Based On Multi-features And Hierarchical-attention-network To Predict Epilepsy

Epilepsy is a common disease caused by abnormal neuronal discharge.Many hidden risks will be triggered when patients suddenly suffers from epilepsy,so it is very significant to predict it.EEG contains a lot of physiological and pathological information,and it is also an important tool for clinicians to diagnose epilepsy,so EEG is widely used in epileptic seizure prediction work.How to extract features from EEG and construct a network suitable for the features we acquired is a common problem faced by researchers.First of all,this paper reviews and analyzes the progress of epilepsy prediction research,and shows the existing research methods and the results they can achieve.We mainly analyzes the methods of EEG signals feature extraction and the classification algorithm in epilepsy prediction up to now.Then the detailed information of the data set used in this paper,the processing method for this data set and the evaluation index of the experimental results used in our experiment are introduced.Thirdly,from the perspective of continuous time relationship and the relationship between various electrodes,this paper designs the parallel convolution kernel to further extract the multi-perspective features of EEG signals,aiming at the primary features of the EEG.The temporal attention and the electrode level attention are introduced to enhance the feature expression effect,and the epileptic seizure prediction model based on multiple convolution kernel and multi-level attention is proposed.Firstly,the EEG signals is decomposed and reconstructed by Discreet Wavelet Transform（DWT）,and the correlation matrices are further obtained on all the sub bands as the primary characteristics of the EEG,so as to reduce the instability that the differences between different patients might bring to the experiment.Considering that the electrode close to the epileptic lesion contains more effective information in the preictal data,the attention machine at the electrode level is introduced to select the more important electrodes,and the temporal attention is applied to highlight the dominant window.Compared with the existing models,our performance is relatively excellent,with the accuracy,sensitivity and specificity respectively reaching 94.72%,93.25% and 96.19%.Finally,based on the comprehensive analysis of the electrode itself and the relationship between this electrode and other electrodes,a method of multi-features and hierarchical attentional epileptic seizure prediction is proposed by introducing a Recurrent Neural Network.Firstly,EEG signals are divided into 6 frequency bands by wavelet decomposition and reconstruction,and two features are calculated in each frequency band,one of which is the correlation matrix describing the relationship between electrodes,the other is a novel splicing feature reflecting the properties of electrodes themselves.Secondly,in order to excavate more useful information,the novel splicing feature integrates the Sample Entropy,Permutation Entropy and Largest Lyapunov index,which describe the complexity of the EEG signal in non-dynamics,and then it is used as one of the input for the network after special normalization.Finally,multiple features of electrodes on six frequency bands are used as inputs to the hierarchical attention network to learn the implied relationship in the timing sequence of EEG signals.Meanwhile,electrode and window attention are introduced to further improve the performance of our network.Compared with other methods,our method achieves better results,with 97.28% accuracy,97.28%sensitivity and 97.47% specificity.