Research On Epileptic Seizure Prediction Based On EEG Spatiotemporal Characteristics

Epilepsy is caused by abnormal discharge of the brain neurons.Accurate prediction of seizures can improve the patients’ health,as well as the quality of their life.Because Electroencephalogram(EEG)can reflect the physiological activities of neural clusters in the brain,it is usually used for seizure prediction.Studies in this field involve multiple disciplines.How to extract useful information from EEG and build models for its pattern recognition are hotspots and difficulties in this field.Firstly,the history and the recent advance of the related field are investigated in this paper,based on which doing our own research.Besides,some introductions are also made for common analysis methods of EEG in epilepsy,as well as the neural network technologies.Secondly,a seizure prediction method is proposed based on an artificial neural network with an adaptive attention mechanism,in which multi-channel EEG is analyzed from the spatial perspective.The method uses a short-time Fourier transformation to transform EEG from the time domain to the time-frequency domain.Compared with the raw data,the time-frequency spectrum of EEG in different periods can be easier distinguished.Then,because there is not a clear definition of epilepsy periods in EEG,an unsupervised learning approach is used to learn its intrinsic features and limit the impact of data labels defined artificially.After that,a supervised learning approach is used to furtherly pre-train the partial weights and fine-tune the whole weights of the model respectively.This process adjusts the network weights from perspectives of both the EEG mode and the EEG labels,which makes the model more robust.Because epilepsy lesions are often located in specific brain regions,EEG collected from electrodes near the lesions contains more epileptic waves.To highlight the dominant electrodes for seizure prediction,the method uses an electrodes attention mechanism to assign weights for electrodes adaptively.Compared with some existing studies,our method achieves a good result.Finally,another seizure prediction method is proposed based on an artificial neural network with multi-timescale,in which EEG is mainly analyzed from the temporal perspective.Wavelet packet decomposition is used to decompose EEG into multifrequency bands,by which more available information could be dug out.After that,the correlation coefficients are calculated among the electrodes in each band for representing the brain synchrony.Then the correlation coefficients are used as the input of a convolutional network to learn the high level features.And the output of the convolutional network is used as the input of a recurrent network to furtherly learn the implicit sequential relationship.Through the convolutional and recurrent networks,spatiotemporal characteristics are considered jointly.Because EEG is a kind of nonlinear and unstable signal,a multi-timescale operation is involved to reduce the impact of its instability.Compared with some other studies,the proposed method has also achieved a good result.