Study Of Seizure Prediction Based On Machine Learning

Epilepsy is a common chronic neurological disease of brain dysfunction,seizures with recurrent and sudden.Although epilepsy itself is not fatal,it can cause slow development of intelligence,mental disease and accident.It brings heavy burden to patients.A part of the epileptic disease can’t be treated at present.If we can achieve real-time automatic prediction of epileptic seizures,we can make timely protection measures for patients with epilepsy,improve patient’s sense of safety,and make up for the lack of current treatment methods.The main purpose of this paper is to design a prediction algorithm based on scalp electroencephalograph(EEG)for epileptic seizure prediction by studying machine learning in the prediction of epileptic seizures,and provide theoretical support for the realization of real-time automatic warning for epileptic seizures.The process of epileptic seizure is the process of epileptic EEG changing from interictal to ictal.The key to the prediction of epileptic seizures is to recognize the transformation of this process.The EEG state of epilepsy is divided into four stages:preictal,interictal,ictal and postictal.Therefore,the problem of epileptic seizure prediction is summed up as a two classification problem,that is,identifying the state of epileptic interictal and preictal.In order to study which factors have important impacts on the prediction of epileptic seizures,this paper studied the effects of different leads,time windows and different features on the results of the different machine learning models,and analyzed the changes of EEG in epileptic interictal and preictal.In the first place,the time-frequency analysis of EEG was carried out by wavelet time-frequency analysis.The features of wavelet energy,spectral power and multi-scale permutation entropy of EEG signal were extracted by wavelet energy,Autoregressive(AR)model and multi-scale permutation entropy,besides,the change of features under different time windows had also been analyzed.Secondly,the distribution and effectiveness of the features were studied by the scatter matrix and the correlation analysis.Finally,the support vector machine(SVM)and Muti-layer Perceptron(MLP)models were used to predict epileptic seizures.According to the test results,the influence of the lead selection,time window,and characteristics on the results were discussed.In addition,this paper constructed a double Long Short-Term Memory networks(LSTM)layers neural network based on LSTM model.The model parameters were tuned by visualization of the training process,and the model was tested with test data.The model achieved 90% accuracy for many cases.Compared with the performance of SVM,MLP and LSTM models,it was found that LSTM has achieved relatively good results in the prediction of epileptic seizures.In addition,the performance of LSTM model was further verified by receiver operating characteristic curve(ROC)and Kolmogrov-Smirnov(KS)values of individual cases.Finally,a smooth filtering was used to reduce the false alarm rate.In summary,the study in this article deepens the understanding of the mechanism of epileptic seizures.By comparing the model results under different conditions,the factors that affect the prediction of epileptic seizures were obtained.The double LSTM layers model constructed in this paper had achieved relatively good results in epileptic seizure prediction,and provided theoretical support for the clinical application of epileptic seizure early warning system.