Comparison of Methods of Artifacts Removal
Regression Methods
Regression methods often assume that the scal potential is a linear combination of brain and other potentials(EOG、ECG、EMG).By subtracting propagated EOG/ECG/EMG from EEG recordings,EEG signals can be recovered.
Regression can also be done in frequency domain based on the concept that subtraction in the frequency domain is equivalent to filtering in the time domain.By eliminating spectral estimates of EOG/EMG/ECG from EEG recordings,it is possible to recover the non-contaminated EEG.
Disadvantages: Both types of regression methods are off-line and rely on EOG/ECG/EMG recordings,which are however,not always available.
PCA(Principle Component Analysis)
This method assumes that each EEG channel recording is simultaneously generated by multiple sources across the scalp. By decomposing multiple channel EEG data into principle components using PCA,the artifactual sources can be identified and removed.
Disadvantages: PCA methods usually failed to completely separate artifacts from cerebral activities,and the orthogonal assumption fro data components,which is always required while using PCA,is hardly satisfied.
ICA(Independent Component Analysis)
ICA was originally developed fro blind source separation(BSS)
Disadvantages: ICA usually requires a large amount of data and visual inspection to eliminate noisy independent components,making the method time-consuming and not suitable for real-time applications.
Wavelet Analysis
It is effective to mesure and manipulate non-stationary signals.In wavelet-based methods,the wavelet thresholding techniques have received significant attention.For this class of methods,wavelet coefficitents at low-frequency sub-bands are corrected by some thresholding functions before signal reconstruction.
Advantages: As an online artifact removal method,the most important advantage of using this method for EEG correction is that it does not rely on either the reference signal or visual inspection.
Disadvantages: It performance is not consistent because the method is sensitive to the selections of wavelet basis and thresholding functions.