Research On Removal Of EOG Artifacts In EEG Based On Deep Learning

During Electroencephalogram(EEG)signals acquisition,EEG signals are very easy to be affected by the artifact(such as Electrooculogram(EOG),Electromyogram(MEG),and Electrocardiogram(ECG))from experimental equipment and participants.In many cases,these artifact signals are much more prominent than neurophysiological signals.If they do not be corrected,they will seriously destroy the EEG signals to measure brain function indicators.Among them,the EOG is caused by blinking and eye movement,which is inevitable and the frequency of blink is also particularly large,compared to other artifacts such as EMG.The EEG signal is easily confused by the EOG artifact,and it is overlapped by the EOG artifact in the frequency domain and time domain.At the same time,the presence of EOG artifact will make EEG visual detection or automatic neurophysiological monitoring,data analysis,EEG signal characteristics Extraction and classification will become very difficult.So far,EOG artifacts generally need visual detection,manual or semi-automatic removal,and this process also requires the participation of experts.Furthermore,additional EOG channels need to be used as reference signals to assist in removing EOG artifacts.At the same time,some methods will need to set thresholds for signals to remove EOG artifacts.However,these ways have some disadvantages.Firstly,it will take a lot of time and effort to complete the process of artifact removal with these ways.Secondly,the way of visual detection will test the professional ability of the experimenter,and there may be neural signals will be mistaken for artifact signals.In addition,the additional EOG electrode will make the subject very discomfort.Based on above disadvantages,this paper will use the EOG artifact as a research object,and then apply the deep learning network model to realize the automatic removal of EOG artifact in EEG.This paper first collects EEG and EOG data separately through laboratory EEG equipment and preprocesses the data,then the EEG and EOG signals are linearly mixed according to the signal-to-noise ratio,and ultimately the analog EEG signals which are interfered by EOG artifact are produced.The analog signals can be used for network training and testing and can be used to better evaluate network models.In this paper AutoEncoder(AE)and Long Short Term Memory(LSTM)network are constructed,and Stacked Sparse Autoencoder(SSAE)and Seq2Seq models are constructed by AE and LSTM respectively.The SSAE and Seq2Seq model are trained respectively by the analog EEG signals which are interfered by EOG artifact,and clean EEG signals are used as labels.And we evaluate the prediction results of the two network outputs.This content is presented as follows:1.In terms of the current methods of detecting and removing EEG artifacts,especially EOG artifacts,this paper investigated and analyzed their research background and development status.At the same time,the characteristics and source of EOG signal were studied and analyzed in this paper.2.In this paper,the experimental process of EEG and EOG data acquisition was designed respectively,and the EEG and EOG data of 12 college students were collected.Then,this paper preprocessed the collected data,including filtering,segmentation,etc.,and made simulation data by linear mixing of EEG and EOG data according to SNR,so as to be used for the training of SSAE and Seq2Seq networks.3.This paper studied the current deep learning Networks,including Recurrent Neural Networks(RNN),Stacked Auto-Encoder(SAE)and Convolutional Neural Networks(CNN),and proposed an algorithm and Network model suitable for the removal of EEG artifacts in this paper through comparative study.4.The TensorFlow deep learning framework was applied to design the scheme of constructing the network model based on the actual situation.The SSAE model with AE network as the basic unit and the Seq2Seq network model with LSTM network as the basic unit were constructed,and these two kinds of neural network models were trained and tested.5.The training and testing process of SSAE and LSTM networks were introduced,and the final prediction results were compared and analyzed.The average correlation coefficient between the predicted value of the SSAE network and the clean EEG label signals reached 0.7940,and that of LSTM network reached 0.8861.It can be seen that the two network models could suppress EOG artifacts well,and could reconstruct EEG signals without EOG artifacts well.Compared with SSAE network,LSTM network model had a better predicted result.Because the structure of the LSTM network in this paper was much more complex than that of SSAE.In this paper,SSAE and Seq2Seq networks have basically achieved their functions in the automatic removal of EOG artifacts from EEG,and have obtained relatively ideal prediction results.The proposed deep learning algorithm provides a powerful technical means for the automatic removal of EOG artifacts and other artifacts in EEG.At the same time,it can effectively avoid the shortcomings of the existing methods for detecting and removing EOG artifacts.