Features selection by genetic algorithm optimization with k-nearest neighbour and learning ensemble to predict Parkinson disease

Abstract

Among the several ways followed for detecting Parkinson's disease, there is the one based on the speech signal, which is a symptom of this disease. In this paper focusing on the signal analysis, a data of voice records has been used. In these records, the patients were asked to utter vowels “a”, “o”, and “u”. Discrete wavelet transforms (DWT) applied to the speech signal to fetch the variable resolution that could hide the most important information about the patients. From the approximation a3 obtained by Daubechies wavelet at the scale 2 level 3, 21 features have been extracted: a linear predictive coding (LPC), energy, zero-crossing rate (ZCR), mel frequency cepstral coefficient (MFCC), and wavelet Shannon entropy. Then for the classification, the K-nearest neighbour (KNN) has been used. The KNN is a type of instance-based learning that can make a decision based on approximated local functions, besides the ensemble learning. However, through the learning process, the choice of the training features can have a significant impact on overall the process. So, here it stands out the role of the genetic algorithm (GA) to select the best training features that give the best accurate classification.