Research On Feature Extraction And Classification Of Epilepsy EEG Signals

Epilepsy is one of the brain diseases caused by hypersynchronization of neurons in the brain.About 50 million people with epilepsy are suffering.For the life and health of epilepsy patients,the detection of epilepsy is very important.Clinically,EEG is generally used as the basis for the diagnosis of epilepsy,but it is a time-consuming and laborious process to determine the epilepsy focus area and the period of each seizure by analyzing the EEG morphology with the naked eye,which may lead to untimely,misdiagnosed and missed diagnosis.It is of great significance to use computer-aided automatic detection algorithms for epilepsy according to the development of information technology.Therefore,in this paper,two automatic detection methods for epilepsy EEG signals were constructed respectively.Based on traditional machine learning methods,this paper proposes an epilepsy detection method based on Step-by-Step Accurate feature selection and Random Forest.The original EEG signal is extracted from the time domain,frequency domain and nonlinear analysis to form an initial feature subset,secondly,the extracted EEG feature set is first selected by the Maximum Information Coefficient method,and then use the Recursive Feature Elimination strategy to further screen the features related to the target task,and then determine the final feature subset through the Adaptive Lasso feature selection method based on the penalty term,finally selected feature subset is used as the input of the random forest algorithm,so as to obtain higher classification accuracy.Based on the deep learning method,a CNN-LSTM combined model is constructed in this paper.The two steps of feature extraction and classification of EEG signals are integrated into an overall model.Two combined models of parallel and serial models are established to realize the automatic detection of one-dimensional epilepsy EEG signals.Determine the best model structure by comparing the classification performance of the two combined models.The combined model can perform experiments and achieve good classification performance without manual design and complex feature extraction.The two proposed classification methods are all verified and compared on the datasets of Bonn University and CHB-MIT.It can be seen that the two automatic detection methods have obtained great detection results and practicability.