EEG Denoising Based on Empirical Mode Decomposition and Mutual Information

Abstract

Empirical mode decomposition (EMD) is a recently introduced decomposition method for non-stationary time series. EMD has an information preserving property so the sum of the decomposed intrinsic mode functions (IMF) can be used to reconstruct the original signal. However, when the signal is corrupted by white Gaussian noise, some of the IMFs may contain most of the noise components. Thus, determining which IMFs have informative oscillations or information free noisy components becomes the main challenge in denoising. In this study, mutual information (MI) is used as a metric to find information free noisy IMFs. Without using an extra reference signal, MI of each IMF and MI of the autocorrelation function (ACF) of each IMF is computed. An adaptive thresholding scheme based on MI scores is applied to decide which IMFs contain most of the white noise. Proposed method is tested on epileptic and normal EEG recordings corrupted by additive white noise to show its denoising capability on stochastic time series.