**4. Classification results**

*Feature selection* was applied to reduce the dimensionality of feature vectors. Feature selection was conducted by running *Wrappers feature subset selection*, which results in an optimal subset of features for a specific classifier. Feature subset selection was accomplished for each classifier. The most relevant groups of features, selected by Wrappers, were the *Mel Frequency Cepstral Coefficients* (MFCC) and the *Tunnable Q-factor Wavelet Transform* (TQWT) features. MFCCs are based on the way the human auditory system works. The computation of MFCCs involves the use of multiple band-pass filters where the filter bandwidth is increased as the central frequency is higher. In the work by Sakar et al. [7], the two most relevant groups of features were TQWT and MFCCs, and the work by Solana-Lavalle et al. [8] is also based on these features.

Multiple classifiers have been applied to the problem of voice-based PD detection such as the *k Nearest Neighbors* (kNN), *Multi-Layer Perceptron* (MLP), *Random Forest* (RF), and the *Support Vector Machine* (SVM). However, after conducting separate studies for male and female populations, it was found that the classifiers, with the highest detection performance, were (1) the *Support Vector Machines* (SVM) with a *radial basis function kernel* (RBF), and (2) the *Multi-Layer perceptron* (MLP) [6]. The highest accuracy reported is 94%, which is a considerable improvement over the previous works that used the same dataset [7, 8]. In addition, the complexity of the last reported model has been reduced from 50 to only 20 used features.
