Comparison of decision tree and stepwise regression methods in classification of FDG-PET brain data using SSM/PCA features

Abstract

Objective: To compare the stepwise regression (SR) method and the decision tree (DT) method for classification of parkinsonian syndromes. Method: We applied the scaled subprofile model/principal component analysis (SSM/PCA) method to FDG-PET brain image data to obtain covariance patterns and the corresponding subject scores. The subject scores formed the input to the C4.5 decision tree algorithm to classify the subject brain images. For the SR method, scatter plots and receiver operating characteristic (ROC) curves indicate the subject classifications. We then compare the decision tree classifier results with those of the SR method. Results: We found out that the SR method performs slightly better than the DT method. We attribute this to the fact that the SR method uses a linear combination of the best features to form one robust feature, unlike the DT method. However, when the same robust feature is used as the input for the DT classifier, the performance is as high as that of the SR method. Conclusion: Even though the SR method performs better than the DT method, including the SR procedure in the DT classification yields a better performance. Additionally, the decision tree approach is more suitable for human interpretation and exploration than the SR method.