Automatic Seizure Detection Based On S-transform And Bidirectional LSTM

Epilepsy is a neurological disease which is caused by repetitive and sudden brain neurons discharge.According to estimation of World Health Organization,more than 50 million people worldwide suffer from epilepsy.It may cause clinical manifestations such as dysfunction,loss of consciousness,and fainting,which seriously endangers people's physical and mental safety.Besides,it adds burden to the family and society.Electroencephalography(EEG)plays an important role in the diagnosis and treatment of epilepsy.EEG records are visually analyzed by experienced doctors to determine the focus area and the beginning and end time of each seizure event.Then they can accordingly determine the treatment plan.Due to the large amount of EEG data,which are usually about 24 hours,manual analysis is time-consuming and labor-intensive with misjudgments and omissions.With the development of information technology,more and more repetitive work can be completed with computer assistance.Therefore,computer-aided technology is employed to achieve automatic detection of epilepsy,which can reduce the burden on staff and improve the accuracy of detection significantly.Automatic seizure detection has been developed for several decades.The earliest seizure detection method utilized "half-wave" to extract EEG features.Nowadays,seizure detection method has gradually evolved into a variety of algorithms such as time-frequency analysis,sparse representation,machine learning and deep neural networks.The basic idea of epilepsy detection method is to extract features of epilepsy EEG and compare them with normal EEG.A general model is the combination of a feature extractor and a classifier.The feature extractor extracts EEG signal features with methods such as time-domain analysis,frequency-domain analysis,wavelet transform,and nonlinear dynamic analysis.Classifiers mainly include support vector machines,random forests,nearest neighbor classification,and artificial neural networks.In this thesis,we propose an efficient seizure detection algorithm which combines S-transform and Bidirectional LSTM.S-transform is a high-performance time-frequency analysis method with multi-resolution analysis and low complexity.It has been widely used in signal processing fields in recent years.The Bidirectional LSTM neural network is a variant of the recurrent neural network.It has excellent time series analysis capabilities and has solved many time series classification such as heart sound detection and stock prediction.S-transform is firstly used to extract the time-frequency features of the segmented EEG signals,and the obtained features are then sent into BiLSTM network.Postprocessing is utilized to achieve higher detection accuracy.To evaluate our method,Freiburg Epilepsy Long-term Database and CHB-MIT are used for testing.The segment-based sensitivity yields 98.09%and 96.55%,respectively.In addition,we have collected several patients from the SDU Second Hospital to build our own database and obtain a sensitivity of 93.31%.The experimental results show that the proposed method can be applied to clinical practice effectively.