Epilepsy EEG Signal Analysis And Research Based On Graph Convolutional Network

As one of the common diseases in neurology,epilepsy is caused by abnormal discharge of nerve cells in brain.It can cause transient disorders of nervous system function.In traditional diagnosis,doctors can rely on their rich work experience to analyze,discuss and diagnose.However,efficiency of manual visual detection is not guaranteed.This has prompted development of computer-based automatic detection technology for epilepsy EEG signal,and it has gradually become a research hotspot around the world.Because seizures are result of synergy of several brain regions.And each channel used to collect EEG signals respectively record activities of different brain regions.So there is a certain relationship between channels.However,this paper have found that most epilepsy EEG research algorithms focus on studying data itself of each EEG channel.Fewer algorithms pay full attention to hidden spatial information between EEG channels and channels.The mining depth for multi-channel epilepsy EEG data is relatively shallow,and analysis dimension is narrow.To a certain extent,it affects final effect of epilepsy EEG automatic detection or seizure prediction.To solve the above problems,this paper takes multi-channel spatial relationship of epilepsy EEG as a breakthrough.From perspective of graph theory,it proposes a specific mapping relationship,and maps interrelationship between channels to graph data for processing.Then,this paper successively carries out modeling from multi-channel spatial level,multi-channel "space + time" level,and multi-channel "frequency + space + time" level.It proposes a "progressive and fusion" epilepsy EEG analysis model based on GCN(Graph Convolutional Networks,GCN).And it deeply explores contribution of multi-channel hidden spatial relationships to epilepsy EEG automatic detection and seizure prediction.The main innovations and contributions of this paper are as follows:(1)Aiming at correlation between multiple channels of epilepsy EEG,this paper proposes an epilepsy EEG automatic detection model based on graph convolutional network.The model proposes a specific mapping relationship from perspective of graph theory.It maps data of each channel of epilepsy EEG to vertices of graph.And it maps correlation between channels and channels to edges of graph.Then this paper constructs a detection model with graph convolutional network.This model makes full use of advantages of graph convolutional networks in processing unstructured data.It can extracts high-dimensional features of relevant data,and it can also deeply explores correlation between channels to provide richer data support for epilepsy diagnosis.In addition,through test of public data set CHB-MIT(Children’s Hospital Boston--Massachusetts Institute of Technology,CHB-MIT),this paper proves effectiveness of the model in automatic detection of epilepsy EEG.(2)Aiming at problem of insufficient dimensionality in analysis of multi-channel spatial relationship,this paper relies on multi-channel hidden spatial relationship obtained in the first stage,and continues to explore law of related data at temporal level.Then this paper proposes a spatio-temporal seizure prediction model based on graph convolutional network.The model includes two parts: graph encoder and space-time predictor,which successively explores multi-channel spatial and temporal characteristics.This strengthens depth of mining and broadens dimension of analysis.After testing,it shows a good classification performance.(3)Aiming at problem of optimizing the model proposed in the second stage,this paper proposes a multi-dimensional enhanced seizure prediction model based on graph convolutional network from multi-level consideration of "frequency + space + time".In terms of improvement,this paper constructs information reconstruction space.It reconstructs data units from frequency band level.And it can realize feature enhancement,update coded representation of graph,and explore space-time correlation.In terms of optimization,this paper improves core part of the space-time predictor.It can reduces model parameters,and improves operating efficiency.In general,the model includes three parts: information reconstruction space,graph encoder,and space-time predictor.And it can explores laws of epileptic EEG signals in multi-band,multi-channel spatial relationship and temporal relationship.Through test of the standard data set CHB-MIT,the model structure is continuously optimized and network parameters are adjusted.The experiment achieve expected effect.In summary,the three tasks mentioned in this paper mainly focus on the study of hidden space relationship between epileptic EEG channels and channels.It combines several technical advantages.Considering multi-channel "space","space + time","frequency + space + time",it correspondingly proposes epilepsy EEG analysis algorithms based on graph convolutional networks.By analyzing richer multi-channel feature data,it can provides more substantial and effective data support for epilepsy diagnosis.And it can also provides more novel analysis ideas for researchers and doctors in the same field.