Hierarchy Graph Convolution Network For Epileptic Detection On EEG

The application of scalp Electroencephalography(s EEG)signal in the diagnosis of epilepsy has been deeply studied and developed.However,previous studies have rarely focused modeling attention on the physical manifestations of epilepsy,while ignoring the graphical structural information present in EEG signals.When a seizure occurs,the electrodes placed on the scalp near the epileptic focus will have obvious and consistent voltage changes,so how to convolute the adjacent electrodes and extract high-level features is the key work.In the dissertation,a novel approach to epileptic detection based on the Hierarchy Graph Convolution Network(HGCN)structure is proposed.Multiple features of time or frequency domains extracted from the raw EEG signals are taken as the input of HGCN.The HGCN extracts the high-level regional features through the position relationship between the electrodes on the scalp.At the same time,the vertical and horizontal adjacent relations between electrodes are processed separately so as to complete the prediction of various classification tasks.On the other hand,the dissertation proposes a Tree Classification(TC)method to improve the robustness of the model under various classification tasks.In the dissertation,a number of validation experiments have been carried out on the dataset of Boston Children’s hospital(CHB-MIT)and the dataset of Temple University Hospital(TUH).The accuracy of the model in the epilepsy 5-class classification task in CHB-MIT dataset is 5.77% higher than the state-of-the-art.The sensitivity and specificity of the model in the epilepsy 2-class classification task in TUH dataset is 2.43% and 19.70% higher than the state-of-the-art,respectively.