FastEMD–CCA algorithm for unsupervised and fast

removal of eyeblink artifacts from electroencephalogram

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

To cite this version:

Ashvaany Egambaram, Nasreen Badruddin, Vijanth S. Asirvadam, Tahamina Begum, Eric Fau-
vet, et al.. FastEMD–CCA algorithm for unsupervised and fast removal of eyeblink artifacts
from electroencephalogram. Biomedical Signal Processing and Control, 2020, 57, pp.101692 -.
�10.1016/j.bspc.2019.101692�. �hal-03489122�

HAL Id: hal-03489122

https://hal.science/hal-03489122v1
Submitted on 21 Jul 2022

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.

Distributed under a Creative Commons Attribution - NonCommercial 4.0 International License

Version of Record: https://www.sciencedirect.com/science/article/pii/S1746809419302733
Manuscript_39c3da81c19847ecb9bb4a1209a2942f

FastEMD–CCA Algorithm for Unsupervised and Fast

Removal of Eyeblink Artifacts from
Electroencephalogram

Ashvaany Egambarama,c,e , Nasreen Badruddina,c , Vijanth S Asirvadamb,c ,

Tahamina Begumd , Eric Fauvete , Christophe Stolze
a Institute of Health and Analytics, Universiti Teknologi PETRONAS, Perak, Malaysia
b Institute of Autonomous Systems, Universiti Teknologi PETRONAS, Perak, Malaysia
c Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS,

Perak, Malaysia
d Department of Neuroscience, Universiti Sains Malaysia, Kelantan, Malaysia
e Laboratoire Electronique, Informatique et Image (Le2i), ERL VIBOT CNRS 6000,

Universite de Bourgogne,France

Abstract

Online detection and removal of eye blink (EB) artifacts from electroencephalo-
gram (EEG) would be very useful in medical diagnosis and Brain-Computer
Interface (BCI). In this work, approaches that combine unsupervised eyeblink
artifact detection with Empirical Mode Decomposition (EMD), and Canoni-
cal Correlation Analysis (CCA), is proposed to automatically identify eyeblink
artifacts and remove them in an online manner. First eyeblink artifact re-
gions are automatically identified and an eyeblink artifact template is extracted
via EMD, which incorporates an alternate interpolation technique, the Akima
spline interpolation. The removal of eyeblink artifact components relies on the
elimination of EEG canonical components obtained through CCA, based on
cross-correlation with the extracted eyeblink artifact template. The proposed
algorithm is evaluated and analysed with respect to its ability in removing eye-
blink artifacts and retaining neural information of the EEG signals. Analysis
proved that the proposed algorithm, FastEMD-CCA, is effective in eyeblink ar-

Email addresses: [email protected] (Ashvaany Egambaram),

[email protected] (Nasreen Badruddin), [email protected] (Vijanth S
Asirvadam), [email protected] (Tahamina Begum), [email protected] (Eric
Fauvet), [email protected] (Christophe Stolz)

Preprint submitted to Biomedical Signal Processing and Control October 13, 2019

© 2019 published by Elsevier. This manuscript is made available under the CC BY NC user license
https://creativecommons.org/licenses/by-nc/4.0/
tifact removal with an average accuracy, sensitivity, specificity and error rate of
97.9%, 97.65%, 99.22% and 2.1% respectively. The algorithm is able to clean
and remove eyeblink artifacts from a 14-channel EEG of length 1 second, at an
average time of 63 milliseconds. This makes it a feasible solution for applications
requiring online removal of eyeblink artifacts.
Keywords: Electroencephalogram (EEG), Enhanced Empirical Mode
Decomposition (FastEMD), Canonical Correlation Analysis (CCA), Eyeblink
Artifact.

1. Introduction

An electroencephalogram (EEG) is used to record and evaluate electrical

potentials generated during cerebral activity. The EEG signal has been in use to
interpret cognitive processes and physiological activity of the brain for medical
purposes and extensively used for various research purposes. An EEG signal
does not only consist of electrical potentials related to brain activities, but it is
invariably contaminated by electrical signals originating from other parts of the
body. These unwanted signals are referred to as artifacts. The superimposition
of these artifacts with the EEG signal could potentially lead to inaccurate EEG
interpretation. This issue is particularly relevant in the medical field where
EEG signals are widely used as a diagnostic tool, thus failing to recognize and
remove artifacts may affect clinical decisions. Therefore, artifact identification
and removal in EEG signal processing are the first and most crucial step.
The most common types of artifacts contaminating EEG signals are the
cardiac artifact, the muscle artifact and the eye blink artifact [1]. The muscle
artifact is caused by muscle movement and contraction, which may take place
when the patient talks or swallows. The cardiac artifact solely arises from the
electrical activity of the heart. Out of all these artifacts, eyeblink artifact is
the one most prominently present in EEG signals as blinking the eyes produces
relatively large electrical potentials around the eyes, hence this work will focus
on identifying and removing it. Eyeblink artifacts appear as spikes with am-

2
plitudes of around 10 times greater than the actual brain signals, noticeable in
the delta wave range and can last up to 200ms to 400ms [2, 3]. The eyeblink
potential propagates and spreads out to all EEG electrodes but in various con-
duction volume - higher conduction near the frontal and parietal regions while
the conduction in the occipital region is very low. The frontal region is the most
prone region to contamination from eyeblink artifacts as it is closest to the eyes.
Fig. 1 shows the positions of EEG electrodes following the 10-20 system. The
Fp1 and Fp2 electrode positions, which are closest to the eyes and highlighted
in Fig. 1, can be used to capture the eyeblink artifacts.

Fig. 1: EEG Electrode Placement

For reliable analysis of EEG signals, it is therefore essential that these arti-
facts be removed. Traditionally artifact removal is done after the EEG signal
has been recorded, either manually or automatically. However, in clinical moni-
toring such as continuous epilepsy monitoring and the brain-computer interface
(BCI), where EEG signals are analyzed and manipulated as they are being
recorded, an online artifact removal solution is required [4]. Various techniques
are available for de-noising purposes, which will be discussed below. The most
common method is to have eyeblink artifact regions identified through man-
ual inspection and these segments are removed. This method can cause a loss

3
of information as the EEG segments being removed may contain useful neu-
rological information. Regression-based methods [5–7] perform a regression or
correlation test between the signal to be processed and a reference signal. For
example, electrooculogram (EOG) signal can be used as the reference signal to
be compared with the EEG signal. The segment of the EEG signal that highly
correlates with the EOG is then assumed to be related to the eyeblink artifacts
and thus removed. However, since EOG also contains some EEG potentials due
to the close proximity of EOG electrodes to the frontal region of the brain, arti-
fact removal via regression methods may also remove important EEG data. In
addition, a reference electrode is obligatory in regression-based methods, which
may cause discomfort to patients when there is an extra pair of electrodes placed
around the eyes especially for longer EEG recordings.
In [8], Principal Component Analysis (PCA) is used to isolate out com-
ponents of the highest variance between EOG and EEG signals. The highest
variance components are the principal components and are classified as eyeblink
components. Similar to regression-based methods, PCA also requires additional
EOG electrodes to be applied. Besides that, it is not able to completely separate
some artifacts from the raw EEG signal in the event that both the eyeblink and
EEG signals have comparable amplitudes [9]. On the other hand, Independent
Component Analysis (ICA) [10, 11] is proven to be able to remove eyeblink
artifacts, as well as artifacts from different sources, but it may not be suit-
able for online applications as visual inspection on the independent components
(ICs) is required to manually identify and select ICs corresponding to artifacts
[12]. To overcome this, some work has been done to automate artifact detection
and removal by combining ICA with other methods like Wavelet or Empirical
Mode Decomposition. However, in all these cases the computational complexity
stands out as a limiting factor for ICA to be used in online applications [13].
The Wavelet transforms on the other hand depends on choosing a suitable de-
composition mother wavelet. The mother wavelet is a function comprising sine
and cosine waves, thus most of the time it will not characterize or adapt to
non-linear EEG signals, producing decomposition errors [12].

4
 Canonical Correlation Analysis (CCA) [14], has been used in muscle and
eyeblink artifact removal and has been proven to be the fastest among other
de-noising techniques discussed above [15]. However, if CCA is to be used alone
to remove artifacts, it still requires an additional reference signal to identify
the artifact events. Instead of using a reference signal, Empirical Mode Decom-
position (EMD) can be used to extract the eyeblink artifact signal from the
EEG signal [15–18]. EMD [19] is an algorithm that decomposes a signal with-
out requiring any pre-knowledge or pattern of interest, unlike other de-noising
techniques. In a comparative study on extracting out a biomedical signal [20],
EMD is proven to be more accurate compared to the wavelet transform. On
top of that, a method combining EMD and CCA (EMD-CCA) in [15], is shown
to outperform CCA, FastICA and EMD-FastICA in terms of artifact removal
accuracy, when evaluated on an EEG signal added with a synthetically gener-
ated eyeblink artifact. Despite the fact that it can accurately remove artifacts
from the EEG signal compared to other techniques, the algorithm is relatively
slow due to its iterative nature.
Most of the techniques on eyeblink artifact removal discussed above are used
only for offline artifact removal. Since applications such as BCI and epilepsy
monitoring require online signal processing, artifact removal methods and algo-
rithms should be capable of online processing. Hence, to cater to online artifact
removal, the methods or algorithm should satisfy a few criteria. The most im-
portant requirement is that the algorithm should be fully automatic without
any expert’s intervention. Secondly, online applications should avoid utilizing
additional electrodes around the artifact originating regions, such as EOG, as it
may cause discomfort and inconvenience to the subject during long-term EEG
recordings. Finally, online implementation requires the artifact removal algo-
rithm to have minimal computational complexity so that the algorithm doesn’t
introduce an unacceptable time delay.
Researchers have studied hybrid techniques to detect and remove eyeblink
artifact from EEG signal which may be useful for online applications [21–23].
Some of these techniques are discussed here. Lawhern et al. in [24] used the

5
Auto-regressive (AR) model for artifact feature selection followed by a Support
Vector Machine (SVM) classifier for training purposes to detect the artifacts.
Nguyen et al. [25], have reported their work on ocular artifact removal by
combining Wavelets and Artificial Neural Network (ANN), and naming their
technique Wavelet Neural Network (WNN). This technique requires an EOG
reference channel to train the ANN classifier. Zhao et al. [13] used Discrete
Wavelet Transformation (DWT) and an Adaptive Predictor Filter (APF) to
remove ocular artifacts from EEG signals. Daly et al. [26] have developed
a software plugin GUI, called the Fully Online and Automated Artifact Re-
moval for Brain-Computer Interfacing (FORCe). This plugin works based on
the combination of Wavelet Decomposition, Independent Component Analy-
sis and thresholding. FORCe runs in MATLAB and it is stated that it can
be used for online BCI applications, making it the only software plugin that
is able to perform significantly faster. Most recently, Tonachini et al. in [27]
has developed an online automatic artifact rejection using artifact subspace re-
construction (ASR), online recursive independent component analysis (ORICA)
and an IC classifier. However, the author has stated that ASR had negligible
effect on eyeblink artifact removal, and the time it took for ORICA to converge
well enough on the blink-related IC for the artifact to be removed is 26 seconds,
which is a significant amount of time.
To the extent of the authors’ knowledge, every online artifact removal tech-
nique discussed above depends on either a dedicated artifact reference recording
or some kind of training data that records artifacts separately for training pur-
poses, which may add some time delay to the techniques in online applications.
This work first focuses on introducing a novel unsupervised eyeblink artifact
detection algorithm which identifies eyeblink artifact regions effectively, assist-
ing subsequent artifact removal process. Secondly, the performance of EMD is
improved with various enhancements to resolve the processing time inefficiency
of the algorithm. Next, the enhanced version of EMD is applied on the most
relevant eyeblink artifact region identified through the unsupervised artifact de-
tection algorithm to extract out a suitable eyeblink artifact template. Finally,

6
our work makes use of the artifact template extracted as a reference in iden-
tifying subsequent eyeblink artifacts instead of relying on an EOG recording.
With the help of the artifact template, the identified eyeblink artifact regions
are subjected to CCA for eyeblink artifact removal in online applications. The
direction of the work is to provide an application-centric solution for online ap-
plications with reasonable/reduced complexity and enhanced performance. The
developed algorithm neither depends on a separate EOG recording or an ex-
pert’s advice for eyeblink events identification, thus removing any constraints in
terms of automation for online implementation. Additionally, no training data
is required beforehand for the algorithm to learn and identify eyeblink artifacts.
The developed algorithm is compared with one of the state-of-the-art methods,
i.e. FORCe, due to its effectiveness in removing eyeblink artifacts and its low
computation time. The next section elaborates the proposed algorithm and the
materials used in this work, while results and discussions are presented in Sec-
tion 3. Finally, we conclude the paper with some recommendations in Section
4.

2. Materials and Methods

2.1. Unsupervised Eyeblink Artifact Region Detection

Since the frontal region of the brain is the nearest region to the position of
eyes, eyeblink artifacts can be easily captured in this region, so the Fp1 and
Fp2 electrodes should hypothetically exhibit high correlation whenever there is
an occurrence of an eyeblink. To validate this theory, the correlation coefficient
is computed between Fp1 and Fp2 in windows of 500 samples (1.95 seconds).
As eyeblink artifacts can last up to 800ms [2, 3], this window size will allow
at least one eyeblink artifact to fall within the window. The test has revealed
that segments of Fp1 and Fp2 without eyeblink artifact produce correlation
below than 0.7, whereas segments containing eyeblink artifact results in higher
correlation, usually more than 0.9 as illustrated in [28] and shown in Fig. 2.

7
 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. 2: Correlation Coefficient, CC, between Fp1 and Fp2 Electrodes

In order to design an automatic eyeblink artifact region detection algorithm,

first a high value of correlation coefficient that is greater than 0.85 is set to
indicate the presence of an eyeblink artifact in that particular window. However,
the eyeblink potentials and starting point of the eyeblink artifact should be
identified for subsequent analysis or artifact removal, which requires a threshold.
The displacement of amplitude is chosen as the threshold criterion as eyeblink
artifacts are in general, higher in amplitude relative to that of the EEG or brain
signal. Therefore, the eyeblink artifact components are expected to produce
higher amplitude displacement compared to uncontaminated EEG potentials.
First, the amplitude displacement from the mean is calculated within an Fp1
window that exhibits a high correlation with Fp2. The displacement distribution
from the mean amplitude is computed using Eq. (1):

Displacement[t] = |X[t] − µ| (1)

where, X[t] is the EEG signal’s amplitude at time t, and for any given window
starting at sample point n, X[t] is evaluated from t = n to t = n + 500, and µ

8
is the mean of that particular window.
An experiment conducted by the authors in [28] has shown that the thresh-
old for eyeblink artifact’s onset point and eyeblink potentials dominating the
EEG window in question can be correctly determined by taking two standard
deviation, 2σ width from the mean of the displacement distribution acquired,
as in Eq. (2). Any absolute value beyond 2σ is classified as an eyeblink artifact
potential and the first sample that exceeds this threshold is considered as the
eyeblink artifact’s starting point.

threshold = mean + 2σ (2)

Later, the onset of eyeblink artifact is moved 100 samples (0.39 seconds)
ahead. The reason for setting the onset point in advanced of 100 samples before
the threshold is to provide a buffer for any subsequent analysis. The end point
of the eyeblink artifact is then set to 256 samples, or 1 second, after the first
sample with an amplitude displacement crossing the threshold. The eyeblink
artifact region is therefore taken to be from the onset of eyeblink till the end
point of the eyeblink. Thus, an eyeblink which can last up to 0.8 seconds (205
samples) in duration completely fit into this window (100+256=356 samples).
Several eyeblink artifact regions are searched and saved in a similar way un-
til any two eyeblink artifact regions exhibit correlation coefficient of more that
0.9 between them. The correlation coefficient value of more than 0.9 is chosen
assuming that a high correlation between the eyeblink artifact regions denotes
repetitiveness or similarity in the blinking pattern of an individual. Hence these
regions with high similarity or correlation will be subjected for further anal-
ysis, which is the EMD algorithm in this research work. Fig. 3 summarizes
the algorithm in a flowchart. Fig. 4 shows the plot of a real EEG signal’s eye-
blink artifact regions identified through the proposed eyeblink artifact detection
algorithm located on the Fp1 channel.

9
 Start

Window size = 500,

EB end = 1st sample of x(n)

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

Find Correlation Coefficient (CC) between EB end = last point

Fp1 & Fp2 window of window samples

NO
CC > 0.85

YES

Compute Amplitude Displacement

Threshold = ( mean + 2 σ ) of Displacement
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

Find Correlation Coefficient (CC) between

current EB region with previous EB regions

NO
CC > 0.8 or 0.9

YES

Apply EMD to these 2 EB regions,

extract out EB template

End

Fig. 3: Flowchart of the Automated Eyeblink Region Identification Algorithm

2.2. Eyeblink Artifact Template Extraction Through Enhanced EMD (FastEMD)

2.2.1. Introduction to Empirical Mode Decomposition (EMD)

Empirical Mode Decomposition (EMD) is an algorithm that decomposes
a signal into multiple oscillating components. The algorithm reiterates itself
until it can isolate the highest oscillating component that remains in a signal.
This is achieved by identifying relative extrema (maximum/minimum) points
in a signal, followed by forming upper/lower envelopes by interpolating these
points and removing the mean of the envelopes from the signal. This process
is called ”sifting”, where it continually sifts out a local high oscillating trend

10
 EB Artifact Regions EEG signal
60
EB Artifact Region

40

20

Amplitude(microvolts)
0

-20

-40

-60

-80
0 1000 2000 3000 4000 5000 6000
Samples

Fig. 4: Eyeblink Artifact Regions Identified through Proposed Method

called the Intrinsic Mode Functions (IMF). Each IMF extracted out from the
original signal is a lower oscillating trend compared to its predecessor. Adding
up all IMFs and the remaining residual signal obtained from the decomposition
would reconstruct the original signal. Each IMF should satisfy the following
criteria as in [19]:

• contains an equal number of extrema and zero crossings, or differ at most

by one

• envelopes of the IMF are symmetric with respect to zero

In general, X(t) is decomposed into multiple oscillating components called

IMFs, xi (t) and a residual component, Rn (t) which is monotonous, as in Eq.
(3):

n−1
X
X(t) = xi (t) + Rn (t) (3)
i=1

Each sifting loop produces the i-th IMF of the algorithm, xi (t). The recursive
sifting discontinues after the algorithm extracts out n − 1 IMFs, the instance

11
where the residual signal, Rn (t) becomes a monotonic trend. The algorithm is
relatively slow because it reiterates itself until the final residual signal becomes
a monotonic function.

2.2.2. Proposed Approaches: Enhanced Empirical Mode Decomposition (FastEMD)

EMD is enhanced to resolve the processing time inefficiency of the algorithm
through approaches discussed in the following subsections.

Envelope Interpolation in EMD

The performance of EMD algorithm through alternative interpolation tech-
niques is discussed in this section. A major concern in EMD’s sifting process
relies on how the upper and lower envelopes are being constructed through in-
terpolation. In online applications, EMD could cause the overall processing time
to increase as the algorithm is iterative and dependent on interpolating large
number of extrema. The interpolation involved would directly consume a lot of
the computer resources, hence EMD can be inefficient while removing eyeblink
artifacts from lengthy EEG signals, especially in online processing.
As to enhance the performance of EMD algorithm, other interpolation tech-
niques were tested and evaluated in another work. Among alternative interpola-
tion techniques investigated are the Cubic Hermite Spline Interpolation (CHSI)
and the Akima Spline Interpolation (ASI). These two interpolation techniques
were investigated in terms of their ability to retain the reconstruction accuracy
after decomposition and their speed compared to Cubic spline interpolation
(CSI), that is used in the classical EMD algorithm. The ASI has produced
the highest correlation coefficient of 0.9063, lowest Root Mean Square Error
(RMSE) of 3.3, lowest percentage root means square difference (PRD) of 44%,
better Signal to Noise Ratio (SNR) of 8.5dB and faster computation time of
0.24s, in decomposing an artificial EEG signal compared to CSI. These results
justify that the ASI technique serves a lower computational burden to EMD
algorithm with higher reconstruction accuracy and shorter computation time as
shown in [29]. Envelope construction through CSI fulfil second-order derivation

12
at every extremum point to ensure continuity and spline curvature smoothness.
Since envelope construction through CSI force two adjacent splines to be con-
tinuous at first and second derivatives, the formed envelopes are susceptible
to overshoots and undershoots. This produces an erroneous mean estimation
during sifting and this error could eventually get transferred and added to the
whole data set on every iteration of EMD’s sifting process, resulting in an in-
accurate and unreliable decomposition. While the envelope construction of ASI
depends only on the slopes of adjacent segments with continuity up to first
order derivative. Although ASI produces envelopes that are not as smooth as
the CSI does, but it demonstrates a better decomposition accuracy. This also
reduces the necessity to solve large system equations which in turn, reduces the
computation time.

Fixed Number of IMFs

Another factor that limits EMD in online applications is the repetitive sift-
ing process required in obtaining the IMFs. Sifting in EMD algorithm can be
classified as redundant in two aspects. First, the algorithm has to repeat sifting
plenty of times before any of the resulting trend satisfies the IMF criteria, and
thus can be classified as an IMF. Secondly, the algorithm has to reiterate itself
multiple times to attain multiple numbers of such IMFs, because it can’t termi-
nate sifting until the residual signal becomes a monotonous function. Therefore,
IMF extraction through repetitive sifting iterations causes EMD algorithm to
be computationally inefficient and slow.
To overcome this issue, the number of IMFs extracted out through EMD
is fixed to a constant number. The higher oscillations in the raw EEG signal
will be isolated out in the first or second IMFs, hence the sum of remaining
IMFs would by default produce an eyeblink artifact trend. Hence, by partially
reconstructing the higher oscillating trends which are lower in amplitude would
yield the EEG trend. Alternatively, low oscillating trends with high amplitudes
are summed together to attain the eyeblink artifact trend. Therefore, EMD’s
algorithm is re-designed to decompose the raw EEG signal up to 5 to 8 IMFs,

13
which is sufficient to segregate out the EEG trend and the eyeblink trend.
2
X
XEEG (t) = xi (t) (4)
i=1
5
X
Xeyeblink (t) = xi (t) + R6 (t) (5)
i=3

This automatically reduces the computation time and the algorithm does not
have to repeat itself until a monotonic residue is acquired. The eyeblink artifact
template obtained by adding up the 3rd IMF onwards with the residual signal
is shown in Fig. 5.

EB Artifact Template
60

40
Amplitude(microvolts)

20

-20

-40

-60
0 50 100 150 200 250 300 350 400
Samples

Fig. 5: Extracted Eyeblink Artifact Template

Optimal Use of EMD in Eyeblink Artifact Template Extraction

In applications that require online monitoring of EEG signals, the applica-
tions could not wait until the entire EEG signal is recorded for analysis, it may
take from few hours to days for an EEG recording to be completed. The eyeblink
artifacts have to be removed instantaneously from the EEG recording as well.
EMD algorithm is not recommended to be applied after the recording is com-
plete, as it may cause a delay in interpreting an EEG signal and applying EMD
on the entire EEG recording will cause the application to get computationally

14
heavy. As an option, EMD can be applied repetitively on short segments, when-
ever an eyeblink artifact event is captured, provided the occurrence of eyeblinks
are known. Unfortunately, EMD gets computationally inefficient and slow on
repetitive application to a huge dataset especially during online recording and
analysis, which may even disrupt the recording task.
To resolve this, two eyeblink regions identified in section 2.1 with cross-
correlation of more than 0.9, indicated with boxes in Fig. 6 are subjected to
EMD separately. EMD is applied only on two most correlating eyeblink artifact
regions, thus keeping the number of EMD applications lowest as possible. These
two eyeblink regions are chosen as these regions are repetitive in terms of the
blinking pattern, which can be assumed as a general eye blinking pattern for that
particular EEG signal. This prevents EMD to be used repetitively, especially
when the EEG signal is processed in an online manner. This method is differ-
ent compared to what is being practised in classical artifact removal technique
through EMD, where EMD will be applied to remove the artifacts whenever an
artifact event is identified. The low oscillating IMFs obtained through EMD are
then added, as in Eq. (5), and averaged out to get an eyeblink artifact template.

EB Artifact Regions EEG signal

60
EB Artifact Region

40

20
Amplitude(microvolts)

-20

-40

-60

-80
0 1000 2000 3000 4000 5000 6000
Samples

Fig. 6: Highly Correlating Eyeblink Artifact Regions Subjected to EMD

15
Stoping Criterion for EMD
The work also adopts a stopping criterion for EMD based on the standard
deviation which was introduced in [19]. The standard deviation (SD) is defined
as the normalized squared difference between two sifting iterations, which is
assumed to indicate consistency between two sifting outputs. The SD value
calculated from two consecutive sifting outputs, yj (t) and yj−1 (t) should be less
than a pre-determined value, normally 0.2 or 0.3 to stop the sifting iteration in
EMD.

k
" #
X |yj−1 (t) − yj (t)|2
SD = 2 (t) ≤ 0.2 (6)
t=0
yj−1
where k is the number of samples in the original signal, X(t).
An updated flowchart of EMD with enhancements discussed above is shown
in Fig. 7.

2.3. Eyeblink Artifact Removal

2.3.1. Introduction to Canonical Correlation Analysis (CCA)
Canonical Correlation Analysis (CCA), utilizes blind source separation (BSS)
technique. As suggested by the name, BBS separates a set of source signals from
a set of mixed signals without any priori knowledge about the source signals or
the weighted mixing components. The linear relationship between two mul-
tidimensional variables is measured. CCA extracts out two sets of source vec-
tors, where the projections of the two multidimensional variables onto extracted
source vectors are maximally correlated.
The application of CCA in the EEG signal is elaborated in this section. The
observed EEG signal, x(t) is classified as the first multidimensional variable,
while the second multidimensional variable is obtained by taking a temporally
delayed component of the original observed EEG signal, y(t) = x(t − 1). As
the BBS implies the observed EEG signal x(t) is a result of weighted mixing W
with clean EEG source signal, S(t):

x(t) = W S(t) (7)

16
 START

i=1:5, j=1
Original signal = X(t) = R0(t)
Input signal = z(t)
Start: z(t) = X(t) = y0(t)

Construct Upper & Lower Envelopes through Akima

spline interpolation,
calculate mean, m(t)

Treat y(t) as input

z(t) = yj(t) Subtract m(t) from input signal:
yj(t) = z(t) - m(t)

Calculate SD between Treat residue, Rn(t) as input

yj(t) and yj-1(t) z(t) = Rn(t)

Is SD < 0.2?
NO

YES

Store IMF as xi(t) = yj(t)

Subtract the IMF from input signal
Rn(t) =Ri-1(t) - xi(t)

NO
Is i > = 5?

YES

END

Fig. 7: Flowchart of Enhanced EMD Algorithm, FastEMD

Hence, the clean EEG sources of Sx (t) and Sy (t) can be estimated by taking

17
the weighted de-mixing matrix, A onto the observed EEG signals:

A = W −1 (8)

Sx (t) = Ax(t) (9)

Sy (t) = By(t) (10)

The source signals, which are considered as the canonical variates u(t) and
v(t) are obtained through linear combinations between the de-mixing matrices
and mean removed observed EEG variables, x̂(t) and ŷ(t), where n is the number
of EEG sample in one channel and p is the number of channels of the EEG
recording:

u1 = a11 xˆ1 + a12 xˆ2 + ... + a1p xˆp

u2 = a21 xˆ1 + a22 xˆ2 + ... + a2p xˆp

un = an1 xˆ1 + an2 xˆ2 + ... + anp xˆp

(11)
v1 = b11 yˆ1 + b12 yˆ2 + ... + b1p yˆp

v2 = b21 yˆ1 + b22 yˆ2 + ... + b2p yˆp

vn = bn1 yˆ1 + bn2 yˆ2 + ... + bnp yˆp

which can be simplified as:

U = AT X̂
(12)
V = B T Ŷ

The purpose of CCA is finding the de-mixing matrices A and B such that the
correlation, ρ between U and V , is maximized, or as large as possible. For exam-
ple, the de-mixing matrices a1 = [a11 , a12 , ..., a1p ]T and b1 = [b11 , b12 , ..., b1p ]T
are computed such that the coefficient of canonical correlation between the first
pair of canonical variates u1 and v1 is maximized:

ρ1 = corr(u1 , v1 ) (13)

where,
u1 = a1 T x̂
(14)
v1 = b1 T ŷ

18
 The second and following pairs of canonical variates are computed in a sim-
ilar way, provided that the second pair of canonical variates are uncorrelated
with the first pair and other pairs of canonical variates. This procedure is re-
peated until enough canonical variate pairs are obtained. CCA was initially
proposed in [30] by Hotelling. In EEG’s artifact removal, CCA was employed
in several works to remove muscle and ocular artifacts. CCA is implemented by
De Clercq et al. [14] to remove muscle artifacts from the EEG signal, followed
by Hallez et al. in [31] with CCA and the blind source separation approach.
Later, Zhao et al. [32] used the Wavelet in combination with CCA to remove
ocular artifacts from EEG. Sweeney et al. [18] then use the Ensemble EMD with
CCA to remove artifact from the EEG signal. On the other hand, M.Soomro
et al. [15] has used the CCA to the entire signal with conventional EMD for
removal of eyeblink artifacts in a short length of EEG signal. In this work, CCA
is applied in windows to obtain canonical components and used with the com-
bination of enhanced EMD for eyeblink artifact removal from real EEG signals
of long durations.

2.3.2. Application of Windowed CCA for Eyeblink Artifact Removal

A sliding window with the length of the eyeblink artifact template is moved
along the EEG signal and each window is cross-correlated with the eyeblink
artifact template extracted, as in Eq. (15):
PN
t=1 X(t)XEB (t)
ρ(X, XEB ) = qP qP (15)
N 2 (t) N 2
t=1 X t=1 XEB (t)

where N represents the length of the window, X(t) represents the contaminated
EEG signal, and XEB (t) represents the eyeblink artifact template extracted
from FastEMD.
Windows that exhibit high similarity scores with the eyeblink artifact tem-
plate are subjected to CCA. CCA estimates the canonical components that
maximize temporal correlation within the specified window. The most perti-
nent artifactual canonical components, U , usually the first row of the canonical
components are forced to become zero in order for it to behave non-artifactual.

19
The artifact-free canonical components are termed as Uclean . Then, clean EEG
segment is reconstructed by taking the inverse of the de-mixing matrix, Ax into
the non-artifactual source, Uclean :

x(t)clean = A−1 Uclean (t) (16)

Fig. 8 depicts the overall block diagram of the proposed method:

IMF 1

EEG Trend

IMF 2

Enhanced IMF 3
EB artifact
Raw EEG Empirical
region
Signal Mode
identification
Decomposition
IMF 4
EB artifact
template
IMF 5

RESIDUAL
SiGNAL

Cross
Raw EEG NO
Correlation > 0.5
Segments
?

YES

Remove EB Canonical
artifact canonical Correlation
components Analysis (CCA)

CLEAN EEG
SIGNAL
EEG signal Unsupervised EB
acquisition artifact region detection EB artifact template extraction EB Artifact removal in windows / segments

Fig. 8: Overall Block Diagram of the Proposed Algorithm, FastEMD-CCA

2.4. EEG Recording/Simulation and Analysis

2.4.1. Synthetic Signals
Synthetic eyeblink and EEG signals are simulated for validation purpose in
MATLAB 2018b. Synthetic eyeblink artifacts, Z(t) can be simulated through
exponential functions with different amplitudes as stated in [29]:
2 2 2 2
Z(t) = 10e−(10t−10) + 10e−(10t−30) + 8e−(10t−45) + 7e−(10t−70) (17)

20
 On the other hand, a synthetic EEG signal can be generated through pink
noise, Y (t) for a duration of 10 seconds, 2560 samples at a sampling frequency of
256 Hz. EEG and eyeblink artifact models simulated through pink noise and ex-
ponential function are shown in Fig. 9(a) and 9(b) respectively. Both synthetic
EEG signal and eyeblink artifact are mixed to acquire a set of synthetically
contaminated EEG signal, X(t) as in Fig. 9(c).

X(t) = Z(t) + Y (t) (18)

(a) Simulated EEG Signal

20

-20
0 1 2 3 4 5 6 7 8 9 10
(b) Simulated EB Artifact
20
Amplitude(uV)

-20
0 1 2 3 4 5 6 7 8 9 10
(c) Contaminated EEG Signal
20

-20
0 1 2 3 4 5 6 7 8 9 10
Time(s)

Fig. 9: (a) Synthetic EEG Signal, (b) Synthetic Eyeblink Artifact, (c) Contaminated EEG
Signal

2.4.2. Real Signals

EEG data sets that were used in this paper were collected at Hitachi, Hatay-
oma site in Japan. EEG data from volunteers were obtained according to the
regulations of the internal review board on Central Research Laboratory, Hi-
tachi, Ltd., following receipt of written informed consent. The approval number
is 20131021-0138. These EEG signals have been primarily collected to conduct a
study on mental stress. Since all recorded signals were contaminated by eyeblink
artifacts, we have re-use these data sets to achieve the goal of this work. These

21
EEG signals are recorded following the 10-20 international standardization with
free electrodes placed on the scalp. The EEG signals were collected from 10 par-
ticipants with 6 recordings from each participant, resulting in 60 EEG data sets.
The participants are aged between 30 and 55 years. All recorded signals are of
different durations, which were recorded at a sampling rate of 256 Hz.

2.5. Performance Evaluation

Any artifact removal algorithm is reflected as efficacious and successful de-

pending on two measures. The first and most important one is how well an
algorithm is able to remove the artifacts, and the second one is how well an
artifact removal algorithm is able to preserve neural information contained in
an EEG signal after artifact removal. On another note, the online eyeblink arti-
fact removal capability can be interpreted through processing time taken by the
algorithm. This is to evaluate the ability of the algorithm in pursuance of online
processing, whether it can achieve instantaneous artifact removal without loss
of neural information.
However, evaluating the performance of any algorithms in identifying and
discarding artifacts is challenging in the absence of ground truths. Hence, the
eyeblink artifacts and EEG signals are artificially generated as discussed in
section 2.4.1. These artificial signals serve as ground truths in carrying out the
performance evaluation, before applying the enhanced algorithm to real sets of
EEG signals. In real EEG signal recordings, the EOG electrodes that capture
eyeblinks are not recorded for convenience purposes. Additionally, there are no
training data with blinking recorded so that the algorithm is fully automatic.
Since EOG is not recorded, validation, if the eyeblink artifacts are removed
turns out to be difficult. Thus, an expert’s advice is sought to substantiate
if the algorithm is able to remove eyeblink artifacts effectively through visual
manual inspection (VMI).

22
2.5.1. Compared Approaches and Evaluation Criteria
The proposed algorithm, FastEMD-CCA is compared with two existing tech-
niques or algorithms. Evaluation on the approaches are performed in MATLAB
2018b on Windows 7 Professional(64-bit OS, 4GB RAM).

FastEMD-CCA and Wavelet Transform

The developed algorithm, FastEMD-CCA is compared with Wavelet Trans-
form, to evaluate the performance exhibited by these algorithms on the syn-
thetically contaminated EEG signal, in Eq. (18). The synthetic EEG signal
and the eyeblink artifacts are simulated for 100 trials for reliability purposes,
and the results are averaged. Wavelet is chosen as it has been extensively used
for eyeblink artifact removal in EEG [12, 33, 34]. The FastEMD-CCA works as
discussed in sections 2.1, 2.2 and 2.3. For Wavelet Transform, the sym9 mother
wavelet from the Symlets family is chosen as it was suggested as resembling EEG
signals the most and would be the most compatible one for de-noising purposes
by Al-Qazzaz et al. [35]. SWT is applied with soft thresholding on the entire
contaminated EEG signal to obtain wavelet coefficients. Larger coefficients are
assumed to correspond to the artifact and smaller coefficients are assumed to
correspond to EEG. The inverse of SWT, ISWT is then applied on the coeffi-
cients corresponding to EEG and artifact to reconstruct the clean EEG signal
and the eyeblink artifact respectively.
The performance of these algorithms in retaining the neural information
in an EEG signal is quantitatively assessed. Reconstructed EEG signals after
artifacts have been removed via these algorithms are validated against synthet-
ically generated EEG signals as ground truths. Ideally, reconstructed EEG
signals should remain intact after artifacts have been removed. The algorithms
are evaluated in terms of correlations coefficient (CC), root means square er-
ror (RMSE) and signal to noise ratio (SNR), in the time domain. Each of the
performance criteria is expressed as confidence intervals for 95% of confidence
level. CCeeg measures the similarity between synthetically generated EEG sig-
nals with its corresponding reconstructed EEG signals after artifact correction,

23
while CCeb estimates the resemblance of removed eyeblink artifacts compared
to synthetic eyeblink artifacts. RMSE measures the removal and reconstruction
error for eyeblink and EEG signals respectively. The RMSE is calculated by
finding the difference between synthetically generated eyeblink artifacts with
removed eyeblink artifacts, RM SEeb and synthetically generated EEG signals
with reconstructed signals, RM SEeeg after processing with the suggested tech-
niques. The SNR is used in this analysis to determine the ratio of signal to
artifact that remains after eyeblink artifact has been removed from the con-
taminated EEG signal. The SNR ratio is calculated before and after eyeblink
artifact removal, using Eq. (23) and (24).

cov(Y, Y1 )
CCeeg = (19)
std(Y ) ∗ std(Y1 )

cov(Z, Z1 )
CCeb = (20)
std(Z) ∗ std(Z1 )

r Pn
t=1 (Y (t) − Y1 (t))2
RMSEeeg = (21)
n

r Pn
t=1 (Z(t) − Z1 (t))2
RMSEeb = (22)
n

 
std(Y )
SNRbefore = 10 log (23)
std(Y − X)

 
std(Y )
SNRafter = 10 log (24)
std(Y − Y1 )

where X(t) represents the synthetically contaminated EEG signals, Y (t) refers
to the simulated/synthetic EEG signals generated using pink noise, Y1 (t) cor-
responds to the reconstructed EEG signals which are free from artifacts, Z(t)
refers to the synthetic eyeblink artifact and Z1 (t) corresponds to the extracted

24
eyeblink artifact. From the performance metrics, 95% of confidence interval
has been estimated so that the probability of the performance is repetitive over
95% of the time, if the evaluation to be repeated multiple times in another time
frame.

FastEMD-CCA and FORCe

Real EEG signals illustrated in section 2.4.2 were used to evaluate the pro-
posed, FastEMD-CCA and the state-of-the-art, FORCe algorithms. The out-
comes are then averaged for these 60 sets of EEG signals. In MATLAB, the
EEG recordings are imported into the workspace and processed automatically
to remove the eyeblink artifacts in windows, with each window is minimally
1s in length for both FastEMD-CCA and FORCe. The FastEMD-CCA algo-
rithm operates as discussed in sections 2.1, 2.2 and 2.3. The FORCe algorithm
first applies wavelet decomposition on each channel of an EEG signal. Result-
ing approximation coefficients attained through wavelet are subjected to ICA
to get independent components, ICs. Next, the artifactual ICs are identified
through several threshold criteria, where ICs exceeding certain threshold values
are classified as eyeblink and electrocardiogram artifacts respectively, and thus
removed. The inverse of ICA is performed to estimate a set of cleaned approx-
imation coefficients. Then, soft thresholding is applied to resulting approxima-
tion coefficients from ICA and detail coefficients acquired through wavelet to
suppress/remove EMG artifacts. Finally, a clean EEG signal is reconstructed.
Since the ground truth are not available for real EEG data sets, the efficacy
of both algorithms in identifying and removing eyeblink artifacts, while pre-
serving the artifact-free EEG segments are verified with the help of an expert,
Neuroscientist Dr. Tahamina, through VMI. The evaluation criteria are derived
from various measures of the binary prediction [36, 37], thus determining the
accuracy, sensitivity, specificity and error rate of the algorithms.

25
3. Results and Discussions

3.1. Evaluation on Synthetic Signals

The proposed algorithm, FastEMD-CCA is compared with Wavelet Trans-

form as elaborated in section 2.4.1, to evaluate the performance exhibited by
these algorithms on synthetically generated EEG signals. Fig. 10(a) - 10(c)
show the reconstructed EEG signal and eyeblink artifact through FastEMD-
CCA. Fig. 11(a) - 11(c) show the reconstructed EEG signal and eyeblink artifact
through SWT.

De-noising via FastEMD-CCA

(a) Contaminated EEG
20

-20
0 1 2 3 4 5 6 7 8 9 10
(b) Reconstructed EEG via FastEMD-CCA
5
Amplitude(uV)

Reconstructed EEG Signal via FastEMD-CCA

-5
0 1 2 3 4 Simulated
5 EEG
6 Signal7 8 9 10
(c) Extracted EB Artifact
20

-20 Extracted EB Artifact via FastEMD-CCA

0 1 2 3 4 5 Simulated
6 EB Artifact
7 8 9 10
Time(s)

Fig. 10: (a) Mixed EEG and Eyeblink Signal, (b) Reconstructed EEG Signal, (c) Extracted
Eyeblink Artifact

The performance metrics obtained for FastEMD-CCA and Wavelet Trans-

form, applied on synthetically generated and contaminated EEG signals are
tabulated in Table 1.

3.1.1. Discussion
The algorithms are evaluated on 100 trials of synthetically contaminated
signals to ensure the performance exhibited by the algorithms are reliable and
repetitive. The confidence interval for 95% of confidence level is determined for

26
 De-noising via SWT
(a) Contaminated EEG
20

-20
0 1 2 3 4 5 6 7 8 9 10
(b) Reconstructed EEG via SWT
Amplitude(uV) 5

Reconstructed EEG Signal via SWT

-5
0 1 2 3 4 5 Simulated
6 EEG
7 Signal 8 9 10

(c) Extracted EB Artifact

20

-20 Extracted EB Artifact via SWT

0 1 2 3 4 5 Simulated7 EB Artifact
6 8 9 10
Time(s)

Fig. 11: (a) Mixed EEG and Eyeblink Signal, (b) Reconstructed EEG Signal, (c) Extracted
Eyeblink Artifact

Table 1: Performance Metrics of Synthetic Signals

Wavelet (SWT) FastEMD-CCA
Techniques
mean ± std (µ ± σ) 95% CI mean ± std (µ ± σ) 95% CI
CCeeg 0.6479 ± 0.0332 0.6413 to 0.6545 0.7478 ± 0.0687 0.7341 to 0.7614
CCeb 0.9272 ± 0.0065 0.9259 to 0.9285 0.9754 ± 0.0055 0.9743 to 0.9765
RMSEeeg 0.7641 ± 0.0272 0.7587 to 0.7695 0.6580 ± 0.0776 0.6426 to 0.6734
RMSEeb 1.9562 ± 0.0894 1.9385 to 1.9740 0.6580 ± 0.0776 0.6426 to 0.6734
SNRafter (dB) 2.6967 ± 0.3560 2.6260 to 2.7673 4.2845 ± 1.2200 4.0424 to 4.5265
SNRbefore (dB) -10.6594 ± 0.00 -10.6594 -10.6594 ± 0.00 -10.6594
Time (s) 0.0226 ± 0.0065 0.0213 to 0.0239 3.2489 ± 3.9818 2.4584 to 4.0386

each of the performance metrics. The 95% confidence level is chosen so that the
estimation of results are statistically sound. CC value normally lies between -1
and 1, in which a value approaching 1 indicates a higher correlation or simi-
larity. RMSE value that approaches zero signifies a more precise and accurate
signal reconstruction, relative to the synthetic signals. The SNR measures the
scale of eyeblink artifacts that have been removed from the noisy EEG signal
and the degree of neural signal preservation. The effectiveness of the evaluated
algorithms in preserving the underlying neural information in an EEG signal

27
can be deduced through CC value that approaches near 1, RMSE close to 0 and
higher SNR value. In this analysis between Wavelet and the proposed technique,
FastEMD-CCA has produced higher CC values on average compared to SWT,
0.7478 in reconstructing the EEG signal and 0.9754 in extracting out the eye-
blink artifact. The error produced by FastEMD-CCA is 14% percent lower than
the error produced by SWT in reconstructing the EEG signal. While in extract-
ing out the eyeblink artifact, FastEMD-CCA has produced an error of 66% lower
than SWT. This indicates that the FastEMD-CCA algorithm is able to remove
eyeblink artifact components appropriately from the contaminated EEG signal
in comparison with SWT. From Table 1, FastEMD-CCA yields very high SNR,
close to 4 dB on average from -10dB before artifact correction, which denotes
a higher ratio of neural information has been preserved. Alternatively, SWT
produced nearly 2dB of SNR on average from -10dB before artifact elimination.
This shows that the FastEMD-CCA is a better choice in removing eyeblink ar-
tifacts, and at the same time, it is able to preserve underlying EEG components
better, by not introducing much distortion to the neural signal. In terms of
computation time, the SWT is way faster than the FastEMD-CCA. It has to be
emphasized here that SWT removes artifacts only from a single channel EEG
signal, hence faster computation time, while FastEMD-CCA performs the ar-
tifact elimination from a multichannel EEG signal. Moreover, SWT is applied
to the entire signal for processing which is not applicable for online applica-
tions, while the FastEMD-CCA algorithm process the EEG signals in windows.
SWT also relies on manual selection of appropriate mother wavelet, comprises
sine and cosine functions, which may not represent a basis function for non-
stationary biomedical signals. Selecting an inappropriate mother wavelet could
lead to inaccuracy in reconstructing artifact-free EEG signals. Furthermore, the
accuracy of SWT is also sensitive to the selection of thresholding function which
could have an effect on preserving or discarding the neural information in an
EEG signal. Considering the performance shown by FastEMD-CCA by means
of accuracy in removing artifacts, it’s used for evaluation in removing artifacts
in real EEG signals.

28
3.2. Evaluation on Real EEG Signals
Results in Table 2 were obtained through offline analysis performed on the
artifact removed EEG signals in an online manner through the proposed tech-
nique, FastEMD-CCA and the state-of-the-art algorithm, FORCe. Fig. 12 and
13 shows an example of an entire EEG signal, reconstructed using FORCe algo-
rithm and the proposed algorithm, FastEMD-CCA respectively. Fig. 14 and 15
show a short portion of the EEG signal, reconstructed using FORCe algorithm
and the proposed algorithm, FastEMD-CCA respectively.

(a)-Observed EEG Signal

100

50

-50
Amplitude(microvolts)

-100
0 1 2 3 4 5 6 7 8
104
(b)-Artifact Free EEG Signal
100

50

-50

-100
0 1 2 3 4 5 6 7 8
Samples 104

Fig. 12: Entire EEG Signal-Reconstructed through FORCe

The average of accuracy, sensitivity, specificity, error rate and computation

time for FastEMD-CCA and FORCe are tabulated in Table 2. The represen-
tation of error, accuracy, sensitivity, specificity and computation time for both
FORCe and FastEMD-CCA are shown in Fig. 16.

3.2.1. Discussion
Accuracy is a measurement of correct detection of eyeblink artifacts by the
algorithms, thus removing them, and also how well the algorithms could retain
the artifact-free EEG segments after artifact correction is performed. The pro-
posed algorithm has achieved an average of 97.9% accuracy compared to 91.7%

29
 (a) Observed EEG Signal
100

50

Amplitude(microvolts)
-50

-100
0 1 2 3 4 5 6 7 8
104
(b) Artifact Free EEG Signal
100

50

-50

-100
0 1 2 3 4 5 6 7 8
Samples 104

Fig. 13: Entire EEG Signal-Reconstructed through FastEMD-CCA

(a)-Observed EEG Signal

100

50

-50
Amplitude(microvolts)

-100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

(b)-Artifact Free EEG Signal

100

50

-50

-100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Samples

Fig. 14: A Portion of the EEG Signal-Reconstructed through FORCe

by FORCe. The higher accuracy level of the proposed algorithm, FastED-

CCA over FORCe by 6.2% clearly reflects the effectiveness of the algorithm in
correctly identifying and removing eyeblink artifacts from EEG signals in on-
line applications. Contrarily, the error rate is the exact opposite of accuracy.
FastEMD-CCA produced an average error rate of 2.10% while FORCe yields

30
 (a) Observed EEG Signal
100

50

-50
Amplitude(microvolts)
-100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

(b) Artifact Free EEG Signal

100

50

-50

-100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Samples

Fig. 15: A Portion of the EEG Signal-Reconstructed through FastEMD-CCA

Table 2: Performance Metrics of Real EEG Signals

Techniques FORCe FastEMD-CCA
Accuracy 91.70 % 97.90 %
Sensitivity 89.47 % 97.65 %
Specificity 98.65 % 99.22 %
Error 8.30 % 2.10 %
Time 85.10 s 19.73 s

8.30%. This denotes that both algorithms are still susceptible to miss out an
eyeblink artifact, however, the proposed algorithm is more reliable in detect-
ing and removing eyeblink artifacts in online applications compared to FORCe.
Sensitivity, on the other hand, is a measurement of how sensitive the algorithms
are in detecting and removing the eyeblink artifacts in comparison with the
actual number of observed eyeblink artifacts. The results indicate the proposed
algorithm, FastEMD-CCA has achieved 97.65% of sensitivity, 8.18% higher than
that of the FORCe algorithm. This shows that FastEMD-CCA could identify
and remove eyeblink artifacts relatively better than FORCe could. The sen-
sitivity of FORCe in identifying and removing the artifacts is 89.47 %. This

31
 Performance Representation
120.00%
100.00%
Percentage

80.00%
60.00%
40.00%
20.00%
0.00%
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58
EEG Datasets

FORCe Error Rate FORCe Accuracy FORCe Sensitivity FORCe Specificity

FastEMDCCA Error Rate FastEMDCCA Accuracy FastEMDCCA Sensitivity FastEMDCCA Specificity

Computation Time (s)

FORCe FastEMDCCA
120.00

100.00

80.00
Time (s)

60.00

40.00

20.00

0.00
389 294 337 330 306 321 345 271 316 335 289 304 359 334 272 346 298 291 331 263 299 359 278 325 337 253 288 339 335 314
EEG Length (s)

Fig. 16: Graphical Representation of the Performance

lower percentage level of sensitivity could have been due to the inability of the
algorithm in identifying some of the artifact events. The identification of arti-
fact related ICs in FORCe during ICA application on the wavelet coefficients
are dependent on manually adjusted threshold values, which classifies or make
a binary decision whether an IC is artifactual. So, having manually adjusted
fix thresholds may lead to detection errors, thereby not removing some of the
artifacts. On a separate note, the performance of the algorithms in retaining
the neural information of an EEG signal is evaluated through specificity. Speci-
ficity is the ratio of undistorted artifact-free EEG segments before and after
artifact elimination is performed. The ideal expectation is to have these por-
tions undistorted after the artifacts have been removed. FastEMD-CCA and
FORCe records an average specificity of 99.22% and 98.65% respectively, which
signifies that both algorithms doesn’t introduce much distortion to the neural
information of the EEG signals under evaluation. From the comparison, it is
clear that FastEMD-CCA has achieved better performance than FORCe on the

32
same set of EEG signals. The average computation time FastEMD-CCA took to
remove eyeblink artifacts from all 14 channels of these 60 EEG data-sets with
an average signal length of 312s ( 5 minutes) is 19.73 seconds, while FORCe
took 85.10 seconds. The computation time of FastEMD-CCA is at least 4 times
faster than that of FORCe.
The results have pointed out that the proposed algorithm, FastEMD-CCA
is highly accurate in removing eyeblink artifacts, proved by accuracy, error rate
and sensitivity measurement. It is also capable of retaining underlying EEG
data in uncontaminated EEG portions which were indicated by the specificity
percentage. Apart from this, the algorithm is also able to remove eyeblink
artifacts when the eyeblink artifacts are in continuous sequence as highlighted
in Fig. 17. The computation time of the algorithm is low as well, with an
average of 63 milliseconds processing time to remove artifacts from 1-second
length of EEG signal with 14 channels (256 samples x 14 EEG channels). This
makes it a feasible solution for applications requiring online removal of eyeblink
artifacts, with very low distortion to the neural signal.

(a) Observed EEG Signal

100

50

-50
Amplitude(microvolts)

-100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

(b) Artifact Free EEG Signal

100

50

-50

-100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Samples

Fig. 17: Continuous Eyeblink Artifacts are Detected and Removed

33
4. Conclusion

This paper has discussed an efficacious algorithm that incorporates unsuper-

vised artifact detection algorithm, enhanced-EMD, and windowed-CCA to auto-
matically identify and eliminate eyeblink artifacts from EEG signals without the
need to have an EOG channel as an artifact reference. The first portion of the
algorithm identifies the eyeblink artifacts without human intervention, thus it
can be also useful for other applications to utilize the eyeblink pattern/template.
Apart from artifact removal, applications, where eyeblink patterns may be use-
ful, are the driver drowsiness detection through eyeblink pattern, stress level
detection using eyeblink pattern/rate and for home applications such as home
light system triggering using eye blinking. The enhanced version of EMD is de-
veloped mainly to improve its processing time as conventional EMD is relatively
slow due to its iterative nature. CCA is then performed in a windowed manner
to characterize an online scenario. The algorithm as a whole is an unsupervised
and a fast approach that performs well in identifying and removing eyeblink
artifacts while preserving the underlying neural information as revealed by the
results. The computation environment for eyeblink artifact removal in EEG
plays a vital role in online applications. Although eyeblink artifact removal is
achievable using various methods, a medium or tool that supports and aid in
speeding up the eyeblink artifact removal process has not been sufficiently stud-
ied. Currently, MATLAB is the most preferred tool used for research purposes,
but for online applications MATLAB may not be a feasible platform. Hence,
future work will be in the direction of executing the algorithm, FastEMD-CCA
in an inexpensive computing environment such as in C++ and GPU based plat-
form as to support online applications.

Acknowledgment

The authors would like to thank the Ministry of Education, Malaysia for sup-
porting this research through the Fundamental Research Grant Scheme, FRGS

34
(FRGS/2/2014/TK03/UTP/02/1) and the Higher Institution Centre of Excel-
lence (HICoE) Scheme.

References

[1] J. A. Urigüen, B. Garcia-Zapirain, EEG Artifact Removal - State of the

Art and Guidelines, Journal of Neural Engineering 12 (3) (2015) 031001.

[2] G. H. Klem, H. O. LuÈders, H. Jasper, C. Elger, et al., The Ten-twenty

Electrode System of the International Federation, Electroencephalographic
Clinical Neurophysiology 52 (3) (1999) 3–6.

[3] S. Sanei, J. A. Chambers, EEG Signal Processing, John Wiley & Sons,
2013.

[4] J. Minguillon, M. A. Lopez-Gordo, F. Pelayo, Trends in EEG-BCI for

Daily-life: Requirements for Artifact Removal, Biomedical Signal Process-
ing and Control 31 (2017) 407–418.

[5] G. Gratton, M. G. Coles, E. Donchin, A New Method for Off-line Re-

moval of Ocular Artifact, Electroencephalography and Clinical europhysi-
ology 55 (4) (1983) 468–484.

[6] J. L. Kenemans, P. Molenaar, M. N. Verbaten, J. L. Slangen, Removal of

the Ocular Artifact from the EEG: A Comparison of Time and Frequency
Domain Methods with Simulated and Real Data, Psychophysiology 28 (1)
(1991) 114–121.

[7] J. Woestenburg, M. Verbaten, J. Slangen, The Removal of the Eye-

movement Artifact from the EEG by Rgression Analysis in the Frequency
Domain, Biological Psychology 16 (1-2) (1983) 127–147.

[8] P. Berg, M. Scherg, Dipole Modelling of Eye Activity and its Application
to the Removal of Eye Artefacts from the EEG and MEG, Clinical Physics
and Physiological Measurement 12 (A) (1991) 49.

35
 [9] T. D. Lagerlund, F. W. Sharbrough, N. E. Busacker, Spatial Filtering of
Multichannel Electroencephalographic Recordings through Principal Com-
ponent Analysis by Singular Value Decomposition, Journal of Clinical Neu-
rophysiology 14 (1) (1997) 73–82.

[10] T.-P. Jung, C. Humphries, T.-W. Lee, S. Makeig, M. J. McKeown,

V. Iragui, T. J. Sejnowski, Extended ICA Removes Artifacts from Elec-
troencephalographic Recordings, in: Advances in Neural Information Pro-
cessing Systems, 1998, pp. 894–900.

[11] R. Vigário, J. Sarela, V. Jousmiki, M. Hamalainen, E. Oja, Independent

Component Approach to the Analysis of EEG and MEG Recordings, IEEE
Transactions on Biomedical Engineering 47 (5) (2000) 589–593.

[12] V. Krishnaveni, S. Jayaraman, S. Aravind, V. Hariharasudhan, K. Rama-

doss, Automatic Identification and Removal of Ocular Artifacts from EEG
Using Wavelet Transform, Measurement Science Review 6 (4) (2006) 45–57.

[13] Q. Zhao, B. Hu, Y. Shi, Y. Li, P. Moore, M. Sun, H. Peng, Automatic

Identification and Removal of Ocular Artifacts in EEGImproved Adap-
tive Predictor Filtering for Portable Applications, IEEE Transactions on
Nanobioscience 13 (2) (2014) 109–117.

[14] W. De Clercq, A. Vergult, B. Vanrumste, W. Van Paesschen, S. Van Huffel,

Canonical Correlation Analysis Applied to Remove Muscle Artifacts from
the Electroencephalogram, IEEE Transactions on Biomedical Engineering
53 (12) (2006) 2583–2587.

[15] M. H. Soomro, N. Badruddin, M. Z. Yusoff, M. A. Jatoi, Automatic Eye-

blink Artifact Removal Method Based on EMD-CCA, in: ICME Inter-
national Conference on Complex Medical Engineering, IEEE, 2013, pp.
186–190.

[16] M. Shahbakhti, V. Khalili, G. Kamaee, Removal of Blink from EEG by

36
 Empirical Mode Decomposition (EMD), in: Biomedical Engineering Inter-
national Conference (BMEiCON), IEEE, 2012, pp. 1–5.

[17] B. Raghavendra, D. N. Dutt, Correction of Ocular Artifacts in EEG

Recordings Using Empirical Mode Decomposition, National Conference on
Communication.

[18] K. T. Sweeney, S. F. McLoone, T. E. Ward, The Use of Ensemble Empirical

Mode Decomposition with Canonical Correlation Analysis as a Novel Ar-
tifact Removal Technique, IEEE Transactions on Biomedical Engineering
60 (1) (2013) 97–105.

[19] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-C.

Yen, C. C. Tung, H. H. Liu, The Empirical Mode Decomposition and the
Hilbert Spectrum for Nonlinear and Non-stationary Time Series Analysis,
in: Proceedings of the Royal Society of London : A Mathematical, Physical
and Engineering Sciences, Vol. 454, The Royal Society, 1998, pp. 903–995.

[20] D. Labate, F. La Foresta, G. Occhiuto, F. C. Morabito, A. Lay-Ekuakille,

P. Vergallo, Empirical Mode Decomposition vs. Wavelet Decomposition for
the Extraction of Respiratory Signal from Single-channel ECG: A Compar-
ison, IEEE Sensors Journal 13 (7) (2013) 2666–2674.

[21] B. Noureddin, P. D. Lawrence, G. E. Birch, Online Removal of Eye Move-

ment and Blink EEG Artifacts Using a High-speed Eye Tracker, IEEE
Transactions on Biomedical Engineering 59 (8) (2011) 2103–2110.

[22] M. K. Islam, A. Rastegarnia, Z. Yang, Methods for Artifact Detection and

Removal from Scalp EEG: A Review, Neurophysiologie Clinique/Clinical
Neurophysiology 46 (4-5) (2016) 287–305.

[23] S. Çınar, N. Acır, A Novel System for Automatic Removal of Ocular Arte-
facts in EEG by Using Outlier Detection Methods and Independent Com-
ponent Analysis, Expert Systems with Applications 68 (2017) 36–44.

37
[24] V. Lawhern, W. D. Hairston, K. McDowell, M. Westerfield, K. Robbins,
Detection and Classification of Subject-generated Artifacts in EEG Signals
Using Autoregressive Models, Journal of Neuroscience Methods 208 (2)
(2012) 181–189.

[25] H.-A. T. Nguyen, J. Musson, F. Li, W. Wang, G. Zhang, R. Xu, C. Richey,

T. Schnell, F. D. McKenzie, J. Li, EOG Artifact Removal Using a Wavelet
Neural Network, Neurocomputing 97 (2012) 374–389.

[26] I. Daly, R. Scherer, M. Billinger, G. Müller-Putz, FORCe: Fully Online

and Automated Artifact Removal for Brain Computer Interfacing, IEEE
Transactions on Neural Systems and Rehabilitation Engineering 23 (5)
(2015) 725–736.

[27] L. Pion-Tonachini, S.-H. Hsu, C.-Y. Chang, T.-P. Jung, S. Makeig, Online
Automatic Artifact Rejection Using the Real-time EEG Source-mapping
Toolbox (REST), in: 2018 40th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC), IEEE, 2018,
pp. 106–109.

[28] A. Egambaram, N. Badruddin, V. S. Asirvadam, E. Fauvet, C. Stolz, T. Be-

gum, Unsupervised Eye Blink Artifact Identification in Electroencephalo-
gram, in: TENCON 2018-2018 IEEE Region 10 Conference, IEEE, 2018,
pp. 2148–2152.

[29] A. Egambaram, N. Badruddin, V. S. Asirvadam, T. Begum, Comparison

of Envelope Interpolation Techniques in Empirical Mode Decomposition
(EMD) for Eyeblink Artifact Removal from EEG, in: IEEE EMBS Confer-
ence on Biomedical Engineering and Sciences (IECBES), IEEE, 2016, pp.
590–595.

[30] H. Hotelling, Relations Between Two Sets of Variates, Biometrika 28 (3/4)

(1936) 321–377.

38
[31] H. Hallez, M. De Vos, B. Vanrumste, P. Van Hese, S. Assecondi,
K. Van Laere, P. Dupont, W. Van Paesschen, S. Van Huffel, I. Lemahieu,
Removing Muscle and Eye Artifacts Using Blind Source Separation Tech-
niques in Ictal EEG Source Imaging, Clinical Neurophysiology 120 (7)
(2009) 1262–1272.

[32] C. Zhao, T. Qiu, An Automatic Ocular Artifacts Removal Method Based

on Wavelet-enhanced Canonical Correlation Analysis, in: Annual Interna-
tional Conference of the IEEE Engineering in Medicine and Biology Society,
EMBC, IEEE, 2011, pp. 4191–4194.

[33] P. S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, Removal of

Ocular Artifacts in the EEG through Wavelet Transform without Using
an EOG Reference Channel, Int. J. Open Problems Compt. Math 1 (3)
(2008) 188–200.

[34] S. Khatun, R. Mahajan, B. I. Morshed, Comparative Study of Wavelet-

based Unsupervised Ocular Artifact Removal Techniques for Single-channel
EEG Data, IEEE Journal of Translational Engineering in Health and
Medicine 4 (2016) 1–8.

[35] N. K. Al-Qazzaz, S. Hamid Bin Mohd Ali, S. A. Ahmad, M. S. Islam,

J. Escudero, Selection of Mother Wavelet Functions for Multi-channel EEG
Signal Analysis During a Working Memory Task, Sensors 15 (11) (2015)
29015–29035.

[36] D. B. Stone, G. Tamburro, P. Fiedler, J. Haueisen, S. Comani, Automatic

Removal of Physiological Artifacts in EEG: The Optimized Fingerprint
Method for Sports Science Applications, Frontiers in Human Neuroscience
12 (2018) 96.

[37] S. ORegan, S. Faul, W. Marnane, Automatic Detection of EEG Artefacts

Arising from Head Movements Using EEG and Gyroscope Signals, Medical
Engineering & Physics 35 (7) (2013) 867–874.

39