Research On Epilepsy Prediction Method Based On Epileptiform Wave Reconstruction And Multiple Feature Selection

Epilepsy is a kind of brain dysfunction disease.Because of its sudden onset,it is difficult to predict and has long been considered unpredictable.Since 2013,scientists successfully predicted epilepsy in a patient before they proposed that epilepsy can be predicted.Since then,epilepsy prediction has become one of the hot issues of neuroscience and clinical research.However,research on epilepsy prediction is slower than that on epilepsy detection.This paper focuses on the problem of epilepsy prediction,and proposes a method based on epileptic wave reconstruction according to Electroencephalogram(EEG)analysis.The general time series prediction problem is to predict the signal size at a certain time in the future according to the current and historical signals.This prediction is not enough for epilepsy prediction,and the latter needs to get EEG for a long time to make a judgment.In addition,epilepsy is a sudden event and EEG is a non-stationary signal,which is difficult to predict by the general time series method.In this paper,the problem is transformed into predicting whether the epilepsy will be seizure within a certain period of time.By dividing EEG into interval,pre-seizure,seizure and post-seizure,the problem of epilepsy prediction is further transformed into the classification of whether the patient is in the interval of epilepsy or pre-seizure.In the clinical practice of epilepsy diagnosis,it is found that epileptiform waves will appear during seizures.Based on this,the author assumes that the component signal of epileptiform wave may have changed in some way before seizures.Based on this assumption,this paper proposes a prediction method based on epileptiform wave reconstruction.Firstly,EEG is decomposed by wavelet transform,and wavelet inverse transform is used to reconstruct three kinds of epileptiform waves,that is,spike wave,sharp wave and slow wave.Secondly,In order to select the appropriate feature combination,this paper comprehensively investigates various typical single features and feature combinations of EEG,intend to find a metric that can measure the correlation between features and class labels,and introduces the concept of information gain into the measurement of continuous features,so as to select the sample entropy,standard deviation and other features of full band signal and three kinds of epileptiform signals as the feature combination of the classifier.Finally,the back propagation neural network classifier is optimized for prediction.In this paper,the CHB-MIT scalp EEG Dataset is used,and the prediction method proposed is combined with 5-fold cross validation to carry out epilepsy prediction experiment.The results show that the performance of the proposed method is as follows,that is,the specificity,sensitivity and accuracy is 95.61%,95.27%and 95.41%,respectively.The results are compared with those of several representative literatures.The results confirm the assumption that the components of epileptiform waves have begun to change in the pre-seizure period,and the selection of multiple feature combination is also appropriate.Furthermore,a similar idea based on epileptiform wave reconstruction is used to test epilepsy detection on the dataset provided by Department of Epileptology at Bonn University,and good results are obtained.