Analysis And Study Of Epileptic EEG Based On Spatio-temporal Feature Learning

Epilepsy is a common neurological disorder characterized by sudden abnormal brain electrical activity.Seizures can lead to mental disorders,motor disorders and even death,so early detection and treatment of epilepsy is crucial.Electroencephalogram(EEG)reflects the electrical signals of neuronal activity in the brain and is important in the diagnosis and treatment of epilepsy.In recent years,with the development of machine learning and deep learning techniques,an increasing number of research efforts have been made for automatic seizure detection.Previous neurology has shown that there are mechanisms of interaction between various brain regions.In contrast,most exiting deep learning algorithms only consider the temporal characteristics of the individual channels of epileptic EEG data itself,while uderestimating the spatial relationships between EEG channels.This makes the extraction of spatio-temporal information from multiple channels of epileptic EEG inadequate and the dimensionality of the analysis narrow,which affects the effectiveness of the final seizure detection.Therefore,fully exploring the temporal and spatial relationships between different EEG channels is the first focus of this thesis.Most existing works focus on single-patient scenarios,with poor performance in cross-patient seizure detection,which is more challenging and relevant.At the same time,in research and practice related to clinical medicine,the duration of normal EEG signals is much longer than seizures,leading to sample imbalance problem.Sample imbalance can lead to a bias of the classifier towards normal EEG,reducing the ability to detect seizures.In contrast,most existing studies use balanced samples to train classifiers.The clinical applicability of these methods need to be improved.In addition,most existing models have high computational complexity,long training time and low detection efficiency.Therefore,improving the performance of automatic seizure detection across patients,reducing the computational complexity and solving the sample imbalance problem are the second key research area of this thesis.In summary,the main contributions of this thesis is described as follows:(1)We proposes an automatic seizure detection method based on Graph Attention Networks(GAT)and Bi-directional Long Short-term Memory(Bi-LSTM)networks.The method exploits the spatio-temporal correlation between the channels of epileptic EEG signals.Specifically,the GAT is used as a front-end for extracting spatial features,exploiting the topology of the different EEG channels.The topology is a graph structure,where the nodes represent the different EEG channels and the edges represent the spatial relationships between the channels.Meanwhile,the Bi-LSTM network is used as the back-end to mine the temporal relationships and make the final decisions based on the state before and after the current moment.Experiments are conducted on CHB-MIT and TUH datasets with ten-fold cross-validation.Extensive experimental results show that the proposed model can effectively detect seizures from the raw EEG signal without additional feature extraction.The seizure detection accuracy is 98.52% and 98.02% for the two datasets,respectively.The performance of the model is better than or comparable to state-of-theart models.(2)To improve the experimental performance of seizure detection across patients,while reducing the computational complexity and addressing the sample imbalance problem,we propose Hybrid Attention Network(HAN)for automatic seizure detection based on the first work.The graph attention network extracts spatial features at the front end and transformer acquires temporal features as the back end.the HAN utilises the attention mechanism to fully extract the spatio-temporal correlation of the EEG signal.The focal loss function is introduced into HAN to deal with the dataset imbalance problem of EEG-based seizure detection.Both single-patient and cross-patient experiments are conducted on the public CHB-MIT dataset.The experimental results demonstrate the effectiveness of the HAN in both experimental settings.The mean detection sensitivity and AUC for the 23 subjects are 97.93% and 98.56%,respectively.In summary,we propose an automatic seizure detection method based on GAT with BiLSTM and an automatic seizure detection method based on HAN.Based on the two methods,single-patient and cross-patient experiments are conducted on the publicly available CHB-MIT and TUH EEG datasets.The experimental results validate the feasibility of spatio-temporal feature learning in the field of automatic seizure detection,providing a scientific reference value for the research of automatic seizure detection algorithms and practical applications in clinical medicine.