An efficient dimensionality reduction method using filter-based feature selection and variational autoencoders on Parkinson's disease classification

Parkinson's disease (Pd) is a progressive disease caused by the loss of brain cells and brings about speech and pronunciation defects during the early stages. This study revealed a Pd classification system based on vocal features extracted from the voice recordings of the individuals and proposed a hybrid dimensionality reduction methods to extract robust features. Proposed method took advantage of the prominent aspects of Variational Autoencoders (VAE) and filter-based feature selection models. Relief and Fisher Score were selected as filter-based methods for their effective performance in handling noisy data while VAE was used as a feature extractor due to the capability of preserving the regular latent space properties during the feature generation. In order to assess the effectiveness of the devised method, multi-kernel Support Vector Machines (SVM) classifier were trained with obtained deep feature representations. The combination of deep Relief features and SVM with multiple kernels distinguished Pd individuals from healthy subjects with an accuracy of 0.916 with 0.772 Matthews Correlation Coefficient (MCC) rates using only 30 features. Compared to results obtained without dimensionality reduction, proposed model provided approximately 9% and 22% improvements on accuracy and MCC rates, respectively. All experimental results showed that models trained with the deep features had higher accuracy and MCC rates with those trained with Fisher Score and Relief selected features. In addition, all models trained with reduced features had higher classification performance than the model without selection. It was also concluded that using multiple kernels in the SVM boosted the classification performance.

Automated summary

This study proposed a Parkinson's disease classification system based on vocal features and introduced a hybrid dimensionality reduction method to extract robust features. By using a multi-kernel Support Vector Machine classifier, the method was able to successfully distinguish Pd individuals from healthy subjects with an accuracy of 0.916.