Selection of dominant voice features for accurate detection of Parkinson's disease

Abstract

Parkinson's disease (PD) is a widespread chronic neurological disease prevalent in old age. Speech is found to be an effective marker for the identification of Parkinson's disease. In the following paper, we have proposed using factor analysis to select meaningful and dominant features from the speech signals, which are relevant for prediction of Parkinson's disease. We infer that along with the jitter variants, shimmer variants and noise to harmonic ratio, pitch period entropy (PPE), the recurrence period density entropy (RPDE), and spread parameters are important in identifying PD. For classification, Support Vector Machine (SVM) is used. The proposed model discriminates Parkinson afflicted individuals from healthy ones with an average accuracy, sensitivity and specificity of about 90%. Further, from the study, it is inferred that sustained phonations carry sufficient information for predicting Parkinson's disease.