Research On Seizure Detection Methods Based On Domain Adaptation

Epilepsy is a neurological disorder characterized by abnormal electrical discharges in the brain neurons,and its seizures are characterized by repeatability and uncertainty,which seriously harm the physical health of patients.In recent years,research on the diagnosis and treatment techniques of epilepsy has received extensive attention and high importance in various fields.The study of automatic detection method for epileptic seizures based on electroencephalogram(EEG)analysis has important clinical significance and value.In order to solve the problem of data distribution differences between datasets,this paper conducts research on methods for seizure detection based on domain adaptation transfer learning theory,and the main research efforts are as follows.(1)A seizure detection method based on manifold embedded distribution alignment is proposed,that is,on the basis of preprocessing and extracting features of multi-lead brainwave signals,manifold subspace learning and dynamic distribution alignment methods are used to narrow the distribution differences between source domain and target domain data,and learn domain invariant classifiers by minimizing structural risks in the source domain to realize seizure detection of target domain data.This method can effectively reduce the difference in data distribution between different patients,and then improve the detection accuracy of the model.Algorithm verification on about 866 hours of EEG data in the CHB-MIT dataset yielded 97.03%.Average sensitivity and 95.78% average accuracy.(2)In order to make full use of the superior feature extraction ability of deep networks to further improve the performance of seizure detection models,this paper combines Res Net-50 network with convolutional block attention and dynamic distribution alignment to establish a seizure detection method based on dynamic distribution depth domain adaptation.First,the original EEG signal is passed through a band-pass filter.Denoising,using short-term Fourier transforms to convert EEG into two-dimensional spectrogram data.Second,the spectrogram is fed into the Rest Net-50 network embedded with the convolutional block attention module to extract depth features while performing dynamic distribution alignment.The method achieves an average sensitivity of 96.77% and an average accuracy of 96.48% in the CHB-MIT dataset.The two epileptic seizure detection methods established in this paper based on domain adaptation theory have good detection performance and practical application prospects,providing new ideas for the development of epileptic seizure detection technology.