Convolutional Neural Network Method For Parkinson's Disease Based On Dysphonia

Parkinson's disease(PD)is a common neurological degeneration disease with a long time course and a significant prevalence.On the current level of medical care,failed to clear the cause of the disease,so can not completely cure Parkinson's disease,only in the early control of the disease development.In the early detection of disease symptoms,if the patients could be treated appropriately that can delay the progression of the disease.The dysphonia is one of the early symptoms of Parkinson's disease.In recent years,the study of Parkinson's disease diagnosis based on dysphonia is one of active fields in diagnosis of Parkinson's disease.In this paper,we propose a deep learning method for Parkinson's disease based on speech features by convolutional neural network.The study of phonological disorders in Parkinson's disease with the combination of depth learning and medical field that promoted the combination of artificial intelligence and speech recognition and played an important role in the diagnosis of Parkinson's disease based on dysphonia.At first,we have improved in the data representation and feature extraction.We proposed that speech is the representation of time-frequency to break through the limitations of feature extraction in the time domain and frequency domain,and then improve the representation.In the speech feature extraction stage,the convolution neural network performs the operation of updating weights,so that the network has the self-learning characteristic and completes the visualization extraction of the features in the speech according to the progressive of the network layer.In subsequential,constructing a multi-layer convolution neural network and expounding the network structure and the fine-tuning stage of the network in detail.In the construction of the convolutional neural network,a convolutional neural network of eight network layers is designed,according to the progressive relationship between the network layer structures and the principle of data passing in the network layers.Through the comparison experiment of the data set and analyzing the experimental results,fine-tuned the configuration parameters of the constructed network.At last,in the experiment of diagnosis based on dysphonia in Parkinson's disease,with the data set of UCI Parkinson's speech and data set of self-recording which we used to train the network and diagnose Parkinson's disease.Experiments show that the theory of convolutional neural network in deep learning not only realized the visualization of features in speech,but also contributes to the extraction of new features,which avoids the bottleneck that features cannot be learned.