Parkinson's Disease Diagnosis Based on Multivariate Deep Features of Speech Signal

Parkinson's disease (PD) is known as neurodegenerative disorder causing speech impairment in patients. Therefore, voice recording has been used as useful tool for diagnosis of PD. For the first time in this study, we have tested the effectiveness of deep convolutional neural network (DCNN) in distinguishing between Parkinson's and healthy voices using spectral features from sustained phoneme /a/ (as pronounced in “car”). Various designs of the DCNN architecture were investigated on raw pathological and healthy voices of varying lengths. This study also investigated the effect of parameters such as frame size, number of convolutional layers and feature maps as well as topology of fully connected layers on the accuracy of the classification outcome. The best network achieved accuracy of 75.7% corresponding on 815 ms of data for distinguishing between Parkinson's and healthy samples. This work has demonstrated that online speech recording has the potential for being used to screening people for Parkinson's disease.