Classification and regression trees (CART) and support vector machines (SVM) have become very popular statistical learning tools for analyzing complex data that often arise in biomedical research. While both CART and SVM serve as powerful classifiers in many clinical settings, there are some common scenarios in which each fails to meet the performance and interpretability needed for use as a clinical decision-making tool. In this paper, we propose a new classification method, SVM-CART, that combines features of SVM and CART to produce a more flexible classifier that has the potential to outperform either method in terms of interpretability and prediction accuracy. Further-more, to enhance prediction accuracy we provide extensions of a single SVM-CART to an ensemble, and methods to extract a representative classifier from the SVM-CART ensemble. The goal is to produce a decision-making tool that can be used in the clinical setting, while still harnessing the stability and predictive improvements gained through developing the SVM-CART ensemble. An extensive simulation study is conducted to asses the performance of the methods in various settings. Finally, we illustrate our methods using a clinical neuropathy dataset.
Keywords: Classification and Regression Trees; Complex Interactions; Ensemble Classifiers; Statistical Learning; Support Vector Machines.