Research On EEG Feature Learning And Seizure Prediction System Based On LSTM

There are 50 million persons suffering from epilepsy all over the world,and about 10 mill ion people in China.Epilepsy results from excessive,spontaneous and synchronous discharge of neurous in the celebral cortex and persistent seizures may lead to permanent injury or even death.It seriously affects the patients' normals lives and job for its sudden and recurrent character.Therefore effective prediction of epilipsy seizure can relieve the agony and further improve the quality of patient's lives.Based on the prediction,many kinds of interference methods are applied to control seizures.Such as short-acting antiepileptic drugs and electrical stimulation and so on.Recently automatic prediction of seizures has become a hotspot in the field of epilepsy research.However the prediction technique is limitied by the algorithm of EEG features extracted from the preictal data.This paper proposes the deep learning algorithon for extracting EEG features from the preictal signals automatically which effectively realizes the prediction of epiloptic seizures.This paper presents a prediction system based on the bidirectional LSTM structure for epileptic seizures.Firstly,all the parellel channels of each patient are jointed and filtered,and varieties of linear features described the eeg signal wave are extracted simultaneously.Then,the Bidirectional LSTM network model is introduced and the feature vectors are putted into the model for automatically learning the features of ictal and preictal signals.With the positive and negative EEG data processed by the network model,The preictal and ictal signals are classified from the seizure signals.Finally,the postprocessing technique including a certain rule is employed to optimize the experimental result and improve the accuracy of the prediction system.This paper uses 9 cases of temporal lobe epilepsy for building the bidirectional LSTM network model.SOP,SPH,sensitive,false alarm rate and average prediction time as statistic measures are utilized to assess the system 24 seizures are correctely predictioned in the totalof 29 seizures and the average false alarm rate is 0.17/h,which proves that the algorithm based the bidirectional LSTM acquires better the prediction performance.The average prediction time is 35.47 min for half of all patients' longer prediction time,which provides sufficient time and psychological preparation before seizures for the patients.On the basis of this algorithm,patients with other focus areas will be also utilized to testify the model and improve the network model further.