Abstract
The unavoidable muscle artifacts pose challenges on reliable interpretation of the electroencephalogram (EEG) recordings, especially for the wearable few-channel EEG, a new emerging scenario. However, the high computational load and low robustness of the existing methods limit its wider applications and performance in artifact removal. Consequently, we propose an efficient and robust muscle artifact removal approach by jointly employing the Fast Multivariate Empirical Mode Decomposition (FMEMD) and CCA for few-channel EEG. The proposed FMEMD-CCA firstly efficiently decomposes the input EEG recordings into several multivariate Intrinsic Mode Functions (IMF) by applying FMEMD. Secondly, all the multivariate IMFs are processed by CCA for computing the underlying sources. Finally, the sources with low autocorrelations are smartly determined as muscle artifacts and rejected, and therefore the other components are reconstructed for EMG-artifact-free IMFs and EEG. Simulated and real data experiments are carried out for verifying the performance of the proposed method. It takes 10 times less computing time in FMEMD-CCA compared with in Multivariate Empirical Mode Decomposition (MEMD)-CCA for 10-s EEG recordings, using the same computer and software. And the accuracy and the average correlation coefficient are highly consistent in both approaches. Furthermore, in contrast to MEMD-CCA, the proposed FMEMD-CCA works more robustly in low sampling rate based on the real data and benchmark.
Highlights
Scalp EEG measures and records voltage fluctuations resulting from the cortical and subcortical neurons with the electrodes placed along the scalp [1]
The proposed Fast Multivariate Empirical Mode Decomposition (FMEMD)-Canonical Correlation Analysis (CCA) is featured by following steps: (i) FMEMD first decomposes the input EEG recordings into several multivariate Intrinsic Mode Functions (IMF); (ii) The obtained intrinsic modes are processed by CCA for computing the underlying sources; (iii) Those sources with low autocorrelations will be smartly designated as muscle artifacts and rejected and (iv) The EMG-artifact-free IMFs and EEG are reconstructed from the retained components
We compare results obtained from the CCA, Multivariate Empirical Mode Decomposition (MEMD)-CCA and FMEMD-CCA approaches
Summary
Scalp EEG measures and records voltage fluctuations resulting from the cortical and subcortical neurons with the electrodes placed along the scalp [1]. The proposed FMEMD-CCA is featured by following steps: (i) FMEMD first decomposes the input EEG recordings into several multivariate IMFs; (ii) The obtained intrinsic modes are processed by CCA for computing the underlying sources; (iii) Those sources with low autocorrelations will be smartly designated as muscle artifacts and rejected and (iv) The EMG-artifact-free IMFs and EEG are reconstructed from the retained components. We compare our proposed algorithm with the state-of-the-art MEMD-CCA EMG-artifact removal approach in scenario of few-channel EEG, and results produced by simulation and real experiments demonstrate improved performances in terms of computation cost and accuracy. Overview of the proposed method, consisting of 5 steps: (1) Decompose: few-channel EEG, in addition to several-channel White Gaussian Noise (WGN) are decomposed into multivariate IMFs; (2) Unmix: compute sources with CCA based on the EEG’s IMFs; (3) Select: identify the muscle artifacts by setting the autocorrelations’ threshold; (4) Reject: reject the muscle artifacts by inverse CCA; and (5) Reconstruct: reconstruct the EMG-artifact-free EEG. The last K ∗ M IMFs corresponding to the M -channel WGN will be discarded
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
