Research On Prediction Method Of Epileptic Seizure Based On EEG Spike Frequency And YOLOv3 Model

Epilepsy is a complex brain disease,which is a clinical manifestation of abnormal discharge of brain neurons.It has the characteristics of repetitiveness and suddenness,which has brought great influence to people’s normal life.If we can predict seizures in advance and take appropriate measures,we can greatly reduce the suffering of patients.In view of this,this article will carry out research on the prediction of seizures,which has practical significance for the prevention and treatment of epilepsy.At present,the commonly used methods for predicting seizures are to analyze the EEG signals of patients with epilepsy to achieve the purpose of prediction.There are two main types of EEG signals,scalp EEG signals and intracranial EEG signals.The scalp EEG signal adopts a non-invasive acquisition method,which will not cause trauma to the brain,and the intracranial EEG signal needs to penetrate electrodes into the skull to perform invasive signal acquisition,which is somewhat traumatic.Therefore,in order to avoid trauma to the brain,scalp EEG signals are mostly used in studies on the prediction of seizures.In this study,scalp EEG signals were selected as the raw data for the prediction of seizures.Epileptic discharges of EEG signals are the pathophysiology basis of seizures,and spike waves are the most characteristic manifestation of epileptic discharges.The International Electroencephalogram and Clinical Neurophysiology Union believes that "spike wave refers to the brain wave that appears on electroencephalography for a short time and is obviously prominent in the background activity,with the shape of spike,the duration of 20-70 ms,the main component is negative phase,and the amplitude is generally more than 100μV".The analysis of the EEG signal data of epilepsy found that the spikes in the pre-seizure period will increase.Based on this,this article analyzes the changes of the spike frequency in the EEG signals of the pre-seizure period.Previous seizure prediction algorithms have problems of high feature extraction complexity,poor generality,and difficulty in real-time prediction.In this regard,this paper proposes a method for predicting seizures based on EEG spike frequency and YOLOv3 convolutional neural network model for spike detection.This method takes advantage of the low complexity epilepsy spike wave feature,and uses two public scalp EEG databases,namely the short-term scalp EEG database of the University of Bonn and the long-term scalp EEG database of CHB-MIT.First,the short-term scalp EEG database was amplified and the epilepsy spike characteristics in the database were labeled for training the YOLOv3 model.Then,the trained model is verified on the long-term scalp EEG database.Through comparison,it is found that compared with SSD and YOLOv3-Tiny training models,the model selected in this paper has faster detection speed and higher prediction sensitivity.Aiming at the problems of high complexity and long training time of the algorithm,this paper has further improved the network structure and training algorithm of the model.The improved algorithm is applied to four different prediction windows of 30 min,60min,90 min,and 120 min to predict the seizures in real time.The experimental results show that the average sensitivity of different prediction windows is 96.52%,the average prediction time is 44.62 min,the average time to detect each EEG is 0.032 s,and the false alarm rate is 0.104/h.It can be seen from the experimental results that the improved algorithm can not only shorten the training time of the model and speed up the detection speed,but also improve the sensitivity of prediction.Finally,the improved algorithm is compared with the phase/amplitude locking method,the attractor state analysis method,and the discrete wavelet transform analysis method.The results show that the improved algorithm has higher sensitivity,longer prediction time,faster detection speed,and lower false alarm rate.The research content in this article provides a new method for real-time prediction of epileptic seizures,and has certain clinical application value.