Epileptic Seizure Prediction Algorithm Based On EEG Signals

Epilepsy,as one of the common neurological diseases,harms more than 50 million patients worldwide.Many studies show that the EEG signals of epileptic patient can be divided into different periods,which provides a possibility for the prediction of seizures.This thesis uses different algorithms to classify the pre-seizure period and inter-seizure period to achieve accurate prediction of seizures,which is of great significance for epileptic patients.The data we used was collected at Boston Children's Hospital(CHB-MIT).We first analyzed the epileptic patient information in the CHB-MIT dataset in detail,and then extracted the EEG signals from pre-seizure period and inter-seizure period of each seizure for all patients.Each segment of EEG signals lasted for one hour.Based on this dataset,the main results of the thesis are as follows:(1)The general model and patient-specific model based on Support Vector Machine(SVM),Extreme Learning Machine(ELM),and Random Forest(RF)were trained using sub-band signals feature,respectively.Each model was evaluated using multiple indicators.It was shown that when the Seizure Occurrence Period(SOP)was 30 minutes and the Seizure Prediction Horizon(SPH)was 10 minutes with the Coefficient of Variation(CV)feature under the single sub-band,the prediction performance of SVM-based patient-specific model and the ELM-based patient-specific model were significantly better than the corresponding general model;The prediction performance of the RF-based general model and the RF-based patient-specific model were more accurate,and the average accuracy rate of the RF-based patient-specific model to predicting seizures can be reached 98%,the average recall rate was 97%,the averaged specificity rate and averaged precision rate were 99%,and the false detection rate was 0/h.(2)The Convolutional Neural Networks(CNN)model consisting of three convolutional blocks was constructed.EEG signals in different periods were first converted into two-dimensional matrix similar to image format as the model inputs.The general model and patient-specific model were trained.False positive rates,AUC,and F2-Score were used to evaluate the model.It was shown that when the SOP was 30 minutes and the SPH was 10 minutes,the prediction performance of the CNN patient-specific model was better than the CNN general model.Ten out of fourteen epilepsy patients can accurately predict seizure using the CNN-based patient-specific model,and the minimum false detection rate was 0/h,the maximum AUC was 1,and the maximum F2-Score was 100%.In this thesis,SVM,ELM,RF and CNN algorithms were used to build the general model and the patient-specific model,and the performance of seizure prediction of models wereevaluated.Our results may provide a theoretical basis for future research the seizure prediction.