Improved EOG Artifact Removal Using Wavelet Enhanced Independent Component Analysis

Abstract

Electroencephalography (EEG) signals are frequently contaminated with unwanted electrooculographic (EOG) artifacts. Blinks and eye movements generate large amplitude peaks that corrupt EEG measurements. Independent component analysis (ICA) has been used extensively in manual and automatic methods to remove artifacts. By decomposing the signals into neural and artifactual components and artifact components can be eliminated before signal reconstruction. Unfortunately, removing entire components may result in losing important neural information present in the component and eventually may distort the spectral characteristics of the reconstructed signals. An alternative approach is to correct artifacts within the independent components instead of rejecting the entire component, for which wavelet transform based decomposition methods have been used with good results. An improved, fully automatic wavelet-based component correction method is presented for EOG artifact removal that corrects EOG components selectively, i.e., within EOG activity regions only, leaving other parts of the component untouched. In addition, the method does not rely on reference EOG channels. The results show that the proposed method outperforms other component rejection and wavelet-based EOG removal methods in its accuracy both in the time and the spectral domain. The proposed new method represents an important step towards the development of accurate, reliable and automatic EOG artifact removal methods.

Keywords: EEG; EOG artifacts removal; discrete wavelet transform (DWT); independent component analysis.

Figure 1

The data processing flowchart of the proposed EOG removal method.

Figure 2

An example for frontal channels (marked by red circles) used for correlation calculation in EOG independent component identification. Top view of scalp with nose pointing upwards, 128-channel Biosemi ABC electrode layout.

Figure 3

Distribution of the normalized weights of the components of 20 EOG contaminated measurements selected from the Klados datasets. The red crosses represent the weight of the EOG (HEOG and VEOG) components.

Figure 4

Distribution of the normalized independent component analysis (ICA) component weights of 10 selected PhysioNet datasets.

Figure 5

Distribution of the normalized ICA component weights of the 22 datasets obtained in our laboratory.

Figure 6

ICA components of a selected Klados dataset.

Figure 7

Sample sections of VEOG (blue) and HEOG (red) EOG components from four selected Klados datasets.

Figure 8

Correction target windows around the detected VEOG blink (a) and HEOG eye movement (b) peaks in the EOG ICA components.

Figure 9

The wavelet decomposition process and calculation of coefficients. Letters F and G represent the output signals of the low-pass and high-pass filters, respectively.

Figure 10

Wavelet decomposition of a target EOG peak signal window within an independent component.

Figure 11

Illustration of the cleaning performance on one artifact contaminated section of the Klados dataset9. The two subplots on the right show the difference of the pure electroencephalography (EEG) data and the wavelet-enhanced ICA (wICA) and proposed method (PM) cleaned signals, respectively. Amplitude scales are different to make difference signal visible.

Figure 12

Comparison of the artifact-free, the contaminated and the PM-cleaned EEG signals of dataset9, channel Fp1.

Figure 13

Distribution of the $\lambda$ (a), difference in signal-to-noise ratio $\mathsf{\Delta }SNR$ (b) and root mean square error $RMSE$ (c) dataset average values obtained with the rejection ICA, wICA and the proposed method. For $\lambda$ and $\mathsf{\Delta }SNR$ the higher, while for $RMSE$, the lower values mean better performance.

Figure 14

Power spectral density distributions of the pure, contaminated versus the ICA rej, wICA and PM method cleaned signals (dataset12, channel Fp1).

Figure 15

The grand average (20 datasets) magnitude squared coherence (MSC) results of the three cleaning methods. Note the higher average performance of our proposed method.

Figure 16

The magnitude squared coherence (MSC) between the pure EEG signal and the contaminated signal as well as the various cleaned signals (dataset12, Fp1).

Figure 17

Distribution of the $\lambda$ (a), $\mathsf{\Delta }SNR$ (b) and $RMSE$ (c) dataset average values for the resting state laboratory measurements obtained by cleaning with the rejection ICA, wICA and PM methods. For $\lambda$ and $\mathsf{\Delta }SNR$ the higher, while for $RMSE$, the lower values mean better performance.

Figure 18

MSC values obtained with different cleaning methods for the resting state laboratory dataset (20 subjects).

Figure 19

A 128-channel EOG contaminated EEG dataset before (a) and after (b) artifact removal.

Figure 20

Topoplot potential map (µV) of a 128-channel EOG contaminated resting state measurement before (left) and after artifact removal (right).

Figure 21

Distribution of the $\lambda$ (a), $\mathsf{\Delta }SNR$ (b) and $RMSE$ (c) dataset average values for the PhysioNEt P300 dataset by cleaning with the rejection ICA, wICA and PM methods. For $\lambda$ and $\mathsf{\Delta }SNR$ the higher, while for $RMSE$, the lower values mean better performance.

Figure 22

Event related potential (ERP) signals computed from artifact-free epochs only (a) and ERP signals computed from all cleaned epochs (b) showing the distorting effects of the cleaning methods on ERP curves. $ER{P}_{clean}^{PM}$ produced the smallest difference in both cases (dataset, electrode Fpz).