A Comparative Study of Acoustic and Clinical Speech Feature Vectors for Parkinson's Disease Detection with a Multilayer Perceptron Classifier

Abstract

Parkinson’s disease (PD) is an advancing neurodegenerative disease that can have speech loss during preliminary stages of the disease, thus allowing detection of the disease using non-invasive and convenient analysis of the voice. In this paper, a comparative analysis of acoustic and clinical speech feature vectors as a method of detection of Parkinson disease is suggested using a multilayer perceptron (MLP) classifier. It evaluates the five sets of features: the traditional Acoustic-Spectral Speech Features (CASSF), statistically pooled Acoustic-Dynamic Speech Features (SP-ADSF), Hybrid Correlation-Nonlinear Acoustic Features (HCNAF) and the Hybrid Acoustic-Dysphonia-Nonlinear Speech Feature set (HADNSF) as well as the state of the art, widely used clinical benchmark of pathological speech analysis with low computational resources the extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS). The experimental results suggest that the MLP classifier can perform optimally when it is trained using SP-ADSF features. The suggested model attains an accuracy of 94.59, a sensitivity of 92.50 and an area under the ROC curve of 0.9799 when using this representation, which reflects strong discriminative abilities of the model to diagnose the Parkinson disease. On the other hand, eGeMAPS and the suggested HADNSF have comparably poor performance, which may be explained by specific features of the sets, the high feature dimensionality, and sensitivity of dysphonia and nonlinear features to processing at the segment level. However, these sets of features are still clinically relevant because their explanations of voice disorders can be read and interpreted and have a pathology-based nature. In general, the findings indicate that performance-based acoustic feature representation and clinically based feature sets are complementary rather competitive. This work highlights the need to have the right balance between discriminative performance and clinical applicability in reliable speech-based systems of Parkinson disease detection.