Research On Detection And Diagnosis Of RBD EEG Based On Time-frequency Deep Network

Parkinson’s disease(PD)is a common neurodegenerative disorder characterized by a large number of motor and non-motor features that can impact on function to a variable degree.Rapid-eye-movement sleep behavior disorder(RBD)is characterized by loss of muscular atonia and prominent motor behaviors,which is considered as subtype of PD.Parkinson’s patients with RBD may lead to cognitive behavior disorder and more serious symptoms.Clinically,RBD is considered to be a subtype of PD.Therefore,diagnosis of RBD in PD is of great significance to patient-specific treatment for PD patients with or without RBD.Currently,polysomnography or RBD screening questionnaire(RBDSQ)is used to diagnose RBD in PD in clinical practice,but polysomnography is collectioncomplicated and labor-intensive and RBDSQ has high rate of missed diagnosis.In this work,an intelligent diagnosis method based on few-channel scalp electroencephalogram(EEG)and time-frequency deep network is proposed for PD patients with RBD.Firstly,the sleep EEG data of the subjects were screened.According to the sleep monitoring report,360 minute EEG of each subject during sleep period was extracted.And the 6-channel scalp EEG data,including Fp1,Fp2,C3,C4,O1 and O2,was segmented in equal length by frequency-domain filtering and time-domain segmentation.After EEG data preprocessing,there are 100 equal time-window data for each subject.Secondly,three time-frequency analysis schemes,including short-time Fourier transform(STFT),continuous wavelet transform(CWT)and Hilbert Huang transform(HHT),are proposed to extract the time-frequency characteristics of EEG time-window data,which is time-frequency spectrograms.Then,time-frequency depth network model is constructed based on EEG spectrogram and Convolution Neural Network(CNN).Batch normalization and L2 regularization are used to optimize the model to increase the inhibition ability of over-fitting,and CBAM is used to improve the model to increase the representation and learning ability of EEG spectrogram.Finally,the deep network model was trained using the spectrogram of EEG time-window data for PD with and without RBD by 3-fold cross validation.segmentation-based classification result of time-frequency deep network is postprocessed to obtain the subject-based classification result.30 idiopathic PD patients and 30 PD patients with RBD are selected and the doctor’s detection results of polysomnography are taken as the gold standard in our study.The accuracy of the classification with segmentation-based sample are91.04% in the validation set.Compared with the six classification algorithms such as CNN,Multi layer perceptron(MLP),Support Vector Machine(SVM),K Nearest Neighbor(KNN),Random Forest(RF)and Extreme Gradient Boosting(XGB),ours time-frequency depth network presents a better classification performance.The accuracy of the classification with the subject-based samples is93.33% in the test set.Compared with the RBDSQ,the novel approach has clinical application value.