Research On Seizure Detection Method Based On Semi-supervised Generative Adversarial Network

Epilepsy is a chronic neurological disorder caused by abnormal discharges of neurons in the brain.Seizures can cause foaming at the mouth,twitching of limbs,and loss of consciousness,which can result in significant physical harm to patients and even be life-threatening.In clinical practice,electroencephalography(EEG)is commonly used for the epilepsy diagnosis,and it requires lengthy monitoring periods,which can be timeconsuming and increase the workload of physicians.Therefore,it is of great clinical significance to investigate automatic seizure detection methods.To address the issue of limited seizure labeling data,we conduct research on seizure detection methods based on semisupervised generative adversarial network in this thesis,and the main research works are as follows.(1)To prevent the discriminator from learning irrelevant areas that could affect the final classification performance,a semi-supervised generative adversarial network is combined with a hybrid attention mechanism for seizure detection.After preprocessing the EEG signals,the semi-supervised generative adversarial network model was constructed,with seizure detection accomplished via post-processing techniques such as mean filtering and threshold comparison.The channel attention module and spatial attention module were added to the discriminator model so that the discriminator learned important features of the seizure EEG signal to improve the detection performance.The automatic seizure detection method achieved an average sensitivity of 92.84% and an average specificity of 94.60%over approximately 975 hours of test data in the CHB-MIT EEG dataset.(2)To build a lightweight seizure data generation network and to further improve the performance of seizure detection,a seizure detection method based on semi-supervised generative adversarial network and knowledge distillation is proposed.The generators of the teacher network and the student network were designed separately using the U-Net structure.Then we calculated the similarity loss and output layer loss to measure the difference in distribution between the two generator outputs and the difference between the predicted probabilities of the two sets of outputs on the discriminator.Knowledge distillation was used to transfer knowledge from the teacher network,which has a complex structure,to the simple structured student network,thus reducing the number of parameters and storage space of the model.Performance tests in the CHB-MIT dataset reached an average sensitivity of 93.30% and an average specificity of 94.11% when the parameter size of the student network was only 1/8 of that of the teacher network,validating the effectiveness of knowledge distillation in seizure detection.