Unsupervised Eye Blink Artifact Identification in

Electroencephalogram
Ashvaany Egambaram, Nasreen Badruddin, Vijanth Asirvadam, Eric Fauvet,
Christophe Stolz, Tahamina Begum

To cite this version:

Ashvaany Egambaram, Nasreen Badruddin, Vijanth Asirvadam, Eric Fauvet, Christophe Stolz, et al..
Unsupervised Eye Blink Artifact Identification in Electroencephalogram. TENCON 2018, Oct 2018,
Jeju, South Korea. pp.2148-2152, �10.1109/TENCON.2018.8650467�. �hal-02140696�

HAL Id: hal-02140696

https://hal.archives-ouvertes.fr/hal-02140696
Submitted on 27 May 2019

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Unsupervised Eye Blink Artifact Identification in
Electroencephalogram
Ashvaany Egambaram, Nasreen Badruddin, Eric Fauvet, Christophe Stolz Tahamina Begum
Vijanth S Asirvadam Laboratoire (Le2i) Department of Neuroscience
Centre for Intelligent Signal and Imaging Research ERL VIBOT CNRS 6000 Universiti Sains Malaysia
Department of Electrical and Electronic Engineering Universite de Bourgogne Kelantan, Malaysia
Universiti Teknologi PETRONAS, Malaysia France
Email:[email protected]

Abstract—The most prominent type of artifact contaminating removal. The most common types of artifacts contaminating
electroencephalogram (EEG) signals is the eye blink (EB) artifact. EEG signals are the cardiac artifact, the muscle artifact and
Hence, EB artifact detection is one of the most crucial pre- the eye blink artifact. The muscle artifact is induced by muscle
processing step in EEG signal processing before this artifact can
be removed. In this work, an approach that identifies EB artifacts movement and contraction during EEG recording which takes
without human supervision and automated varying threshold place when the subject talks and swallows. The pattern of the
setting is proposed and evaluated. The algorithm functions on artifact purely relies on the degree of muscle movement and
the basis of correlation between two EEG electrodes, Fp1 and contraction. Cardiac artifact on the other hand is an effect of
Fp2, followed by EB artifact threshold determination utilizing the electrical activity of the heart. The influence of cardiac artifact
amplitude displacement from the mean. The proposed approach
is validated and evaluated in terms of accuracy and error rate on the scalp is typically low, with low amplitude, repetitive
in detecting events of EB artifacts in EEG signals. Analysis and are of regular pattern. The eye blink (EB) artifact is the
has revealed that the proposed approach achieved an average strongest and most prominent type of artifact recorded along
of 96.6% accuracy compared to a conventional method of with EEG, hence this research will focus on identifying it. EB
identifying EB artifacts with a fixed constant threshold. artifacts appear as spikes with amplitudes of around 10 times
Keywords—Electroencephalogram, EB Artifacts, Automated
Threshold. greater than the actual brain signals, noticeable in the delta
wave range and can last up to 200ms to 400ms [2], [3].
A lot of methods have been developed in the past
I. I NTRODUCTION
for automatic identification or detection of EB artifacts in
The Electroencephalogram (EEG) signal has been in use EEG signal. One of the easiest and preferred ways for EB
to interpret cognitive processes and physiological activity of artifact recognition is by simply using an amplitude threshold
the brain for medical purposes such as diagnosing epilepsy, [4]. This method determines if an EB artifact present in an
sleep disorders, coma, encephalopathies and brain death. The EEG segment if the amplitude in the segment exceeds a
EEG signal is also being extensively used for various research predetermined amplitude threshold. However, the amplitude
purposes, for example in neuroscience, cognitive science, cog- of eye blinks may vary depending on the blinking strength of
nitive psychology, neurolinguistics and psychophysiological an individual, hence this method may not identify EB artifacts
research [1]. In addition, EEG signal is being progressively in- from an individual who exhibits gentle blinks which are lower
vestigated for real time applications such as continual epilepsy than the threshold value. The other common way of detecting
monitoring and brain computer interface (BCI). Hence, it is EB artifacts is through feature based identification which
important to extract meaningful neurological information from determines the presence of EB artifact in an EEG segment
EEG signal to facilitate its interpretation. However, EEG sig- after extracting certain features. Some of the common features
nals are often contaminated with undesired non-neurological used are the kurtosis, maximum absolute value, entropy-based
information which causes deviations in the signal of interest. features and second-order difference in [4], [5], [6], [7]. As a
These undesired signals are termed artifacts. Superimposition general rule, these artifact detection features require a certain
of these artifacts with EEG signal could potentially mystify threshold value to classify or make a binary decision whether
EEG’s interpretation. This is particularly relevant in medical or not an EEG segment is contaminated by EB artifact. As
field where EEG signals are being used as a sole diagnostic elaborated earlier for the amplitude threshold, applying a fix
source, thus failing to recognize artifacts may severely affect threshold value for the features discussed may lead to detection
clinical decisions. Therefore, artifact identification in EEG sig- errors due to the individual variance in blinking pattern and
nal processing is the first and most crucial step prior to artifact blinking strength. As a result, the threshold values may need
to be tailored for every individual which will be impractical artifacts are higher in amplitude in nature relative to that of
in real time applications. This study proposes a new approach the EEG or brain signal, hence the EB artifacts are expected to
to automatically identify EB artifacts with varying threshold yield higher displacement from mean amplitude in comparison
without any supervision on the EEG signal. The performance with uncontaminated EEG potentials. A threshold value is
of the proposed approach is measured by validating if the required in the case of recognizing the starting event of an EB
proposed approach is accurate in identifying EB artifacts in artifact. First, the windows of Fp1 channel that were confirmed
comparison with the use of a fix threshold. containing EB artifacts identified through steps illustrated in
The paper is organized as follows: Section II illustrates section III-A are subjected to attain displacement distribution
the methodology, Section III elaborates the proposed algo- from the mean amplitude using:
rithm, results and discussions are included in Section IV and
V. Finally, Section VI provides the conclusion.
Displacement[t] = |[X[t] − µ]| (1)
II. M ETHODOLOGY
A. Evaluation on Real EEG Signal where, X[t] is the EEG signal’s amplitude at sample t = n
10 sets of EEG signals collected at Universiti Teknologi until t = n+500, n is any sample point of the signal where the
PETRONAS, (UTP) are used to evaluate the proposed al- window starts, and µ is the mean of that window or segment.
gorithm. EB artifacts can be clearly captured in the frontal An EB artifact event detection algorithm within a win-
channels, Fp1-Fp2 electrodes of the EEG recordings. Hence, dow or segment confirmed containing an EB artifact is then
the proposed algorithm is evaluated on the frontal channel, designed. The onset of EB artifact is assumed 100 samples
Fp1, of these EEG signals. All recorded signals are of different (0.39 seconds) before the first sample of amplitude displace-
durations, with a sampling frequency of 256 Hz. The evalu- ment more than the threshold, within the window. The reason
ation was performed using MATLAB R2016b in Windows 7 setting the onset point in advanced of 100 samples before the
Professional (64 bit OS) with a 4GB RAM. threshold is to provide a buffer for any subsequent analysis.
The end point of the EB artifact is found by aligning a 1
III. P ROPOSED A PPROACH second (256 samples) window apart from the first sample of
A. Unsupervised EB Artifact Region Detection amplitude displacement more than the threshold. A 1 second
window is chosen, because an eye blink can last up to 0.8
In an EEG signal recording, the EB artifacts are primarily
seconds (205 samples) in duration which can fit well into this
captured in the frontal electrodes, Fp1 and Fp2. This is because
frame (100+256=356 samples).
the frontal electrodes are positioned close to the eyes. Another
logical point to note here is, both eyes of any individual blinks An experiment was conducted to define classification
simultaneously, hence the Fp1 and Fp2 electrodes should criteria to classify EEG potentials and EB artifacts. The
theoretically unveil high correlation whenever eyes blink. To threshold value for EB artifact for windows containing EB
validate this theory, the correlation between Fp1 and Fp2 artifacts are checked and determined by this experiment. From
channels of an EEG signal is computed in segments of 500 the displacement distribution, a mean (µ) and a standard
samples (1̃.95 seconds) per segment. This window size is deviation (σ) are acquired. Figure 2 shows an example of
chosen so that at least one EB artifact can be captured in displacement distribution by setting the threshold value to be
this window. Figure 1 shows an example of Fp1’s and Fp2’s any displacement values beyond 1σ from the mean, while
recordings plotted out and their corresponding correlation figure 3 shows EB artifacts components plotted in red when
coefficient values in each segment. threshold is set beyond 1σ.
The test has revealed that segments of Fp1 and Fp2
without EB artifact produces correlation below than 0.7, 100
Displacement Distribution for Threshold greater than 1

whereas segments containing EB artifact results in higher 90

correlation, usually more than 0.9. Thus, existence of EB Normal distribution:
80
mu = 9.73735
artifact in a particular segment or window is supported by sigma = 10.4853
70
the high correlation coefficient value, which proves the theory.
60
Following this, segments of the EEG signal contaminated with
Samples

50
EB artifacts are identified. However, EB artifact components
40
should be identified for subsequent analysis or artifact re-
30
moval, which requires a threshold. Next section will discuss on
1
the automated threshold level determination on every window 20

that was identified containing an EB artifact. 10

0
B. Automatic Thresholding to Identify EB Artifacts -30 -20 -10 0 10 20 30 40 50
Amplitude Displacement
Displacement or deviation of amplitude from the mean
is chosen as a threshold criteria to classify the EB artifact’s Fig. 2. Displacement Distribution for Threshold greater than 1σ
onset point. The displacement of amplitude is chosen as EB
 Correlation Coefficient (CC) between Fp1 & Fp2
100
Fp1
Fp2
50

0.45 0.97 0.96

0.19 0.99 0.98 0.98 0.97
0.78 0.98 0.97
-50 0.60 0.33 0.98
0.69 0.54
0.62 0.97

-100

-150

-200
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Samples

Fig. 1. Correlation Coefficient, CC, between Fp1 and Fp2 Electrodes

EB Locations (Threshold = + 1 ) Displacement Distribution for Threshold greater than 2

60 100
EEG signal
EB Artifact Region 90
40
Normal distribution:
80
mu = 9.73735
20
sigma = 10.4853
Amplitude(microvolts)

70
0
60
Samples

-20 50

40
-40
30
-60
EEG detected as EB artifact 20
-80 2
10

-100 0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 -30 -20 -10 0 10 20 30 40 50
Time Samples(s) Amplitude Displacement

Fig. 3. EB Artifacts for Threshold greater than 1σ Fig. 4. Displacement Distribution for Threshold greater than 2σ

From Figure 3, it can be seen that some of the EEG 60

EB Locations (Threshold = + 2 )
potentials are identified as EB artifacts when the threshold is EEG signal
40 EB Artifact Region
set beyond 1σ from the mean.
Figure 4 shows the displacement distribution of the same 20
Amplitude(microvolts)

segment by setting the threshold value to be any displacement 0

values beyond 2σ from the mean, while figure 5 shows EB
-20
artifacts components plotted in red when threshold is set
beyond 2σ. -40

From Figure 5, it can be seen that all EB artifacts are -60

identified as EB artifacts when the threshold is set beyond 2σ
-80
from the mean.
Figure 6 shows the same displacement distribution by -100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
setting the threshold value to be any displacement values Time Samples(s)
beyond 3σ from the mean, while figure 7 shows EB artifacts
components plotted when threshold is set beyond 3σ. Fig. 5. EB Artifacts for Threshold greater than 2σ
From Figure 7, it can be seen that the EB artifacts are
not identified as EB artifacts when the threshold is set beyond
3σ from the mean. EEG window in question can be correctly determined by
Therefore the threshold for EB artifact dominating the taking two standard deviation, 2σ width from mean of the
 Displacement Distribution for Threshold greater than 3 Start
100

90
Normal distribution: Window size =500,
80
mu = 9.73735 EB end =1st sample of x(n)
70 sigma = 10.4853

60
Samples

Window (Fp1 & Fp2) =EB end +Window size

50

40
Find Correlation Coefficient (CC) between EB end =last point of
30 Fp1 & Fp2 window window samples

20

10 3 CC >0.9
NO

0
-30 -20 -10 0 10 20 30 40 50 YES
Amplitude Displacement
Compute Amplitude Displacement (Deviation)
Threshold =( mean +2 ) of Displacement
Fig. 6. Displacement Distribution for Threshold greater than 3σ Distribution

EB start =1st sample >threshold

EB onset =100 samples before EB start
EB end =1 second after EB start,
EB region =EB onset to EB end,
Store this region as an EB artifact region

End

Fig. 8. Flowchart of the Automated EB Artifact Identification Algorithm

TABLE I
C ONFUSION M ATRIX - EB A RTIFACT D ETECTION

Detected
EB Artifact Clean EEG
EB Artifact TP FN
Observed
Clean EEG FP TN

Fig. 7. EB Artifacts for Threshold greater than 3σ

the EEG signals with measures derived from the confusion
matrix, such as the accuracy and error rate.
displacement distribution acquired. Error: The error rate, ERR is computed with mis-
detection of EB artifact and EEG divided by total number of
threshold = mean + 2σ (2) detections. Values approaching zero denotes better error rate.
Figure 8 summarizes the EB artifact detection algorithm FP + FN
ERR = (3)
in a flowchart. TP + TN + FN + FP
IV. R ESULTS Accuracy: ACC is the opposite of error rate, where it is
This section aims to measure the accuracy of the pro- the ratio of correct EB artifact and EEG detections by total
posed approach in detecting EB artifacts compared to the use number of detections. The best accuracy is 1.
of a constant or common threshold. The constant threshold TP + TN
is evaluated by fixing the threshold values of more than ACC = (4)
TP + TN + FN + FP
10uV, 20uV, 30uV, 40uV and 50uV, whichever amplitude
The results obtained from the proposed method in com-
displacement that exceeds these thresholds are considered to
parison with constant thresholds in detecting EB artifacts
be EB artifacts.
correctly, applied on the Fp1 channel of 10 EEG signals are
The binary prediction is used to test the detection of EB
tabulated in Table II.
artifacts, which produces below outcomes and the confusion
matrix as in Table I: V. D ISCUSSIONS
• True positive (TP): correct EB artifact detection As elaborated in Section III-A, an automated EB arti-
• False positive (FP): clean EEG identified as EB artifact fact detection algorithm is designed. The proposed approach
• True negative (TN): correct clean EEG identification detects EB artifacts without any human supervision. First
• False negative (FN): EB artifact identified as clean EEG EEG segments containing EB artifacts are recognized with the
The efficiency of the proposed algorithm compared to concept of correlation between EEG electrodes, Fp1 and Fp2.
fix a constant threshold, is validated by manually inspecting Secondly, to address the issue of fixing a constant threshold,
 TABLE II
C OMPARISON OF ACCURACY

Accuracy
Proposed Fix Thresholds
Signal
(Automated Varying Threshold) 10uV 20uV 30uV 40uV 50uV
EEG 1 94.4% 97.2% 91.7% 77.8% 50.0% 50.0%
EEG 2 98.9% 98.9% 97.8% 89.9% 57.3% 42.7%
EEG 3 97.9% 92.5% 98.9% 96.3% 49.7% 21.4%
EEG 4 97.5% 94.9% 94.9% 75.9% 27.8% 26.6%
EEG 5 96.9% 96.1% 97.7% 98.4% 97.7% 70.3%
EEG 6 94.3% 94.3% 94.3% 100.0% 98.7% 98.7%
EEG 7 97.9% 94.7% 97.9% 96.8% 97.9% 97.9%
EEG 8 97.0% 96.3% 94.0% 94.8% 92.5% 65.7%
EEG 9 94.8% 89.6% 95.7% 96.5% 87.0% 73.0%
EEG 10 96.2% 93.7% 94.9% 94.9% 94.3% 81.0%
AVERAGE 96.6% 94.8% 95.8% 92.1% 75.3% 62.7%

an automated and varying threshold is determined for every EB artifact detection algorithm proposed in this paper is
EEG segment containing EB artifact using the amplitude dis- accurate in identifying EB artifact events in an EEG signal.
placement. From Table II, the accuracy achieved in detecting On the other hand, the algorithm is also consistent in detecting
EB artifacts by the proposed technique is higher in average, EB artifacts across different EEG signals compared to a con-
96.6% compared to 94.8%, 95.8%, 92.1%, 75.3% and 62.7% ventional algorithm which is fixed with a constant threshold to
achieved through fixing thresholds between 10uV to 50uV detect EB artifacts. Thus, it can be concluded that the proposed
of amplitude displacement. The individual accuracy value for algorithm is a reliable solution in detecting EB artifacts across
each EEG signal of the proposed automated threshold reveals all types of EEG signals which may have individual variance
the accuracy has not fluctuated much, between 94% to 99% due to blinking pattern and strength.
compared to the constant thresholds, between 89% to 99% for
ACKNOWLEDGMENT
10uV, between 91% to 99% for 20uV, between 75% to 100%
for 30uV, between 27% to 99% for 40uV, and between 21% The authors would like to thank the Ministry
to 99% for 50uV. The constant threshold has achieved 100% of Education, Malaysia for supporting this research
accuracy for EEG set 6, which means the constant threshold through the Fundamental Research Grant Scheme, FRGS
of more than 30uV for EB artifact detection suits well for (FRGS/2/2014/TK03/UTP/02/1) and the Higher Institution
this data set. On the other hand, the 30uV threshold has only Centre of Excellence (HICoE) Scheme.
achieved 75.9% of accuracy for EEG set 4. This indicates the R EFERENCES
performance of fixing a constant threshold is purely dependent
[1] J. A. Urigüen and B. Garcia-Zapirain, “Eeg artifact removalŮstate-of-
on the nature of the EEG signal and is not consistent across all the-art and guidelines,” Journal of neural engineering, vol. 12, no. 3, p.
EEG signals. While the automated threshold changes for every 031001, 2015.
window, whereby the threshold will correspond to the varying [2] G. H. Klem, H. O. LuÈders, H. Jasper, C. Elger et al., “The ten-twenty
electrode system of the international federation,” Electroencephalogr Clin
nature of the EB artifacts. Hence, threshold is automatically Neurophysiol, vol. 52, no. 3, pp. 3–6, 1999.
determined for every window without setting any specific [3] S. Sanei and J. A. Chambers, EEG signal processing. John Wiley &
value. Sons, 2013.
[4] H. Nolan, R. Whelan, and R. Reilly, “Faster: fully automated statistical
thresholding for eeg artifact rejection,” Journal of neuroscience methods,
VI. C ONCLUSION vol. 192, no. 1, pp. 152–162, 2010.
Conventional EB artifact detection algorithms depends [5] G. Barbati, C. Porcaro, F. Zappasodi, P. M. Rossini, and F. Tecchio,
“Optimization of an independent component analysis approach for artifact
on constant thresholds or constant features to make a binary identification and removal in magnetoencephalographic signals,” Clinical
decision to recognize if an EEG segment contains EB artifact Neurophysiology, vol. 115, no. 5, pp. 1220–1232, 2004.
or not. In this paper, an unsupervised EB artifact detection [6] A. Mognon, J. Jovicich, L. Bruzzone, and M. Buiatti, “Adjust: An
automatic eeg artifact detector based on the joint use of spatial and
algorithm in EEG signal is proposed and evaluated. The temporal features,” Psychophysiology, vol. 48, no. 2, pp. 229–240, 2011.
algorithm relies on the concept of correlation between two [7] A. Klein and W. Skrandies, “A reliable statistical method to detect
EEG electrodes and the amplitude displacement range when eyeblink-artefacts from electroencephalogram data only,” Brain topog-
raphy, vol. 26, no. 4, pp. 558–568, 2013.
there is an event of EB activity. Based on the results and
discussions, it is apparent that the automated and unsupervised