Automatic Seizure Detection Based On Stransform And Deep Belief Network

Epilepsy is a chronic neurological syndrome characterized by repeated attack,which can be caused by inheritance,various kinds of neurological disorder and systemic diseases.Epilepsy may occur in all age brackets and its clinical features mainly include the sudden loss of consciousness,overall rigidity,convulsion,somatic or limb shakes like electric shock.In the microcosmic level,epileptic seizures result from the abnormal discharge of the brain nerve cells,which is clinically recorded by EEG,and EEG helps in the treatment and research of epilepsy by doctors.On the one hand,EEG contains much physiological and pathological information thus playing an important part in the analysis and treatment of epilepsy,on the other hand,EEG generally contains dozens of hours of data which need medical staff to read based on their clinical experience,which means analyzing EEG is a time-consuming and tedious work,moreover,the accuracy of seizure detection may be cut down because of medical staffs' subjective thoughts and tiredness.In order to improve the efficiency and accuracy of seizure treatment and diagnosis,the development of information technology brings us automatic seizure detection.It can read and analysis EEG through computers and detection algorithms automatically,which greatly reduces the work burden in manual EEG reading,thus increasing the effectiveness of seizure treatment.In this paper,we propose a novel automatic seizure detection algorithm named Deep Belief Network based on deep learning.DBN which is made up of many Restricted Boltzmann Machines is a classical algorithm in deep learning.At present,DBN has been widely used in many fields,such as object-oriented modeling,feature extraction,recognition and so on.The algorithm framework proposed in this paper is as follows:first we divide the original EEG signal into segments and then the epochs are analyzed in the time-frequency domain through S-transform and a batch of features can be extracted from each epoch.Then the features are processed through linear normalization in both the training and testing stage.After that,the normalized features are put into the self-defined deep belief network,and after extracting and abstracting layer by layer,the network output two kinds of labels,named seizure and unseizure.At last we apply post-processing to the minus between the two outputs to promote the detection accuracy.The database used in this experiment comes from the Epilepsy Center of the University Hospital of Freiburg,and we evaluate the proposed method through 71 seizure events from 18 patients.The proposed method applies the classical deep learning algorithm into seizure detection,thus boosting the development of seizure detection technology.