Prediction Of Epileptic Seizures Based On Convolution Neural Network

Epilepsy is a chronic disease of transient brain dysfunction,the main cause of which is the excessive discharge of brain neurons.Because epileptic seizures can lead to abnormal behaviors such as mental disorders,transient loss of consciousness,and convulsions,there is a serious threat to the safety and health of patients.Therefore,if the pre-ictal state can be identified,the patient can be protected in advance to avoid the patient suffering from sudden seizuresElectroencephalogram(EEG)records the potential changes in brain neuron activity and is one of the main tools used to analyze the characteristics of seizures.The characteristics of epileptic EEG signals in different periods have obvious differences.Most epileptic prediction algorithms identify EEG signals in pre-ictal according to this difference,so as to achieve the purpose of seizure prediction.On the one hand,the existing algorithms define the pre-ictal period as the time interval one hour before the seizure,which is a big error for the prediction of clinical seizures.On the other hand,existing algorithms are often overly dependent on artificially designed feature parameters,and due to the specificity of different individual EEG signals,it is easy to make the learned algorithm not generalized.Therefore,it is particularly important to study an algorithm that can accurately extract features in the early stage of seizure recognitionThe main research work of this paper is divided into the following parts:1.Research on the algorithm of EEG signal classification based on convolutional neural network and support vector machine.First,a method of dividing the epilepsy state that equalizes the pre-ictal is proposed.Then,based on the frequency domain information of the EEG signal,the average amplitude spectrum(MAS)feature on each sub-band is extracted.Finally,the obtained MAS feature map is input into the Convolutional Neural Network(CNN)for feature extraction,and the Support Vector Machine(SVM)is used to classify the epileptic state.Experiments show that the classification algorithm combining CNN and SVM based on MAS feature achieves 86.25%classification accuracy on the CHB-MIT database.2.Research on seizure prediction algorithm based on improved CNN model fusion and vali-dation of clinical EEG data.First,the CNN for the first part is improved from the activation function layer and the normalization layer.Then based on the improved convolutional neural network model,a stacking integration method is designed,and the adaptive weighted feature fusion is used for the probability output of each CNN,and applied to the prediction of actual clinical data.The results show that the recognition rate of the model based on Stacking fusion and adaptive weighted fusion on the CHB-MIT dataset is 88.1%;and the classification accuracy of the model after fusion in the actual clinical data is 67.4%,the mean of F1 is 0.667,the average AUC value is 0.763.