EEG Analysis Of Parkinson’s Disease Using Time-frequency Analysis And Deep Learning

The high incidence and disability rate of Parkinson’s disease(PD)has attracted wide attention in the whole society.Early clinical symptoms of Parkinson’s disease is not obvious,the general onset in the middle and late stage.Therefore,accurate diagnosis of Parkinson’s disease in the early optimal treatment stage is of great clinical significance.As an early biomarker of neurodegenerative disease,REM sleep behavior disorder is important for the early diagnosis,prevention and treatment of Neurodegeneration.For patients with REM sleep behavior disorder,Parkinson’s condition is often more complex and the motor symptoms are more severe,and if such patients can be identified early and given aggressive long term follow up and treatment,so on the one hand can delay the emergence of patients with obvious movement disorder,the greatest degree of assurance patient’s quality of life,on the other hand can reduce the severity of the movement disorder as far as possible,bring improvement to the patient’s prognosis.Some neurological disorders have been shown to respond to changes in EEG activity associated with the onset of the disease.Therefore,this article focuses on the early diagnosis of Parkinson’s patients and the differentiation of REM sleep behavior disorder in Parkinson’s patients,to construct an EEG classification model for Parkinson and patients with REM sleep disorder.In this paper,we propose three kinds of methods to study Parkinson’s disease.By using time-frequency analysis methods and feature extract methods,two classification methods called Tunable Q-factor Wavelet Transform combined with feature extract(TQWT + feature)and Wavelet Packet Transform combined with feature extract(WPT + feature)were obtained.By using original signals as the input of the deep learning method,a classification method called original signals combined with Deep Residual Shrinkage Network(Original data +DRSN)was obtained.By combining time-frequency analysis with deep learning,tunable Qfactor wavelet transform with deep residual shrinkage network(TQWT-DRSN)and the wavelet packet transform with deep residual shrinkage network(WPT-DRSN)were obtained.The dataset used in this paper is from the clinical sleep data set of Shaanxi Provincial People’s Hospital.TQWT-DRSN and WPT-DRSN have the best experimental results in 2-class,3-class and 4-class classification problems,the accuracy was 99.92% and 99.86%,respectively.The accuracy was 97.81% in the WPT-DRSN model with 3-dim input and 92.59%in the four-classification task,higher than 95.20% and 90.46% of TQWT-DRSN.The method proposed in this paper can be used as a means to monitor the progress of Parkinson’s disease and facilitate the early diagnosis,treatment and prognosis of patients with Parkinson’s disease.The idea of constructing the model is also applicable to the analysis and research of other non-stationary signals,which can bring some guiding significance to the classification of other diseases.