| Epilepsy has become the second largest neurological disease after headaches,and epilepsy patients account for nearly 1% of the global population.The seizures of epilepsy will bring great pain and danger to the patient,and also bring a heavy burden to the patient’s family and society.If the prediction of epileptic seizures is accurate,and the timely administration of drugs to terminate the seizures,it is of great significance to improve the quality of life of patients and reduce the burden on family and society.As an important tool for studying epilepsy,electroencephalogram(EEG)has become a research hotspot in epileptic seizure prediction.However,there is no clinically applicable epileptic seizure prediction system based on EEG,because there are three key problems that have not been resolved:(1)The problem of too many channels for EEG signal acquisition.Too many channels are not conducive to wearing for a long time,and the amount of calculation is too large.(2)The question of whether to choose deep learning algorithm or traditional machine learning algorithm as the classifier in the prediction of epileptic seizures with few channels.(3)The problem of channel selection method in epileptic seizure prediction.In response to the above problems,this thesis is based on CHB-MIT,a long-term epilepsy scalp EEG database.Research has been conducted on key issues such as channel selection,classifiers selection,etc.,in order to realize the lightweight of the epileptic seizure prediction system and develop epileptic seizure prediction equipment suitable for clinical use.The main work and results are as follows:(1)We performed data preprocessing and data segmentation on epileptic EEG signals,and extracted delta(1～3Hz),theta(4～7Hz),alpha(8～13Hz),beta(14～30Hz),gamma1(31～ 55Hz),gamma2(55～110Hz)relative power of six frequency bands as frequency domain characteristics,and eight time-domain features,including kurtosis,skewness,and decorrelation time,are extracted as the feature set for epilepsy prediction.(2)In epileptic seizure prediction using full-channel EEG data,the performance of four classifiers,K-nearest neighbor(KNN),support vector machine(SVM),long short-term memory(LSTM)network,and convolutional neural network(CNN),is compared and analyzed.The result shows that the epileptic seizure prediction result based on SVM is better than the result based on KNN.In the comparison of LSTM structures,the performance of the bidirection long short-term memory network(BiLSTM)as a classifier is better than that of the LSTM.In the comparison of classifiers for epileptic seizure prediction,the results obtained by CNN have the highest sensitivity,the results obtained by SVM have the highest specificity,accuracy and largest variance,and the results obtained by the Bi-LSTM have the smallest variance.The average sensitivity,average specificity,and average accuracy of epileptic seizure prediction results based on the three classifiers are all above 96%.In comparison with other methods,the results obtained by using SVM and CNN are in a leading position,which proves the effectiveness of the epileptic seizure prediction method used in this thesis.(3)A channel selection method based on the original data of the channel-the minimum variance method is proposed.The interictal EEG data of each patient is formed into a matrix,and the channel variance is arranged from small to large,and this sequence is applied to the preictal EEG data matrix,the desired EEG channel can be selected.Then,in the rough selection of EEG channels,it is finally determined that the SVM is more suitable for the prediction of epileptic seizures with fewer EEG channels.Using sensitivity as the standard,the average optimal number of EEG channels for epileptic seizure prediction in 24 patients was determined to be 5.8.The channel selection method proposed in this thesis is compared with the genetic algorithm,backward elimination method,maximum entropy method and other channel selection methods,which proves the effectiveness of the method proposed in this thesis. |
PDF Full Text Request