Patient-Independent Automatic Seizure Detection Research Based On Deep Learning

Seizure is a common neurological disorder whose pathogenesis is a transient disorder of the nervous system caused by abnormal discharges of nerve cells,which is extremely harmful and may lead to fatal damage.In traditional seizure diagnosis,doctors rely on their expertise to diagnose seizure by analyzing Electroencephalogram(EEG),which is time-consuming and laborious.Therefore,automatic seizure detection technology has become a hot research topic.Deep learning-based seizure detection algorithms are used to implement automated analysis of EEG.During the research process,it is found that there is a spatially synergistic response of electrical signals released from various brain regions.Besides,different brain regions interact with each other.Exploiting this spatial topology between brain regions is of great help to seizure detection algorithms.However,the spatial topology has often been overlooked in past studies.Meanwhile,existing automatic seizure detection technologies have significant performance degradation when applied to unknown patients.The emergence of these problems has affected the clinical application of existing automated seizure detection technologies.This study addresses existing problems of automatic seizure detection technologies and proposes solutions to assist doctors in diagnosing and protecting patients’ lives and health.The main research contents are as follows.(1)To deal with the problem of insufficient utilization of spatial relationships between different channels of EEG signals,an automatic seizure detection model based on Graph Attention Networks(GAT)and focal loss is proposed.The model converts the channels of the original EEG signal into graph structures according to their correlations.And the graph structures together with the EEG signal are used as training data.The feature maps of the model are updated by a feedforward neural network following a self-attention mechanism.Meanwhile,in order to deal with the classification decision bias caused by the positive and negative sample imbalance problem of EEG data,a focal loss function is employed to balance positive and negativesamples.The performance of the model was tested on the public CHB-MIT dataset.The average accuracy,sensitivity,specificity,F1-score and AUC are 98.89%,97.10%,99.63%,98.33% and 99.06%,respectively.This work broadens the research on the application of graph structure in seizure detection and strategies to cope withimbalanced data.(2)In order to tackle the performance decay problem in the application of automatic seizure detection techniques for unknown patients,this study further investigates patientindependent seizure detection based on the first research and proposes a multi-view patientindependent seizure detection model based on information bottleneck attribution.Multi-view features are first extracted in the frequency domain as well as the time-frequency domain.An adversarial learning framework is proposed for discriminative representation learning from the extracted features and the original EEG data.Generic seizure features are obtained through generators,discriminators and the reconstructed signal constraints,which in turn enable seizure detection on unknown patients.Also,information bottleneck attribution is introduced to enhance the interpretability of the model and improve the confidence of the prediction.The performance of the model was tested on the public CHB-MIT dataset as well as the TUH EEG dataset.The average accuracy,sensitivity,specificity,F1-score and AUC are 76.36%,77.42%,76.32%,76.12%and 82.07%,respectively.This task broadens the study of the application of seizure detection technologies to unknown patients and strategies to cope with model interpretability.In summary,the above two works of this thesis focus on automatic seizure detection technology and propose effective solutions tothe problems identified during the study.This study contributes to the auxiliary diagnosis during seizure treatment and provides new ideas for the development of automatic seizure detection technology.