clinically established diagnostic criteria and also for seeking new criteria in different autopsyconfirmed Parkinsonian disorders. Since differentiating various types of Parkinsonian disorders can often present great difficulties due to the overlapping of clinical signs and symptoms, we present a strategy for extracting additional attributes from our database. From the clinical point of view, we obtained interesting results that confirm the importance of already established diagnostic criteria, but we also found some attributes (signs and symptoms) which deserve closer clinical observation. The compatibility of results obtained by our method with those from already established clinical criteria speaks in favor of the validity of the method. © Springer-Verlag 2004.