Automatic Recognition Of Epileptic EEG Based On Multi-feature Fusion And Ensemble Learning

Epilepsy is a common neurological disease,and abnormal discharge of brain nerves is the main cause of seizures.Epilepsy is extremely harmful,which occurs randomly and may occur multiple times in a day.The seizures often manifest as loss of consciousness,convulsions,and mental disorders.Patients with epilepsy in an unattended situation may suddenly and unexpectedly die due to various factors,including biting the tongue,secretions or vomits blocking the respiratory tract,causing suffocation,uncontrolled body stiffness,paralysis,and the like.The incidence of epilepsy is high,and the age group is widely distributed,including children,adolescents and the elderly,but the incidence of children and adolescents is the highest.Both men and women are at risk for this disease,and men are more likely to have this disease than women.EEG is a very important tool in the study of epilepsy,providing information that other physiological methods cannot provide.Specific waveforms will generate during seizures such as spikes,sharps,and complex waves can be reflected by EEG.Since the acquisition of EEG signals is often performed simultaneously in multiple locations,the EEG generated is a multi-channel EEG.This paper aims to extract suitable features in the time-frequency domain to solve multi-channel data association problems,and uses ensemble learning methods to predict epileptic seizures based on the characteristics of fusion.On the basis of summarizing the automatic recognition methods of epilepsy EEG at home and abroad,this paper firstly describes several methods with integrated learning ideas:Bagging,Boosting and Stacking.Then analyzing the existing EEG feature extraction methods.On this basis,combined with the results of the EEG time domain and frequency domain analysis of the patients,the correlation coefficient matrix and its eigenvalues?in the time and frequency domain are used to solve the multi-channel correlation problem and fuse with the amplitude and phase features in the frequency domain.Then,the random forest model is usedto compare the prediction effects of different frequency segments and different features,and the subsequent frequency bands and features are determined.An automatic recognition model of epilepsy EEG was established by using Bagging-based random forest and Boosting-based gradient Boosting decision tree and XGBoost.Finally,using the Stacking method to integrate the three models,the prediction error caused by individual differences is successfully reduced,and the effect of the model is improved.The results of this paper show that the characteristics of the amplitude,phase and time-frequency domain correlation coefficient matrix and their eigenvalues in the multi-channel EEG frequency domain can effectively improve the prediction effect;the tree model-based method has good for epilepsy detection.The AUC of the three models reached more than 87%,but due to individual differences,the three models performed poorly on some test sets;the AUC of the Stacking integration method reached 92%,which can effectively reduce the error due to the difference between individuals.Therefore,the research in this paper can effectively help identify the epileptic seizures,on the one hand,it can greatly reduce the cost of prediction;on the other hand,it can help doctors reduce the workload,improve the diagnostic efficiency and reduce the possibility of misdiagnosis,so that patients can get timely and effective treatment,reduces the burden on doctors and patients.