Low Complexity Automatic Stationary Wavelet Transform for Elimination of Eye Blinks from EEG

Abstract

The electroencephalogram signal (EEG) often suffers from various artifacts and noises that have physiological and non-physiological origins. Among these artifacts, eye blink, due to its amplitude is considered to have the most influence on EEG analysis. In this paper, a low complexity approach based on Stationary Wavelet Transform (SWT) and skewness is proposed to remove eye blink artifacts from EEG signals. The proposed method is compared against Automatic Wavelet Independent Components Analysis (AWICA) and Enhanced AWICA. Normalized Root Mean Square Error (NRMSE), Peak Signal-to-Noise Ratio (PSNR), and correlation coefficient ( ρ ) between filtered and pure EEG signals are utilized to quantify artifact removal performance. The proposed approach shows smaller NRMSE, larger PSNR, and larger correlation coefficient values compared to the other methods. Furthermore, the speed of execution of the proposed method is considerably faster than other methods, which makes it more suitable for real-time processing.

Keywords: SWT; electroencephalography; eye blink; skewness.

Figure 1

Examples of signals and histograms for: clean EEG (a), blink artifact (b) and contaminated EEG (c). S is the skewness value.

Figure 2

The block diagram of the proposed method.

Figure 3

Examples of simulated data from CHB-MIT database: $x\left(n\right)$—pure EEG, $r\left(n\right)$—eye blink artifact, $z_1\left(n\right)$—contaminated EEG with $a=0.75$, $z_2\left(n\right)$—$a=1.0$, $z_3\left(n\right)$—$a=1.5$, and $z_4\left(n\right)$—$a=2.0$.

Figure 4

Mean ± std of NRMSEs per different T values.

Figure 5

Examples of approximation coefficients and corresponding skewness values for a contaminated EEG signal from CHB-MIT database in SWT domain.

Figure 6

Box plots of NRMSE and correlation coefficient between pure and filtered EEG for simulated data by all methods: (a,b) are for CHB-MIT, (c,d) are for EEG-MAT databases.

Figure 7

Examples of eye blink cancellation in simulated EEG signals from: EEG-MAT (a), and CHB-MIT (b) databases: z(n)—contaminated EEG, r(n)—real eye blink artifact, x(n)—pure EEG, $x_1\left(n\right)$—filtered EEG by the proposed method, $x_2\left(n\right)$—filtered EEG by the AWICA and $x_3\left(n\right)$—filtered EEG by the EAWICA.

Figure 8

Examples of the PSDs for the pure and the filtered EEG signals by all methods for simulated data using CHB-MIT (a) and EEG-MAT (b) databases.

Figure 9

PSNR curves as the function of NRMSE for filtered EEG signals: CHB-MIT (a) and EEG-MAT (b) databases. ASWT outperformed the other algorithms because in each subplot, the points associated with the largest PSNR and the smallest NRMSE were achieved by ASWT.

Figure 10

Examples of eye blink cancellation in real EEG signals from: BCI Competition 2008—Graz Data Sets 2a (a), and BCI 2011 left/right motor imagery (b). z(n)—EEG contaminated with eye blink, $x_1\left(n\right)$—filtered EEG by the proposed method, $x_2\left(n\right)$—filtered EEG by AWICA and $x_3\left(n\right)$—filtered EEG by EAWICA.