Research On EEG Sleep Staging Based On Mode Fusion And Deep Neural Network

EEG（Electroencephalogram）can detect abundant information of brain activity.The combination of EEG and artificial intelligence is one of the hottest research fields.Epilepsy diagnosis,fatigue driving and anesthesia detection based on EEG signals have achieved good results.Sleep is a basic physiological process which plays an important role in the recovery of human body.Some sleep related diseases,such as autism,insomnia,and schizophrenia,can be diagnosed by analyzing sleep stages.Traditional sleep stage method is time-consuming and laborious.The fast sleep staging algorithm based on EEG has become the focus of this paper.EEG is a kind of non-stationary and nonlinear signal,and EEG signal itself is very weak.It needs an accurate and effective algorithm to extract feature information from it for the classification process of classification network.The traditional feature extraction method is time-consuming and does not meet the timeliness requirements of fast staging.Moreover,due to the lack of adequate prior knowledge,it is easy to miss the key features.Deep learning networks such as convolutional neural networks（CNN）have powerful data analysis and mining capabilities.In this paper,deep learning network is introduced into the study of sleep staging.In order to make use of the complementary information between different signals,we also add the method of modal fusion in this study.And the long short term memory（LSTM）neural network is selected as the classification network.In this paper,based on EEG and deep learning network,three different sleep staging models are proposed to make analysis on the the SLEEP_EDF₂013 dataset.The first method decomposes EEG signal by using complementary empirical mode decomposition（CEEMD）,and then uses simplified residual network（ResNet）to extract features,and send it to classification network for classification.In the second method,the inverted frequency feature is extracted by triangle filtering and other feature is extracted by parallel CNN（PCNN）respectively,then two feature vectors are combined to classify.In the third method,in order to improve the accuracy of non-rapid eye movement phase 1（N1）classification,Electrooculogram（EOG）is introduced to fuse with EEG signals.The main works and innovations of this paper are as follows:（1）A sleep staging framework based on the combination of CEEMD and simplified residual network（ResNet）is proposed.First,the original EEG signal is decomposed by CEEMD to get IMF components.Then different numbers of IMF are combined with the original EEG signals and sent them to simplified ResNet.Finally,fuse the characteristic features extracted by simplified ResNet and input them into LSTM network with Focal loss attached to classify.Compared with the traditional feature extraction method and only using the EEG signal for classification,the classification result of this method has been significantly improved.（2）Two sleep staging algorithms based on feature fusion are proposed.The first feature fusion method is based on the combination of triangle filtering and parallel convolutional network（PCNN）.It combines the traditional manual feature extraction method with deep learning network to obtain more distinguishing features.The second method is mainly based on modal fusion.EEG and EOG physiological signals are extracted and characterized by PCNN respectively for feature fusion,and then sent the fused features into a bidirectional recurrent network for classification.The accurary of the algorithm is improved by taking advantage of the information complementarity between different modes.The experimental results show that the accuracy of N1 period identification have been significantly improved by the two methods we proposed...