4.3 Simulation Studies
4.3.3 Simulation Results using Real Life EEG Signals
To assess the efficacy of proposed SSA-ANC technique on the real life contaminated EEG signals, we consider EEG data recorded for BCI motor imagery study [83, 134]. As this data is recorded for movement imagery task, there will be a α (8−12Hz) component present in the measured EEG signal. To evaluate the proposed SSA-ANC technique on this data, 8 sec of EEG epoch contaminated by EOG artifact is considered. Fig. 4.8.
shows the estimated EOG reference signal and corrected EEG signal by DWT-ANC.
It is noticed from Fig. 4.8(b)-(c) that the reference and corrected EEG signals using DWT-ANC are poorly reconstructed, shown in circle. The reason is that the morpholo- gies of the wavelet function and eye blink at that instant are not matched. From these results it is noted that the performance of the DWT-ANC is depends on the selection of mother wavelet function. In the proposed SSA-ANC technique, after performing first
0 1 2 3 4 5
−2 0 2
(a)
0 1 2 3 4 5
−2 0 2
(b)
0 1 2 3 4 5
−2 0 2
Time in sec
(c)
Figure 4.6: (a) synthetic contaminated EEG signaly(n)forSN R= 1, (b) extracted reference signals for ANC using DWT and (c) extracted reference signal for ANC using
SSA.
0 1 2 3 4 5
−1 0 1
(a)
0 1 2 3 4 5
−1 0 1
(b)
0 1 2 3 4 5
−1 0 1
Time in sec
(c)
Figure 4.7: (a) Synthetically contaminated EEG signal y(n) forSN R = 1, (b) cor- rected EEG signal by the DWT-ANC and (c) corrected EEG signal by the proposed
SSA-ANC.
Chapter 4 Introduction 53
466 467 468 469 470 471 472 473
−1 0 1
(a)
466 467 468 469 470 471 472 473
−1 0 1
(b)
466 467 468 469 470 471 472 473
−1 0 1
Time in sec
(c)
Figure 4.8: (a) contaminated EEG signal, (b) estimated reference signal for ANC using DWT and (c) corrected EEG signal by DWT-ANC.
0 10 20 30 40 50 60 70
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Index of eigenvectors Local mobilit m v
Figure 4.9: Local Mobilities of eigenvectors for window lengthL= 68.
466 467 468 469 470 471 472 473
−1 0 1
(a)
466 467 468 469 470 471 472 473
−1 0 1
(b)
466 467 468 469 470 471 472 473
−1 0 1
Time in sec
(c)
Figure 4.10: (a) contaminated EEG signal, (b) estimated reference signal for ANC using SSA and (c) corrected EEG signal by proposed SSA-ANC.
two steps of SSA, local mobility of each eigenvector is computed. Fig. 4.9 shows the local mobility of each eigenvector and dotted line indicates the threshold0.1. From Fig.
4.9 it is found that eigenvectorsv1,v2andv3 are the basis vectors to construct the EOG signal (reference) for ANC. The reconstructed reference and corrected EEG signals using proposed SSA-ANC technique are shown in Fig. 4.10. Comparing the reference signals obtained by DWT-ANC (Fig. 4.8(b)) and proposed SSA-ANC (Fig. 4.10(b)), SSA out- performs in reconstructing the EOG signal than the DWT. In conventional methods, the extracted EOG signal is simply subtracted from the contaminated EEG to obtained artifact free EEG signal. However, this technique fails to track the signal changes due to the eye blink. Hence use of ANC enhance in the process of estimating EOG artifact.
Fig 4.11(a) shows the reconstructed EOG signals at output of SSA and adaptive filter blocks. The improvement in the estimated EOG signal can be clearly observed in Fig.
4.11(b). From the above simulations it is concluded that, unlike DWT-ANC, where the performance is depends on the morphology of the EOG artifact, the proposed SSA-ANC performance does not depend on the morphology of the EOG signal, hence is a suitable candidate for the removal of EOG artifact from the EEG signal. The efficiency of the proposed SSA-ANC technique in removing the EOG artifact can be observed from the power spectral density (PSD) of the contaminated and corrected EEG signals, shown in Fig. 4.12. It is clear from Fig. 4.12 that the proposed SSA-ANC technique efficiently
Chapter4: Removal of Eye Blink Artifact from Single Channel EEG Signal using
SSA-ANC Technique 55
466 467 468 469 470 471 472 473
−1
−0.5 0 0.5 1
(a)
r2(n) ˆ r1(n)
466 466.5 467 467.5 468
−1
−0.5 0 0.5 1
Time in sec
(b)
r2(n) ˆ r1(n)
Figure 4.11: (a) estimated EOG reference signal by SSA r2(n) and adaptive filter outputˆr1(n), (b) zoomed version of above plot between time interval466to468sec.
attenuate the EOG signal and is justified as small power in frequency band 0.5−4Hz than the PSD of the corrected EEG signal by DWT-ANC. In order to show that the pro- posed SSA-ANC technique is preserves the EEG α component, we define a parameter called, mean absolute error (MAE) and is given by
M AE=
∑l
k=j |pe(k)−py(k)|
l−j (4.4)
where,peandpyare the PSD of the corrected and contaminated EEG signals respectively andjandlare the specific frequency bands. First, we applied DWT-ANC and proposed SSA-ANC techniques on the five EEG epochs which are fully contaminated by EOG artifacts and compute the PSDs of contaminated and corrected EEG signals. Later MAE for the EEGα band 10−12Hzis computed. Table. 4.1 shows the MAE for both techniques and is clear that on average MAE for proposed SSA-ANC results 0.1671dB error, which is smaller than the DWT-ANC. Hence we can conclude that proposed SSA- ANC technique is not altering the EEGα component.
The proposed SSA-ANC technique is also applied on the CHB-MIT scalp EEG database [3, 8] to remove EOG artifact in online perspective as shown in 4.2. Fig 4.13 and 4.14 shows the estimated EOG signals at the output of SSA and ANC blocks. After careful
0 5 10 15 20 25 30 35 40
−50
−45
−40
−35
−30
−25
−20
−15
Freq in Hz
Mag in dB’s
(a) (b) (c)
Figure 4.12: PSD of (a) contaminated EEG signal, (b) corrected EEG signal by DWT-ANC and (c) corrected EEG signal by SSA-ANC.
Table 4.1: Mean Absolute Error (MAE) of Two Techniques.
Segment DWT-ANC (dB) Proposed SSA-ANC (dB)
1 0.4382 0.1949
2 0.7456 0.2072
3 0.2440 0.2152
4 0.1057 0.0980
5 0.5501 0.1206
Average 0.4167 0.1671
observation of Figs. 4.13 and 4.14 it is noticed that SSA partially estimated the EOG artifact and is fully reconstructed by the adaptive filter. Once again the combination of SSA-ANC technique enhances ability of the EOG artifact removal.
Chapter 4 Introduction 57
0 20 40 60 80 100 120
−1
−0.5 0 0.5 1
(a)
20 22 24 26 28 30
−1
−0.5 0 0.5 1
(b)
Time (s)
x(n) r2(n)
Figure 4.13: (a) Contaminated EEG signal (green) x(n) and the estimated EOG r2(n)by SSA technique, and (b) zoomed version of (a) between20−30stime period.
0 20 40 60 80 100 120
−1
−0.5 0 0.5 1
(a)
20 22 24 26 28 30
−1
−0.5 0 0.5 1
(b)
Time (s)
x(n) ˆ r1(n)
Figure 4.14: (a) Contaminated EEG signal (green)x(n) and estimated EOG rˆ1(n) at ANC, and (b) zoomed version of (a) between20−30stime period.