The Research On Preprocessing And Sleep Staging Based On Multi-channel EEG Signal

Sleep plays an irreplaceable role in the maintenance of the body's immune system and ease the fatigue of the brain,is an important physiological activity of human.With the accelerated process of urbanization,incidence of sleep disorders in the urban population increased year by year,which has become a serious problem in the world.It is the premise for accurate sleep staging to evaluate sleep quality and cure sleep disorders.It is very important for clinical significance.This is the objective and effective method for sleep staging based on analyzing EEG signals,it has being a topic in the field of sleep medicine of the world.The purpose of this paper is to improve an algorithm of preprocessing and sleep staging based on multi-channel EEG signals,the realizing automatic sleep staging algorithm is designed in three aspects of signal preprocessing,feature optimal selection and feature parameter classification.The main innovation and contents of this paper are as follows:The first,the general S transform has three disadvantages on discerning artifact automatically?computational complexity and selecting threshold,and it is difficult to processing multi-channel EEG signals.This paper proposed an algorithm to reduce the computational complexity combined with extended Infomax and DOST,which used fractal dimension to discern artifact automatically and applied Otsu method to avoid the loss of useful signal because of the excessive denoising,is then ready for further sleep staging.Second,the general algorithm of sleep staging has two disadvantages on redundant features?high demand of labeled sample.So this paper proposed an algorithm of sleep staging combined with ACO and semi-supervised classification in order to reduce the demand for the labeled samples.The two areas of this section are as follows:1.This section use ACO multi-domain feature selection algorithms to obtain the optimal feature subset which maximize the efficiency of sleep staging for eliminating redundant features and improving efficiency.This algorithm improves the classification accuracy for next step2.This section use weighted voting method combined co-training and active learning strategy after obtaining the optimal feature subset to improve the link of3.ALKLSS algorithm.The ultimate aim of the algorithm is to reduce the demand of labeled samples and guarantee the accuracy of classification.It used MIT-BIH datasets to validate these algorithms.The result demonstrates that the denoising algorithms in first section reduces the computational complexity for processing multi-channel EEG signals;the improved algorithm in second section improves the accuracy by 16.83 and 8.59 percentage than ALKLSS and LS-SVM.It demonstrates the feasibility of this algorithm for the low number labeled samples.