Classification of Parkinsonian Syndromes from FDG-PET Brain Data Using Decision Trees with SSM/PCA Features

Table 5

The LOOCV performance for various types of classifier. Features used were the subject scores obtained after applying the SSM/PCA method on all subjects included in the datasets. () Note that for LDA only 90% of the features were considered because of the classifier’s restrictions while constructing the covariance matrix. For easy reference, the feature preselection results for C4.5 already presented in Table 2 are included.

Dataset	PD-HC	MSA-HC	PSP-HC
Nearest neighbors	76.3	76.9	80.0
Linear SVM	78.9	92.3	88.6
Random forest	63.2	61.5	71.4
Naive Bayes	65.8	71.8	71.4
LDA ()	50.0	61.5	65.7
CART	57.9	53.8	85.7
C4.5	63.2	74.4	82.9