A novel octopus based Parkinson’s disease and gender recognition method using vowels

Abstract

The Parkinson’s disease (PD) is one of the widely seen and most important neurological disorders worldwide. With the development of the technology, many machine learning methods have been presented to recognize PD automatically. In order to recognize PD and gender, vowels have been widely used and many papers have been presented for solving these problems in the literature. In this study, a novel octopus based feature extraction network is presented and the proposed octopus is a multiple pooling method. In this method, minimum, maximum, maximum-minimum, average, variance, median, skewness and kurtosis pooling methods are used. These eight pooling methods consist the leg of the octopus. In this article, a vowel recognition method is proposed using the proposed octopus pooling method. The proposed method contains preprocessing, feature extraction, feature selection, classification and post processing phases. In the preprocessing, the proposed octopus method is applied to signal to generate octopus signal. Singular Value Decomposition (SVD) is utilized as feature extractor and the features are extracted using original vowel signal and the signals of the octopus. In order to feature selection, neighborhood component analysis (NCA) is used to remove redundant features. In the classification phase, support vector machine with various activation functions (linear, cubic, radial bases function), 1NN with Manhattan distance, tree and logistic regression are utilized. To obtain individual results, the proposed post processing algorithm is applied to validation predictions. In order to show success of the proposed method, a vowel dataset is used. This dataset contains PD disease vowels and there are gender labels. By using the proposed octopus based method, PD, gender and both PD and gender recognitions are performed. The proposed method achieved 99.21%, 98.41% and 97.62% accuracy rates for gender, PD and gender and PD classification respectively using 1 nearest neighbor (1NN) classifier. The space complexity of the proposed method was calculated and was found as Onlogn. These results clearly indicated that the proposed solves three problems with high success rates and low computational complexity.