Automatic Seizure Detection Researches Based On Transformer

Epilepsy is a fatal condition that can cause momentary brain malfunction.Although its pathogenesis is unknown,it could seriously damage a person’s body.Automatic seizure detection technology has grown significantly as a research priority in the area of epilepsy diagnosis with the development of technology.The automatic detection algorithm for epilepsy proposed in this paper can not only replace traditional manual detection methods,but also save a lot of time and energy,thereby helping doctors to diagnose epilepsy.In response to the limitations of convolutional neural networks(CNN)in global feature extraction and the insufficient utilization of spatial and temporal features in each channel of epileptic EEG signals,this paper conducted research work and proposed two seizure detection algorithms based on Transformer.Which can achieve automatic analysis of EEG.The research can solve the problems existed in the current CNN technology in epileptic EEG detection.It can explore the spatiotemporal relationship between EEG channels in space and time.The research is of great help to the development of seizure detection algorithms.At the same time,the research results of this article promote the clinical application of deep learning models in seizure detection.In order to solve the problems existed in automatic seizure detection technology,this paper proposes corresponding solutions to assist doctors in diagnosing and protecting patients ’ life and health.The main research contents are as follows :(1)Automatic seizure detection based on interactive fusion of local and global features(CNN and Transformer feature fusion).CNN algorithm has excellent local feature extraction capabilities but weak global feature extraction capabilities.Aiming at the problem of insufficient feature extraction in seizure detection using existing CNN techniques,an interactive method is proposed using Transformer’s good global feature extraction ability.This algorithm can solve the shortcomings of CNN in feature extraction.Firstly,CNN and Transformer are used to extract local and global features of EEG.Then a feature interaction layer is used to interactively fuse the two types of information.The enhanced feature representation is input into the classifier for seizure and non seizure detection.Experiments are conducted on the CHB-MIT dataset.The average sensitivity,AUC,F1 score,specificity,accuracy is 97.70%,98.14%,97.90%,97.60%and 98.76%,respectively.(2)An automatic seizure detection based on spatio-temporal transformer.The interactive fusion method for local and global features has high time complexity and long training time.By analyzing the the characteristics of EEG signal,the spatial relationship and temporal relationship of different EEG channels,we propose an algorithm for automatic seizure detection based on spatio-temporal Transformer.This method first extracts the spatial features between channels,and then extracts the temporal features on each channel.The spatial and temporal features are combined and put into the classifier.Patient-specific seizure detection experiments are carried out on the public dataset CHB-MIT.The average sensitivity,AUC,average F1 score,specificity,accuracy is 97.46%,97.98%,97.48%,97.52% and 98.63%,respectively.In summary,in order to sovle the problems found in the research of automatic seizure detection technology,the paper proposes two solutions based on the popular Transformer.The experimentals results provide assistance in the auxiliary diagnosis of epilepsy and provide a new idea for the development of automatic seizure detection.