The Research Of Seizure Type Classification Algorithm Based On Improved Common Spatial Pattern And Two-dimensional Feature Selection

Epilepsy is a disorder of the electrical activity of neurons in the brain,resulting in lasting damage to the nervous system,with long-term,repeated and cumulative characteristics.Seizures are brief involuntary convulsions of a part of the body or the entire body(partial seizure or generalized seizure).Electroencephalogram(EEG)is a graphic record of the brain’s continuous electrical activity with voltage fluctuations via multiple electrodes and is considered the gold standard for assessing EEG background and detecting seizures.However,due to the complexity of EEG signals,clinicians are still unable to judge disease progression quickly and accurately.Correctly identifying seizure types can help clinicians develop a diagnosis and treatment plan,thereby reducing the potential for future seizures and epilepsy complications.This thesis focuses on the analysis and processing of EEG signals of six seizure types,and proposes an integrated model of multi-domain feature extraction and two-dimensional feature selection based on EEG signals.The main research contents are as follows:1.This thesis studied a feature extraction algorithm based on multi-class specific bands common spatial pattern(MSBCSP).Approximation approximation diagonalization(JAD)algorithm is applied with common spatial pattern(CSP)algorithm in multiple classification tasks.The energy of intrinsic mode function(IMF)is extracted by complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN).Next,linear discriminant analysis(LDA)and random forest(RF)were used for feature selection.A penalized ensemble method is developed to deal with unbalanced data and improve the generalization ability of the model.Specifically,a penalty strategy is first introduced to deal with unbalanced data,which punishes the majority classes by giving a lower weight to the majority classes and a higher weight to the minority classes.Then weighted voting strategy was introduced to combine logistic regression(LR)algorithm with Light GBM(LGB)algorithm to classify epileptic seizure types.2.This thesis studied a seizure type classification algorithm based on two-dimensional feature selection.In the study of epileptic seizure classification based on machine learning,the number of features directly affects the performance of the model.In order to reduce the number of features on the premise of ensuring the model performance,a classification algorithm of epileptic seizure type based on two-dimensional feature selection was proposed.Firstly,a brain network of phase-locking value(PLV)was constructed after removing a channel by EEG signals,and the weighting degree,clustering coefficient,harmonic centrality,module degree,proximity centrality and feature vector centrality of the 20 networks were calculated.The importance of each network was calculated based on these six features.Then the contribution of each channel is obtained according to the matrix operation.At the same time,the time domain,frequency domain and spatial characteristics of EEG signals were extracted,and the individual contribution of features was calculated by using RF,and then two-dimensional iterative selection was carried out according to the channel contribution and individual contribution of features.Finally,the feature set after iterative selection is used to train the punishment integration model to realize the classification of epileptic seizure types.