Research On Feature Selection And Ensemble Methods For Epileptic Seizure Classification

Epilepsy is a neurological disorder that characterized by the recurrence of seizures.This seizure poses a serious risk of injury,which limits the mobility and independence of a patient.In epileptic seizure detection,the electroencephalogram(EEG)signal is widely used,since seizure diagnosis could be performed by recognizing the EEG signal abnormalities.The epilepsy diagnosis using EEG with computer aided approaches can alleviate problems such as prone to error in matter of visual inspection and time consuming.The computer aided epileptic seizure study and research can be divided into three categories as follows:1)focus on preprocessing methods,2)focus on feature selection or feature transformation and 3)focus on detection and classification methods.Unfortunately,many epileptic seizure detection and classification research mostly focus on preprocessing and the detection and classification methods.There exists still some space to explore the processing of EEG features.Furthermore,in detection and classification methods which utilized multi classifier system,the researchers only using step generation and fusion,while in selection step is not considered.Thus,in this dissertation,two frameworks are proposed for epileptic seizure detection and classification which are dedicated to obtaining best subset features.The key constributions of this dissertation are summarized as follows:1)In chapter 3,adaptive hybrid feature selection within ensemble bagging(AHFSE)is introduced.The goal of AHFSE is to explore the optimal combination of features using an adaptive rank-aggregation method.AHFSE improves the performances of detection and classification by investigating the data from multiple views.Specifically,AHFSE applied the genetic algorithm as an adaptive process in rank-aggregation to obtain the optimal combination of feature selection in every bootstrap.With wavelet transform,AHFSE obtains significant performance improvements when compared with single feature selection methods,ensemble methods and other wavelet transform based methods.The limitation of AHFSE is that it needs longer time in learning or training,due to genetic algorithm,which could be replaced with other faster algorithms,such as particle swarm optimization which is free to use combination and mutation.Although our algorithm needs longer time in learning models,when all learning models are generated,the real time epileptic seizure detection and classification is computationally cheap and fast.2)In chapter 4,multi-objective stacking hybrid within ensemble(MOSHE)is proposed.MOSHE focuses on maximizing the diversities of multiple classifiers in the ensemble and adaptively selecting the optimal combination of classifiers.MOSHE transforms the original space to a new space where there exist significant divergences between features to enhance the performance of classification and prediction.The optimum combination of classifiers is obtained by adoping Non-domainted sorting genetic algorithm approach.Additionally,MOSHE utilizes the disrete waevelet transform and tunable-q transform to achieve remarkable performance improvement,compared with other MCS methods.The limitation of MOSHE which is need longer time in learning process could be solved by applying parallel processing.Also,NSGAII could be replaced with a multi-objective approach which able to automatically balance the crossover and mutation probability.