Diagnosis of Parkinson's disease from continuous speech using deep convolutional networks without manual selection of features

Abstract

Parkinson's Disease (PD) is a relatively common neurodegenerative disabling disease. It affects central nervous system with profound effect on the motor system. The most common symptoms include slowness, rigidity and tremor during motion. It has been suggested that the vocal cords are among the first one to be affected and thus the speech is affected at very early stage of the disease and continues to deteriorate as the disease progress. In this work, we focus on automating the process of diagnosis from continuous native speech by removing the necessity of a trained personal from the diagnosis process. This is done by using an adaptation of Convolutional Neural Network (CNN) architecture for one-dimensional signal processing (i.e. raw speech signal) on a relatively small training set. This is a continuation to previous works where we showed (i) that this task can be achieved by using manually-extracted features of the speech (such as formants and their ratios) and (ii) by using an automatic process of auditory features extraction, where the features were selected by signal processing specialist.