Sleep staging from the EEG signal using multi-domain feature extraction

Abstract

Abstract The analysis of the electroencephalogram (EEG) can yield much useful information about brain function, including indications of sleep stage. During the process of EEG analysis, feature extraction is one of the most critical technical aspect. Traditional EEG feature extraction methods are mainly based on single domain analysis. However, due to the highly non-stationary and nonlinear characteristics of the EEG, it is difficult to extract comprehensive information only from single domain analysis. In the present study, a novel feature extraction method was proposed based on the multi-domain analysis of the EEG. Fifteen characteristic parameters were extracted based on the multifractal detrended fluctuation analysis (MF-DFA), visibility graph algorithm (VGA), frequency analysis and nonlinear analysis. Ten optimal parameters of the fifteen parameters were selected by the genetic algorithm (GA). Then the Least Squares-Support Vector Machines (LS-SVM) were used to classify the sleep states. The cross validation results demonstrated that multi-domain feature extraction method can obtain more useful information in the EEG signal. Compared to the frequency domain parameters, nonlinear parameters and time domain parameters, the predictive accuracy of sleep staging classification with optimal multi-domain parameters improved 11.08%, 10.76% and 6.40% respectively.