DeepVoice: A voiceprint-based mobile health framework for Parkinson's disease identification

Parkinsons disease (PD) identification has at-tracted a lot of attention in recent years. However, there is still no standardized and convenient way to identify PD, because most researchers are only focusing on promoting identification accuracy. With the recent development of mobile health, a feasible mobile application to achieve PD identification is highly demanded with a small amount of information but can provide reliable results. To this end, we propose DeepVoice, a voiceprint-based PD identification application simultaneously integrating deep learning and mobile health. DeepVoice works by collecting a short period of monosyllabic voice through a mobile health App on a smartphone. Specifically, we propose the Joint Time-Frequency Analysis algorithm to enhance the voiceprint feature in spectrogram domain. We also develop a customized convolutional neural network (CNN) to complete the final identification. We evaluate our proposed DeepVoice on input data format, the length of the input voice and neural network architecture in a large PD dataset. Experimental results show DeepVoice could successfully achieve PD identification with an accuracy of 90.45±1.71% with only 10 seconds long audio segment. Our study also reveals that the smartphone-based mobile health application is feasible for PD identification.