Epileptic Seizure Prediction Based On Improved LSTM Network

Epilepsy is a common chronic brain disease in our daily life.With the increase in the number of diseases worldwide,epileptic diseases are becoming more and more serious to humans.Therefore,it is necessary to study the prediction of epileptic seizures.In the epilepsy prediction framework,due to the improper setting of the number of hidden layer units and learning rate in the traditional network model,there will be over fitting and instability,which will lead to poor network prediction effect.In this paper,the sample entropy is selected as the feature extraction method,and the electronic search algorithm(ESA)based on the idea of electron energy transition in physics is improved.Combined with the traditional long-term and short-term memory network(LSTM)as the classification model in epileptic seizure prediction,the epileptic prediction model is constructed.The main contents of this paper are as follows:1)Before the prediction of epileptic seizures,it is necessary to extract the features of epileptic EEG data.The wavelet transform is used to preprocess the EEG signal,and the subband containing the characteristics of spike wave and spike wave is decomposed.The wavelet energy,sample entropy and standard deviation are used to calculate the eigenvalue of the signal.And use support vector machine(SVM),extreme learning machine(ELM),LSTM network and three kinds of feature extraction method combined with experimental analysis,the results show that the sample entropy method is better than the other two methods,so select the sample entropy as the experimental feature extraction method.2)Network super parameter is often an important factor affecting network performance.In view of this drawback,this paper proposes an improved adaptive ESA algorithm combined with the traditional LSTM network model to make up for the poor debugging ability of LSTM network to super parameter,and constructs an E-LSTM network model for epileptic seizure prediction.The improved ESA algorithm is used to optimize the network parameters,including the number of hidden layer units and learning rate,so that the E-LSTM network can reach the best state of the network model.Using E-LSTM network to test the model,the average sensitivity,the average specificity and the average accuracy are 91.11%,89.52% and90.40%,respectively.3)In order to verify the generalization ability of the models,four network models were tested by cross validation method,and the prediction effect of the models was compared.Under different datasets,the sensitivity,specificity and accuracy of E-LSTM model are improved by 3.9%,3.54% and 4.64% respectively compared with LSTM model.Therefore,the E-LSTM network proposed in this paper can be used as a network model for prediction and classification to realize the prediction of epileptic seizures.