Biomedical Signal Processing and Control 68 (2021) 102741

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Review of challenges associated with the EEG artifact removal methods

Wajid Mumtaz *, Suleman Rasheed, Alina Irfan
Department of Electrical Engineering, School of Electrical Engineering and Computer Science (SEECS), National University of Sciences and Technology (NUST),
Islamabad, Pakistan

A R T I C L E I N F O A B S T R A C T

Keywords: Electroencephalography (EEG), as a non-invasive modality, enables the representation of the underlying
EEG artifact removal method neuronal activities as electrical signals with high temporal resolution. In general, the EEG artifact removal
Independent component analysis methods have been considered as a fundamental preliminary step during EEG analysis. However, the associated
Challenges of artifact removal methods
challenges of EEG artifact removal methods should be addressed carefully, to fully utilize the data. This
EEG preprocessing methods
EEG artifact correction methods
manuscript is based on the notion that the full capacity of the EEG artifact removal methods can be achieved
while addressing the associated challenges well. Because these methods could enhance the inferences deduced
from the EEG data. The focus of this manuscript is to elaborate challenges (e.g., the algorithm-specific challenges
and general challenges) of the EEG artifact removal methods. Considering the challenges, the manuscript has
presented recommendations to address them. The manuscript also provides information on Matlab and Python-
based toolboxes developed for EEG preprocessing. In addition, this manuscript provides a brief account of the
EEG artifact types along with an overview of the EEG artifact removal methods. In short, this manuscript pro­
vides information on various EEG artifact removal methods and the recommendations provided serve as
guidelines for the selection of suitable tools and methods for EEG artifact corrections.

1. Introduction of the artifacts, it is difficult to extract only the artifacts without the loss
of actual neuronal data. In addition, there are methods that could not be
Electroencephalography (EEG) has been considered as a standard used for online applications. The online processing requirements could
means to record human brain activity in the form of electric pulses as a be imposed by the brain-computer interface (BCI) and neurofeedback
function of time. The EEG data have shown great potential as a diag­ applications and preferably for single-channel, which is a challenging
nostic and monitoring tool for various clinical applications such as the condition [6]. Furthermore, the EEG signals are very weak, typically in
quantification of anesthesia levels before/during a surgery [1], the the 20-mV range, and thus require amplification. However, amplifica­
diagnosis of epilepsy [2], and the prediction of occurrence of an tion of the signal leads to an amplification of the artifacts too. These
epileptic seizure [3], the neurofeedback applications for autistic patients challenges highlight the importance of an artifact handling stage in the
[4] and the neuro-rehabilitation [5]. However, the EEG has been EEG signal analysis pipeline that would remove the artifact activity from
suffering from many inherent challenges such as the removal of the raw EEG signals while preserving the neuronal activity of the brain [7].
additive noises (i.e., the EEG artifacts) that could be generated by However, the EEG artifact removal methods should overcome the
different noise sources such as the muscle movements or due to the challenges posed by the composite nature of the EEG artifact types.
electric line interferences. The EEG artifact removal methods are mainly As explained in Fig. 1, different EEG applications have different re­
used to clean the artifacts from the EEG data. The success of EEG artifact quirements of accuracy, speed, reliability, and ease of use of the subject.
removal methods enables the full utilization of the EEG data for clinical The trade-offs between these 4 factors would eventually decide which
and industrial applications. artifact removal algorithm would be most suitable for a particular type
The EEG artifact removal methods have encountered various chal­ of Application. This makes it even harder to select a single EEG artifact
lenges. These challenges could be either because of the complexity of the removal algorithm as a general best algorithm. For a brief review of
methods or could be because of the nonlinearities of the noise being different applications of EEG signals, we encourage our readers to read
added in the EEG signal. For example, because of the ‘nonlinear’ nature the following papers [8–10]. For Clinical Diagnostic applications (like

* Corresponding author.
E-mail address: [email protected] (W. Mumtaz).

https://doi.org/10.1016/j.bspc.2021.102741
Received 4 March 2021; Received in revised form 16 April 2021; Accepted 7 May 2021
Available online 13 May 2021
1746-8094/© 2021 Elsevier Ltd. All rights reserved.
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Fig. 1. Different Applications of EEG Signals along with the Desired Requirements. Comments are based on 4 factors only: Accuracy, Robustness, Real-Time, and
User’s Ease.

Seizure Detection, Alzheimer’s Disease Detection, and Brain Injury reliability so practitioners must do a speed-accuracy tradeoff to select an
Detection) the accuracy and reliability of the results outweigh the re­ optimal method based on available computational resources and
quirements of speed and user’s ease which makes Hybrid methods a very real-time constraints of the application. For Neuromarketing applica­
suitable choice. For BCI and Neurofeedback applications (like Wheel­ tions (like Customer’s Response Prediction, Wine Preference, and
chair Control, Exoskeleton Control, and Neuroprosthetics) the speed of Automotive Brand Preference), there are no hard constraints on speed
the algorithm matters as much as its accuracy, ease of the user, and and accuracy and user’s ease should be the primary focus. Ease of use

Fig. 2. General and Algorithm Specific Challenges associated with EEG Artifact Removal Algorithms and their Remedy.

2
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

means that the overall experiment protocol should be comfortable for Table 1
the user which suggests that the algorithm should use a minimum A Brief Summary of Artifact Removal Algorithms.
number of channels with minimum set up time. This may suggest using Type Algorithms Comments and Citations
algorithms that don’t require a lot of calibration and can work well with
Analog Methods The subtraction of artifacts A potentiometer circuitry is
a small number of channels ideally without the need for a reference from the recorded EEG data. used to record and combine
channel. artifacts to subtract them
Due to the above-mentioned reasons, it has almost been 50 years from the contaminated EEG
since researchers started exploring EEG artifact removal methods, and records. The popular citations
are provided here: [19,20,21,
still to date there is no consensus on which algorithm is optimal for a 22,23,24]
particular application. Therefore, the researchers must thoroughly study Regression Estimation of artifact and Regression-based methods
the pros and cons of each of these algorithms from multiple aspects (e.g., subtraction from the EEG were proposed both for time
automatic vs manual methods, online vs offline methods, suitability for data. and frequency domains. Also
proposed as an improved
a particular application, etc.) to decide the best choice. The details of
version of the analog
these algorithms are beyond the scope of this paper and these excellent methods. Popular citations
review papers can be used as a reference for details: [7,11–14] and [15]. are provided here: [25,26,27,
Moreover, there also exists other EEG artifact removal algorithms as 28,29,30,31,32,33]
well but we have kept the scope of our work to only the most used and Adaptive LMS, NLMS, RLS, etc Adaptive filtering suits EEG
Filtering preprocessing because of the
well-known EEG artifact removal techniques. flexible nature of their
This manuscript has provided a detailed review of different chal­ adaptive transfer function of
lenges associated with EEG artifact removal algorithms. According to the filters. The popular
our knowledge, this manuscript is the first of its kind that is solely citations are provided here:
[34,35,36].
dedicated to challenges associated with EEG artifact removal algorithms
BSS Independent Component ICA assumes that the data
and elaborates both algorithm-specific and general challenges associ­ Analysis (ICA) recorded by the surface
ated with these methods. Most recent reviews [14] and [15] have only electrodes is a linear
discussed different methods and their implementation details. Some combination of sources inside
older review papers have also discussed some of these challenges as a the brain. Popular citations
are provided here: [37,38,
subsection of their paper or as a few paragraphs in the discussion section
39].
and don’t cover a wide range of algorithm-specific and general chal­ BSS Canonical Correlation CCA is another BSS method
lenges associated with EEG artifact removal algorithms. For example, Analysis (CCA) that utilizes correlation to
Jiang et al. [12] and Islam et al. [11] include a discussion on artifact separate different sources of
EEG activity. It finds the basis
removal algorithms based on only four algorithm-specific challenges;
vectors of 2 sets of variables
additional reference channel, automatic, online, and applicability to a such that the correlation
single EEG channel. Similarly, Mannan et al. [7] discussed different between their projections
challenges of artifact removal algorithms in the discussion section but a onto the basis vectors is
detailed discussion specifically on each of these challenges wasn’t in the mutually maximized [7,40,
41,42,43,44].
scope of their work. Eventually, an explicit and detailed discussion on
BSS Principal Component Analysis PCA assumes orthogonality
different algorithm-specific and general challenges associated with EEG (PCA) between the brain activities
artifact removal algorithms comes up as a research gap. Therefore, as and the artifact sources [45,
shown in Fig. 2, the objective of this manuscript is to highlight different 46]. In case if amplitudes of
artifact and neuronal activity
challenges of EEG artifact removal methods and provide recommenda­
are similar then PCA doesn’t
tions on each one of them. perform well in separating
The manuscript has been divided into different sections: section 2 these types of artifacts [7].
elaborates on different types of artifacts commonly found in the EEG Frequency Wavelet Transform (WT) WT decomposition is mainly
data. Section 3 briefed the artifact removal methods. Sections 4 and 5 Decomposition Decomposition based on decomposing the
time domain signal to
talk discuss the algorithm-specific and general challenges of the EEG
different frequency
artifact removal methods. Section 6 gives recommendations for each of components utilizing the full
these challenges and discusses publicly available Python/MATLAB capacity of the EEG data and
toolboxes. thresholding is applied to
identify artifact related
components and then the
2. Different types of artifacts artifact free signal is
reconstructed from non-
In general, the EEG signal artifacts may be broadly categorized as artifactual components only
physiological and non-physiological artifacts. [11,47,48,49,50]
Frequency Empirical Mode EMD is mainly a data-driven
Decomposition Decomposition (EMD) approach and subjective to
1. Physiological Artifacts: These artifacts are also known as internal/ individual artifact types. For
intrinsic artifacts and are related to physiological sources of the example, template matching
human body like ocular artifacts (eye blinks and movements), mus­ based autodetection and
EMD-based artifact removal
cle artifacts (muscles movement, jaw/head movement, chewing),
method [51].
and cardiac artifacts (related to heartbeats). These artifacts some­ The hybrid methods have
Hybrid methods involve more
times may also be referred to as Electrooculogram (EOG), Electro­ enjoyed the combined
than one algorithm such as
myogram (EMG) and Electrocardiogram (ECG), electrocardiograph Hybrid Methods ICA and wavelet
benefits of individual
(EKG) corresponding to the sensors/techniques that are used to algorithms that implicated a
decomposition implicated in
win-win situation. A brief list
measure these signals. the wICA method.
is cited here.
2. Non-Physiological Artifacts: These artifacts are also known as
(continued on next page)
external/extrinsic artifacts and their sources are related to external
factors like environment noise or poor experimentation protocols.

3
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Table 1 (continued ) Decomposition (EMD). On the contrary, the hybrid methods involve a
Type Algorithms Comments and Citations combination of these single methods, e.g., wavelet-based ICA methods
[16].
Adaptive filtering and blind
source separation (AF-BSS)
[52], adaptive filtering and 3.3. Single vs multiple channel methods
wavelet transform (AF-WT)
[53,54], Adaptive Filtering There could be some situations for which the single-channel methods
and Empirical Mode
perform better than the multiple channels such as the EEG-based anes­
Decomposition (AF-EMD)
[55], Wavelet Transform and thesia monitoring may need a single-channel artifact removal method.
Blind Source Separation On the other hand, the ICA-based methods could be more suitable for the
(WT-BSS) [56], Empirical multiple channel scenarios.
Mode Decomposition and
Blind Source Separation
(EMD-BSS) [57], Blind Source
3.4. Offline vs online/real-time methods
Separation with Support
Vector Machines (BSS-SVM) The offline scenario may include EEG data recordings for training
[58,59], etc. deep learning architectures e.g., the epileptic EEG records. On the
contrary, the neurofeedback applications might need online methods e.
These artifacts include power line noise (50/60 Hz), electrode mal­ g., an autistic patient training during game playing.
function (due to floating electrodes such as loose connection with the
scalp or high impedance electrodes), electromagnetic interference 3.5. Manual vs automatic methods
(due to other electrical devices placed nearby), and variation in
impedances due to the slow drying of the conductive paste. The manual methods may include visual inspection of the artifacts
and performance of manual deletion. The automatic methods enable
The non-physiological artifacts can be handled well by following auto-detection and auto-correction of the artifacts as is the case with the
strict experimental protocols and precise recording systems augmented wICA method [16].
with simple, linear filtering (e.g., the notch filter for line noise removal).
On the contrary, the physiological artifacts are harder to remove as their 3.6. Linear vs non-linear methods
spectrum often overlaps with the underlying brain activity, requiring the
employment of advanced methods for artifact handling (removal/ EEG artifact removal methods may also be categorized as linear and
reduction) [12]. non-linear methods based on how they apply the correction. For
example, regression is a linear method while deep learning-based arti­
fact removal methods are non-linear. There is also a scope to explore
3. EEG artifact removal methods
non-linear analysis-based approaches [17,18] in this regard.
Table 1 provides citations and a brief description of specific types of
In the literature, various artifact reduction methods have been pro­
different artifact removal methods. The interested readers can directly
posed. An update on the EEG artifact removal methods can be found in
jump to the relevant citations for a detailed description of each method.
these excellent review articles [14,15]. This section provides a brief
account of different categories of the methods. The EEG artifact removal
4. Algorithm specific challenges
methods can be categorized into general and specific categories. Table 1
provides the specific categories of the methods. On the other hand, the
4.1. Requirement of reference channel
general categories can be made according to different requirements of
algorithms e.g., whether the method requires a reference channel or not,
The literature has evidenced various methods that require a refer­
single or hybrid methods, etc. These categories help the readers in un­
ence channel to complete the artifact correction, e.g., the analog
derstanding the different challenges mentioned in the later sections. A
methods, regression-based methods, adaptive filtering-based methods.
list of such categories is provided as follows.
These methods exploit the extra information provided by an additional
reference channel like EOG/ECG as an estimate of relevant physiological
3.1. Reference channel vs non-reference channel methods artifacts (ocular, cardiac) and subtract them from the raw EEG data to
reduce/remove physiological artifacts.
The EEG reference channel methods require an extra reference The requirement of an additional reference channel poses the chal­
channel to estimate the artifacts. Once the artifacts are identified, a lenge of handling the placement and extra noise being associated with
subtraction from the EEG recordings can be done. On the other hand, the such a channel. For muscle artifacts, although both the ECG and EMG
non-reference channel methods may perform the decomposition of the sensors are good at picking muscle activities, still muscle movements
raw EEG signals into different components or transform them to another produce very dynamic artifacts. Especially, the EMG sensors must be
domain so that thresholding can be employed to eliminate artifactual placed at multiple locations to fully cover these dynamics (as there is no
components and then reconstruct the corrected signal using only the single source of muscle artifacts), which isn’t practically feasible. Hence,
remaining components [12]. While the former ones have the advantage the artifact removal methods that don’t need a reference channel such as
of exploiting extra information from the reference channel to estimate the wavelet transform, and the blind source separation algorithms
artifacts, the latter ones don’t need a reference channel, so they are more should be the preferred choices for removing the EMG and ECG artifacts.
convenient for the user and have a broader scope. An additional reference channel might create discomfort for the study
participants because of the extra gel paste that is applied to the skin
3.2. Single vs hybrid methods before attaching EOG/EMG electrodes.
From the algorithmic point of view, the performance of reference
This categorization is based on how many individual algorithms are channel dependent algorithms depends upon the robust/correct
involved in the pre-processing pipeline. Popular single methods for EEG recording of reference signals. In case, a reference channel malfunctions
artifact removal include Linear Regression, Adaptive Filtering, Wavelet or because of a loose connection, severe effects on the overall pre­
Transform, Blind Source Separation (BSS), and Empirical Mode processing pipeline could not be avoided.

4
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

The EOG reference channels are normally placed at the Fp1 and Fp2 expertise and it should be easy for the operator to use it without in-depth
locations yet they might record the neuronal activities as well and could knowledge of the workings of the algorithms.
not serve as pure EOG reference. The methods that use EOG signals as a The performance of the manual or semi-automatic methods depends
reference for ocular artifact removal suffer from bidirectional contami­ on the expertise level of the operator. For example, the ICA-based semi-
nation error [60]. These methods assume that the pure EEG signal and automatic methods could perform better if operated by an expert EEG
EOG signals are uncorrelated which isn’t a valid assumption. EEG sig­ analyst. The manual step to finding bad independent components (ICs)
nals can be contaminated by EOG signals and vice versa and this bidi­ requires domain expertise and one must visualize the ICs by using
rectional interference would lead towards artifact removal errors. topographic maps, power spectrum, and time-domain characteristics to
Because the EOG signals might have captured some neuronal activity; find ICs corresponding to artifacts which is not a trivial job. An expert
therefore, the removal of the EOG also implicated the removal of the operator could handle the shortcomings associated with the ICA-based
neuronal information as well. semi-automatic methods.
Moreover, algorithms that need an extra reference channel to
4.2. Applicability to single vs multi-channel EEG data remove artifacts demand a higher level of expertise as compared to their
alternatives. The operator must be cautious about electrode pop and
The performance of many EEG artifact removal methods depends malfunction errors of the reference channel as any noise introduced via
upon how many EEG channels or electrodes are used while recording the the reference channel would affect the underlying EEG signal as well due
data. Therefore, it is not necessary for an algorithm that performs well to the dependence of the artifact removal algorithm on this channel. In
on multi-channel EEG data to perform well on single-channel EEG re­ clinical applications, a well-trained person is usually available who can
cordings and vice versa. Ideally, an EEG artifact removal algorithm perform manual EEG artifact removal and can take care of the extra
should be independent of the choice of electrode cap (i.e., number of channel if applicable. However, for other applications like BCIs, it is
channels) and should work equally well for a single channel or for always preferable to make the system independent of the expertise of the
multiple EEG channels. However, empirically it has been observed that operator.
the performance of certain EEG artifact removal algorithms, especially
Blind Source Separation based algorithms like ICA rely heavily on the 4.5. Requirement of calibration
number of electrodes and they become better at removing artifacts as the
number of channels increases. On the contrary, it also has the disad­ Calibration is the preliminary step of some EEG artifact removal al­
vantage that ICA can’t remove artifacts using a single EEG channel. It is gorithms to fine-tune their parameters or threshold settings. Regression
also worth mentioning that algorithms like linear regression and adap­ based algorithms require a calibration stage to compute Artifact Prop­
tive filtering although require an extra reference channel, yet they can agation Coefficients (Betas) to find the contribution of a noise source
be applied to single-channel EEG data effectively. from reference channel to different EEG channels. After calibration,
Multi-channel data has more information than single-channel data these parameters are used to subtract the effect of physiological artifacts
but in recent years, single-channel EEG devices have risen in demand from EEG data. Similarly, some automated versions of wavelet analysis-
and use because of their usability for measurement and their portability. based artifact removal methods require calibration for proper threshold
Applications involving clinical diagnostics, prosthetics, wheelchair setting.
control, etc. usually involve EEG readings from multiple EEG channels It should be noted that extra precautions need to be taken during the
but applications like driver’s drowsiness detection or home healthcare calibration phase as the parameters learned during the calibration runs
application usually involve a single EEG electrode. In these situations, would have a direct impact on all the subsequent analyses. Moreover, if
the researchers are bound to use EEG artifact removal algorithms that an application requires calibration every time, we have to use it then
can work on a single EEG channel. that would be very inconvenient for the end-user.

4.3. Automatic vs manual 4.6. Real-time constraints

Automated methods are always a preferable choice over manual Neurofeedback and BCIs are among the key applications of EEG that
methods unless they compromise the accuracy and reliability of the would greatly improve the quality of life of people with motor disabil­
system. Manual methods of artifact removal may involve visual in­ ities. However, these applications demand real-time control without
spections and deletion of artifact data; the semi-automatic methods may compromising on accuracy. EEG signal artifacts have magnitudes com­
involve both manual and automatic operations. Whereas fully automatic parable or larger in magnitude than the underlying brain activity that
methods don’t need human intervention for their processing. The might alter the result of the classification stage of the BCI. As a result, the
manual intervention makes manual and semi-automated methods rela­ control command of the wheelchair or prosthetic arm could cause
tively time-consuming, and they aren’t easily scalable for large EEG inconvenience for the user. Therefore, accurate artifact removal algo­
datasets. Similarly, for real-time and online applications it is mandatory rithms that support real-time operations should be considered.
to use an automated method, so one must be careful about this aspect of For real-time applications, low latency is almost as important as
the different EEG algorithms as many popular EEG artifact removal al­ being accurate, so accuracy-cost trade-offs are made in these scenarios.
gorithms are manual or semi-automatic. Regression and Adaptive Filtering are relatively fast, but the require­
The ICA-based artifact removal is a semi-automated method and may ment of extra channels makes them uncomfortable for the user so
involve two stages: first, it computes independent components (ICs) despite their low computational cost they aren’t usually used in BCI and
from the raw EEG data, and then the operator must visualize each of Neuroprosthetics applications. On the contrary algorithms like EMD and
them (using topographic maps and power spectrum) to see which ones BSS although don’t have a reference channel yet are computationally
correspond to sources other than neural activity e.g., eyes, muscles, etc. intensive. The researchers must make trade-offs between speed and
and then the second stage involves reconstructing the EEG data from accuracy. Moreover, In a recent review [7], it has been found that hybrid
only non-artifact ICs. The second stage requires manual visualization of methods are outperforming single methods and the EEG research com­
bad ICs which is an issue towards automation of ICA. munity is shifting from single methods to hybrid methods however these
methods are a combination of 2 or more single methods that increases
4.4. Expertise of operator the computational complexity.
The Choice of Programming Language also directly influences the la­
An ideal EEG artifact removal should require minimum user tency of the final model. MATLAB, Python, and C++ are key

5
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Table 2
Summary of the algorithm-specific challenges.
Method Need of Reference Applicable to Single Expertise of the Calibration Fully Automated Realtime * Applicable to all types
Channel EEG Channel Operator of Artifacts

Required to estimate Medium Yes Yes Yes No

Regression artifact propagation Yes Reference Needed to Once calibrated it can Theoretically possible,
coefficients (betas) Channel estimate betas work automatically but practically not
feasible for muscle
Required to estimate
artifacts due to
Adaptive noise and subtract it
Yes Precautions No Yes Yes unavailability of a
Filtering from raw EEG in
single robust EMG
feedback
reference
Mostly No Not Applicable High No Mostly No
Manual Bad IC
A few variants use Blind Source
Selection. Some
reference channel separation theory
advanced
ICA for automatic bad based methods Manual bad IC No Yes
Some variants
Independent assume that the selection
automate it by
Component (IC) number of artifact
adding an additional
rejection sources should at
computational stage
least be equal to
CCA No Easy No Mostly Yes Yes Yes
number of channels
PCA No Easy No Most Yes Yes
Wavelet implementations of
No Yes Easy No No Yes
Analysis these methods use
automated
EMD No Yes Easy No No Yes
thresholding
Mostly Not Required Mostly Yes Mostly Easy Mostly No Mostly Yes Mostly No Mostly Yes
Needed only if the Would need Calibration
Not applicable only Extra Fail for muscle
first stage algorithm reference channel required only if
if Blind Source computational artifacts only If
Hybrid ** requires a reference precautions if first first stage
Separation (BSS) complexity as it is a regression or adaptive
channels e.g., stage algorithm algorithm require
algorithms are at combination of 2 filtering are first stage
regression or requires a calibration e.g.,
first stage single methods algorithms
adaptive filtering. reference channel linear regression
*
Realtime suitability of an algorithm also depends on the available computational resources and hardware specific optimizations. Here we have just given a relative
comparison. For practical applications, one should always decide based on their computational requirements.
**
For Hybrid Methods exact comments would depend upon the choice of individual methods in the pipeline and the order in which they are applied.

competitors in the domain of EEG-related software development. Python 5. General challenges

has the advantage of being an easy-to-use high-level programming
language which shortens the development time. However, it is much 5.1. Selection of an evaluate criterion
slower than MATLAB and C++ and not the first option for real-time
applications. MATLAB and C++ are faster but have very large devel­ Comparing performances of different EEG artifact removal algo­
opment time as compared to Python. Moreover, MATLAB also has some rithms seems an important challenge that arises because of the lack of
great EEG toolboxes, but the biggest issue is that it is not open source. proper validation strategies. In general, the validation may be per­
Therefore, one should have a fair understanding of the available formed in two ways.
computational resources to have an idea of which language should be The first approach utilizes real EEG data. From a composite EEG
chosen. signal, it is difficult to quantify either the pure EEG signal or noise.
Hence, it is difficult to compute objective evaluation metrics like signal-
4.7. Different algorithms optimal for different artifact types to-noise ratio (SNR), Mean Square Error (MSE), etc. Therefore, the most
popular way for inspecting the performance of artifact removal algo­
An ideal EEG artifact removal algorithm should be able to deal with rithms on real EEG data is still by visual inspection. This way of evalu­
all types of artifacts. However, it is challenging to use a single method ation is neither scalable nor objective measures. Although some
for all types of EEG artifacts. In the past, algorithms were used to remove researchers have proposed some methods [13] to use objective evalua­
a particular type of artifact. In recent years, the trend has shifted towards tion measures on real EEG data, there is no consensus on a single
studying artifact removal algorithms that can remove all types of arti­ approach and visual inspection is still the most popular way of
facts [7]. inspecting the performance of artifact removal algorithms on real EEG
Algorithms that need a reference channel are only suitable for ocular data.
and cardiac artifacts as EOG and ECG references give reliable mea­ The second approach is to use synthetic or simulated EEG data. The
surements of underlying ocular or cardiac activity. Muscle artifacts can’t benefit of this approach is that we know the pure EEG signals and so we
be practically removed by EMG sensors as they are quite dynamic, and it can estimate noise from the contaminated signal. Hence, the evaluation
is really hard to use a single muscle’s EMG signal as a true representative metrics like SNR, MSE, etc. can be computed. However, it is hard for the
of all muscle artifacts. On the other hand, algorithms that don’t need a simulated data to be a true representation of actual raw EEG signals.
reference channel can theoretically deal with all types of artifacts e.g., Moreover, most of the simulated datasets used by researchers mostly
ICA is a popular choice in this regard, and it can deal with all types of deal with one type of artifact only which makes it hard to compare al­
artifacts. Table 2 provides a summary of the application-specific gorithms for multiple types of artifacts.
challenges.
5.2. Lack of open-source EEG artifact datasets

One of the key challenges for researchers developing artifact removal

algorithms is that there is no single EEG artifact dataset that is accepted

6
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Table 3 embedded systems developers working in the industry and there is a lack
List of Open-Source EEG Artifact Datasets. of proper guidance and training in academia in this regard. This gap in
Dataset Title Artifact Nature Ref Dataset Link software and hardware is often responsible for academic prototypes not
Types being able to create their worth in the market. And the EEG community
A semi-simulated Ocular Simulated [61] https://data.me needs to also focus on optimizing their solutions from hardware aspects.
EEG/EOG dataset ndeley.com/data
for the comparison sets/wb6 5.5. Special challenges of machine learning based algorithms
of EOG artifact yvr725d/4
rejection techniques
EEG eye artifact Ocular Real [62] https://osf.
Recently there has been a lot of research done in applications of
dataset io/2qgrd/ machine learning especially deep learning in the domain of EEG signal
EEG dataset Ocular, Real – https://github. processing. However, the increase in performance comes with an extra
contaminated with Head/Jaw com/inabiyoun computational cost and requirement of large training data which isn’t al­
artifact/noise Movement i/EEG_dataset_for_
ways a case especially if we talk about brain signals. There is a need to
artifact-noise_det
ection further explore Transfer Learning and Data Augmentation Strategies for
The TUH EEG Artifact 5 Types Real [63] https://www.isip. EEG signals to deal with data scarcity issues.
Corpus (TUAR) piconepress.co Transfer Learning is a technique that uses a pre-trained model
m/projects/tuh trained on a similar task as a baseline and then its parameters are fine-
_eeg/html/dow
nloads.shtml
tuned for the new task. This approach helps us train machine learning
models in a shorter amount of time with less training data. Unfortu­
nately, transfer learning in EEG signal processing faces a special chal­
by the EEG community. Researchers usually use simulated datasets with lenge that the computer vision community didn’t have to face. In
only a single type of artifact that is not a true representation of actual computer vision tasks, it is easy to resize images to fit the input size of
EEG signals. Others might collect some data under controlled experi­ the pre-trained model however, it isn’t possible to train an EEG-based
mental conditions, but they are left with the option of only visual in­ model from another model when the electrode cap and configuration
spection as discussed in the previous section. To make things even of EEG channels aren’t exactly identical.
harder, most of the time researchers don’t open source their artifact Another way to increase the data synthetically is to apply different
datasets so it is really hard to actually compare different algorithms data augmentation techniques on the available EEG dataset. For images,
across studies which go back to the issue of reproducibility. Some au­ it is easy to realize data augmentation strategies like resizing, reshaping,
thors also mention that data can be provided if requested, so one should rotation, cropping, flipping, etc. which increase the amount of available
try to contact them to show interest in their dataset. Table 3 lists some of training data and so we can extract more insights from the existing
the open-source EEG Artifact Datasets that we found available for public dataset. Unfortunately, It is not very easy to realize such transformations
use. for multi-dimensional non-stationary time-series data like EEG.
Deep learning-based methods are often criticized for their lack of
5.3. Lack of open-source implementations explainability and often referred to as “Black Box”. The main reason is
that, unlike traditional algorithms, the feature extraction is automatic so
Among the key barriers in rapid progress in EEG artifact removal even we may get above 95 % accuracy, we may not know what features
techniques is the lack of open-source implementations. Among the key are used to arrive at this decision and the model might have memorized
reasons for the rapid development of computer vision algorithms is the unintended and irrelevant information as well that would deteriorate
availability of open-source and pre-trained object classification, detec­ the results in deployment. These papers [64,65] are a good starting for
tion, and segmentation algorithms. The benefit of this practice is that detailed exploration of these issues in deep learning models. Similarly,
researchers don’t have to reinvent the wheel and the focus remains on neural networks are also a hot spot for adversarial attacks [66] as many
solving their actual problem instead of focusing on writing the code people try to use some latent feature representation to exploit the weak
again. Unfortunately, this practice of sharing the code implementations spots in the deep learning algorithm. So explainable and secure AI is also
is still not widely adopted in the EEG community. Although some of the a big concern.
popular algorithms like ICA can be found in most of the EEG analysis Moreover, data privacy and ethics concerns are much greater for ma­
software/libraries. However, if someone wants to explore other artifact chine learning based methods as we must use data from users to train our
removal algorithms that aren’t well known then he/she must have to machine learning system and people are very sensitive about their data
implement that from the scratch. To make things worse, it is more especially when it comes to sharing their brain activity.
difficult to integrate a custom algorithm to be a part of an EEG pro­
cessing pipeline of existing libraries and toolboxes and this integration 5.6. Lack of benchmarks/competitions
requires more time and coding effort that could have been used in
solving the problem. One of the main reasons for very fast development in the field of
computer vision was the availability of large amounts of open-source
5.4. Lack of hardware optimized implementations datasets that were often a part of a competition. So, whenever a new
model is proposed then researchers compare their results on test sets of
Lack of hardware optimized implementations makes it difficult for well-established datasets which makes it easy to benchmark their pro­
researchers to transform their EEG-based prototypes into commercial posed methods. Unfortunately, in the EEG community, there is no
products. Most of the time, when a researcher writes a piece of code in consensus on a single artifact removal dataset as a benchmark which
academia then its sole purpose is to demonstrate that proof of a concept makes it difficult to compare across different models.
of an idea which is the first stage of any product development. However, Among many other reasons, one of the benefits of competitions/
it is very rare for successful prototypes to go to production as this re­ benchmarks is that the train set, and test set are similar for everyone. In
quires more coding effort to optimize the code for real-world deploy­ literature, one of the worst mistakes of applying machine learning
ment. The commercial viability largely depends upon the choice of cost- blindly to a new application is to use the same training set for valida­
effective hardware to implement the code and optimizing the software tion/testing, which leads to information leakage so one might get away
for that platform is a necessary step in this process. The practice of with publishing his excellent results in reputed venues however, this
designing hardware optimized algorithms is usually only adopted by model would be of no practical use. So, having a consistent and separate

7
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Table 4 Table 4 (continued )

Comparison of Popular EEG Artifact Removal Algorithms in Literature. Study Year Type of Type of Competing Winners
Study Year Type of Type of Competing Winners Artifact Data Algorithms
Artifact Data Algorithms
Different
[67] 2021 Ocular, Real and CCR, AWCCR, AWCCR Winners for
Muscle, Simulated ICA, Wavelet (Automatic different levels
External ICA Wavelet of
Common contamination
Component [84] 2011 Ocular Simulated Adaptive ICA SOBI based
Rejection) Filtering, PCA, ICA for real-time
[68] 2021 Muscle Real and EEMD-ICA, VMD-CCA ICA Variants
Simulated EEMD-CCA, (Hybrid, [85] 2009 Ocular Real Adaptive LMS based
VMD-ICA, Variational Filtering (LMS, Adaptive
VMD-CCA Mode RLS), ICA (ex- Filtering
Decomposition- ICA, SOBI),
CCA) Time-Varying
[69] 2020 Ocular, Real WPD based Wavelet Packet Adaptive
Muscle Method, ICA Decomposition Method
and Variants, (WPD) based [86] 2008 Muscle Real and ICA Variants AMUSE based
Movement Wavelet Method Simulated (AMUSE, SOBI, ICA
Transform Infomax, JADE)
[70] 2019 Ocular Real and EAWICA, ICA- EAWICA [87] 2007 Ocular Simulated Adaptive Adaptive
Simulated W (Wavelet Filtering, Time Filtering
Transform and Domain
ICA Hybrid) Regression
[71] 2019 Ocular Real ASR, rASR Riemannian [60] 2004 Ocular Real and Regression, Regression and
Artifact Simulated PCA, ICA PCA
Subspace [88] 1998 Ocular, Real and PCA, ICA ICA
Reconstruction Muscle Simulated
(rASR)
[72] 2019 Ocular Real ASR, ICA, PCA Artifact
Subspace training and test set ensures to overcome this problem. Moreover, if we
Reconstruction try too many different settings with the same validation set then this
(ASR)
might lead to overfitting the validation set that would not give optimal
[73] 2018 Ocular and Real and SuBAR, Surrogate based
Muscle Simulated Wavelet Artifact results on the test set (or real-world deployment).
Thresholding, Rejection
CCA-EMD (SuBAR)
[74] 2018 Ocular, Real and BSS-REG (BSS- BSS-REG 5.7. Reproducibility issues
Muscle Simulated Regression), (Hybrid)
ICA, Regression Although the lack of open-source data or code is the obvious reason
[75] 2018 Ocular, Real and ICA, CCA Multi-Channel
behind reproducibility concerns. However, there is much more to the
Muscle, Simulated Wiener Filter
and (MCWF) list. This subsection provides a brief account of these issues.
Movement Reproducibility requires uniformity that can be achieved by pro­
[76] 2018 Muscle Real Multiple ICA Extended moting uniform data storing methods. Added to the challenge, there is
Variants Infomax based no agreement on a single data storing format and different libraries, and
ICA
[77] 2017 Ocular Real EYE-REG, EYE- Eye artifact
EEG Headset providers output data in different formats. Therefore, it is
SUB, MARA, subspace hard to use some library/framework that is outside the pools of re­
EYE-EEG, subtraction sources provided by the headset manufacturer. Useful research time can
REGICA (EYE-SUB) be saved if such resources are available and open-sourced.
[78] 2017 Ocular Simulated HMM-AF, ICA Hidden Markov
It is common to observe that the authors might miss providing
Model with
Adaptive detailed implementation aspects in their publication. A well-written
Filtering (HMM- methodology section should provide a clear view of different
AF) modeling and design choices and it should enable its reader to re-
[79] 2017 Ocular Real Eye’s Ballistic EBPM implement the same idea. However, this is not the case with most of
Physiology
the research work, and many times researchers don’t share things like
based Method
(EBPM), ICA Which EEG headset and software were used to acquire data? What
Variants experiment timing protocols (e.g., cue onset, button press, etc.) were
[80] 2016 Ocular Real ANC Scheme, Novel Adaptive followed? Which library/toolbox was used? Which preprocessing tech­
ICA, ASR Noise
niques were followed? Which types of features were used? Which data-
Cancellation
(ANC) Scheme split strategy was used (random split or k-fold)? How did they perform
[81] 2016 Internal Real Signal Space tSSP for external cross-validation? Which loss function did they choose? What Hyper­
and Projection artifacts ICA for parameters optimization was performed? What was the evaluation
External (SSP), ICA internal artifacts metric? etc. Without these pieces of information, it is hard for new re­
[82] 2013 Ocular Real NMF, ICA Non-negative
searchers to re-implement the idea. Table 4 provides a comparison of
Matrix
Factorization popular EEG artifact removal algorithms.
(NMF) for fewer
EEG Channels 6. Discussion
EMD for highly
Real and ICA, CCA, EMD,
[83] 2012 Muscle contaminated
Simulated WT 6.1. Recommendations for algorithm specific challenges
data

‘Reference channels’ are a feasible choice if we are using EEG

8
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Table 5
Summary Table with respect to methods and algorithm-specific challenges.
Method Need of Reference Applicable to Single EEG Expertise of the Calibration Automatic Real- Applicable to all types of
Channel Channel operator Time artifacts

Regression Yes Yes Medium Yes Yes Yes No

Adaptive Yes Yes Medium No Yes Yes No
Filtering
ICA No No High No No No Yes
CCA No No Easy No Yes Yes Yes
PCA No No Easy No Yes Yes Yes
Wavelet No Yes Easy No Yes No Yes
Analysis
EMD No Yes Easy No Yes No Yes
Hybrid No Yes Easy No Yes No Yes

recordings for a critical task as the extra information provides us more automated without any manual intervention, so it is easy for the oper­
insights related to the artifact and so it is better to use this information to ator to use that method without knowing the detailed working of the
make a robust EEG analysis system. This is the case in clinical di­ underlying technique. Second, the algorithm shouldn’t use an extra
agnostics where the accuracy of the results weighs more than the con­ reference channel to estimate artifact activity. This would avoid extra
venience of the subject and we only need to record the signal once a precautions associated with the reference electrode.
while, so we don’t have to permanently attach these extra electrodes to As far as the ‘challenge of calibration’ is concerned, exploration of
the user for a long period of time. However, for other applications like calibration-free methods should be the priority. However, if an algo­
prosthetic arm control, wheelchair control, neurofeedback, etc. we must rithm performs good results, then calibration can be an option because it
record the brain activity regularly so having an extra electrode always doesn’t influence the online operation of the overall pipeline and we just
attached to one’s head (in case of EOG artifacts) would be very have to compromise the comfort of the user for the gain of accuracy.
inconvenient. Linear Regression uses an extra reference electrode for calibration of
The ’bidirectional contamination’ error can be handled. One simple artifact propagation coefficients so extra precautions associated with the
solution to this problem is to apply a low-pass filter on the EOG signal reference electrode should again be considered.
before using it as reference [89] which is based on the assumption that Removal of artifacts from EEG data demands processing power and
most of the high-frequency content in the EOG belongs to EEG signal so computational time. This can become an obstacle in ‘real-time applica­
removing that would reduce the bidirectional contamination. However, tions’ where a delay incurred due to pre-processing can be undesirable.
there is no consensus on the threshold frequency of low pass filters and For such real-time applications, further modifications to the artifact
some studies have shown that cerebral artifacts propagate to removal stage can be explored. One of these possibilities is to define a
low-frequency EOG (alpha and beta bands) as well, so simple low pass specific frame of EEG data and to run the artifact removal algorithm only
filtering isn’t an optimal solution in that case. To tackle this challenge, on the given frame. This stems from the findings of previous research in
Wallstrom et al. [90] utilized an Adaptive Bayesian filtering technique EEG signals that analysis of only the initial portion of an EEG signal can
that reduces bidirectional contamination errors substantially and might provide accurate results [96], motivated by the observation that medical
be a better option if bidirectional contamination must be avoided. practitioners classify brain signals as normal or abnormal using only the
Using a ‘simulated/virtual reference channel’ can be an option if first few minutes of the data. Performing artifact removal on a smaller
someone must apply an EEG artifact removal algorithm that requires a segment of data would reduce the time and complexity of resources
reference channel, and it is not possible/feasible to use an actual needed before the data is available for the actual application at hand.
reference channel. To simulate EOG activity, we may use EEG signals This can remarkably reduce the processing delay for real-time
recorded from prefrontal electrodes FP1 and FP2 as these electrodes applications.
show the greatest correlation with EOG activity. However, this might Extraction of the first few minutes of the EEG recordings is also
lead to the loss of some underlying neuronal activity captured by these promising beyond delay reduction. This can help reduce the non-
electrodes and might not be an option if we are interested in studying the physiological artifacts, because once the EEG electrodes are placed on
neuronal activity of the prefrontal cortex. A simulated ECG channel can the scalp, the impedances due to external factors begin to vary with time
be obtained if the recorded data also have MEG (magnetoencephalog­ owing to the gradual drying of the conductive gel. This causes variations
raphy) signals from a gradiometer or magnetometer. MNE python [91] in the EEG signal which are not representative of the actual neural ac­
is an open-source library that has built-in functions to use a simulated tivity of the brain. Therefore, the initial segment of the data is hence the
channel as a reference for artifact removal and its documentation can be most representative of brain activity.
referred to for further implementation details. The issue of ‘applicability of an algorithm to only a certain type of
Regarding the ‘automation’ aspect of Blind Source Separation algo­ artifact’ can be resolved by focusing further research on those algorithms
rithms, one possible direction is to combine an extra computational that can deal with all types of artifacts ideally without the need for a
stage with the output of ICA to make finding bad Independent Compo­ reference channel. A possible research direction would be to investigate
nents (ICs) automatic. One of the options is to train a machine learning algorithms that don’t treat each artifact separately and can differentiate
classifier and apply it to topo maps of ICs to automatically find bad ICs automatically pure EEG from an artifact. For example, Jafari et al. [97]
[92,93]. Another option is to combine an ICA with another EEG artifact combined ICA with multi-instance learning to classify bad ICs without
removal algorithm like wavelet analysis that would make it a hybrid explicitly labeling all types of artifacts and used all artifacts as a single
method i.e., wICA [94,95]. However, both these methods also introduce class. Table 5 provides a summary of the methods and the algorithm
extra computational cost to the overall EEG artifact removal pipeline so specific challenges.
this factor should also be considered while automating Blind Source
Separation algorithms like ICA.
An ideal EEG artifact removal algorithm shouldn’t ‘require high 6.2. Recommendations for general challenges
expertise of an operator’. From our understanding, fulfilling this
requirement has two prerequisites. First, the method should be fully To have a ‘uniform and standard evaluation protocol’ for different
artifact removal algorithms, [7] suggests a 3-stage evaluation

9
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

procedure. At the first stage, the algorithms are evaluated on the

simulated dataset so that we can compare different algorithms based on
objective evaluation measures like Signal to Noise Ratio (SNR). At the
second stage, self-recorded real EEG signals should be used to visually
analyze the artifact removal performance of the best performing algo­
rithms at the last stage, the algorithms should be evaluated on a standard
large EEG database. We believe that this approach can be an ideal
protocol for the performance evaluation of artifact removal algorithms.
There is a ’need for a publicly available EEG artifact removal dataset’
that can be used as a benchmark for the evaluation of different artifact
removal algorithms. From our findings, Temple University Hospital’s
TUH EEG Artifact Corpus [63] is the largest public EEG artifact dataset
of real clinical EEG recordings with annotations of 5 different types of Fig. 3. To-Do List for Reproducible EEG Research.
artifacts and it has the potential to be used as a benchmark for future
studies. Recently [98], performed a benchmark study on classifying the with only 1% accuracy degradation. Similarly, Gul et al. [104] imple­
artifactual EEG signal to one of these 5 artifact types and this work can mented a fixed point real-time and online implementation of EMD al­
be used as a baseline for further work. However, their model only gorithms on FPGAs. However, these implementations are very specific to
classifies artifacts among one of five types and doesn’t repair them. a few variants and a few hardware. Future researchers should focus on
There is a need to evaluate EEG artifact removal algorithms on this making hardware optimized implementations of hybrid methods as they
dataset. are outperforming single methods in terms of accuracy and reliability
The EEG community should learn from the data science and machine and if we can implement them for real-time then they could also be used
learning research community who have put a lot of focus on ‘the open- in BCI and Neurofeedback applications like Robotic Arm Control and
source implementations of the recent articles and it is highly recom­ Wheelchair Control. One should also be careful that these hardware
mended, if not mandatory, for authors to share their code files. It is also optimized solutions are only desired in real-time EEG applications and
encouraged to make well-documented instructions for others who might are worth exploring only if the EEG artifact removal algorithm is
be interested in their work and might want to reuse their ideas for implemented in a real-time EEG application that eventually would lead
personal use. Fortunately, the latest publications in the EEG domain to commercial products.
often share their code and implementation details which is a good sign ‘Data Scarcity’ is the main concern with machine learning especially
for new researchers. Open Science Foundation (https://osf.io/) and deep learning-based EEG artifact removal algorithms. Researchers are
Github (https://github.com/) are among the popular tools used by re­ exploring options of transfer learning and data augmentation strategies
searchers to share their work. Similarly, there are some other initiatives to tackle this challenge. [105] provides an overview of different data
like https://www.paperswithcode.com/ that keep track of machine augmentation techniques for deep learning-based EEG analysis. There is
learning-related research articles whose code is open source and also an opportunity to further explore unsupervised or semi-supervised
“benchmark” performance of different algorithms on a common dataset. algorithms that don’t require a lot of training data as compared to su­
For example, this URL https://paperswithcode.com/sota/eeg-on-seed-iv pervised techniques. Consider the case of the Automated-ICA algorithm
shows the performance of 4 different papers that used the SEED-IV [93] that applies a supervised classifier on the results of the ICA algo­
dataset [99] and compared them with each other. Similarly, mother of rithm (i.e., Independent Components) which also need training data for
all BCI benchmarks (a.k.a moabb) [100] https://github.com/NeuroTech learning. However, Mur et al. [106] recently proposed unsupervised
X/moabb is another such initiative that is focused more on BCI related learning based automated and online ICA artifact removal method
experiments with a goal to build an open-source and comprehensive whose performance was comparable with other state-of-the-art EEG
benchmark of popular BCI algorithms on freely available BCI datasets. artifact removal algorithms and doesn’t require a training dataset.
So, a beginner may have an idea of which algorithm works best for a ‘High Computational Cost’ is also a big concern for deep learning-
dataset and there can be a consensus on how different BCI practitioners based methods that make them unsuitable for real-time applications.
report their results. However, recently a lot of work is done from a hardware-software
To have a ‘consistency in EEG data storage and data sharing protocols’, codesign perspective to optimize the performance of neural networks
the neuroimaging community has proposed “Brain Imaging Data by utilizing minimum resources in real-time constraints. Azghadi et al.
Structure” BIDS [101] that builds on the idea of the need for a single [107] work can be a great starting point to explore biomedical appli­
standard and intuitive way of organizing, storing, and sharing neuro­ cations of hardware-based neural network accelerators. Moreover, since
imaging data. The official website https://bids.neuroimaging.io/ has deep learning based methods do feature extraction automatically so they
some relevant resources and BIDS compatible datasets can be found on also shorten the overall pipeline of a typical ML system and so they can
OpenNeuro Website https://openneuro.org/. EEG-BIDS [102] is an still be a good choice depending upon the desired computational re­
extension of the BIDS concept for electroencephalographic (EEG) data quirements of the task at hand.
and it is encouraged by many researchers to adopt this convention while ‘Reproducibility’ is a hot research topic in the machine learning
storing and sharing EEG data. Moreover, famous software platforms like community and EEG researchers should follow their footsteps towards
MATLAB and Python also have libraries that can easily import data from making their research reproducible. In this regard, we found [108] to be
BIDS format and so researchers can focus on the analysis part instead of a useful article that has mentioned 10 not-to-do things for Machine
writing scripts for loading data. Learning researchers while reporting results and most of them are also
There is also a need for ‘hardware optimized EEG artifact removal al­ valid for EEG analysis. Fig. 3 provides a list of 10 ‘to-do things’ for the
gorithms’. Fortunately, in recent years, researchers have focused on this EEG community to address the reproducibility issues.
research area to make these algorithms suitable for real-time applica­
tions. Consider the work of Jafari et al. [97] who combined ICA with a
multi-instance learning approach to automate bad ICs selection and then 6.3. Popular EEG artifact removal libraries and toolboxes
improved the execution speed on embedded hardware from 282 s to 8 s
which is quite impressive. Kardon et al. [103] implemented FPGA based This subsection provides a brief account of the ‘python-based libraries
custom Canonical Correlation Analysis (CCA) engine for a BCI applica­ for automatic removal of artifacts’ for the EEG data. Python, being an
tion and decreased the inference time from microseconds to milliseconds open-source and general-purpose programming language, offers several

10
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

Table 6 B. Westover, J. Dauwels, Automated adult epilepsy diagnostic tool based on

Open Source Matlab-Based EEG Artifact Removal Toolboxes. interictal scalp electroencephalogram characteristics: a six-center study, Int. J.
Neural Syst. (2021), 2050074, https://doi.org/10.1142/S0129065720500744.
Study Methods [3] K. Rasheed, A. Qayyum, J. Qadir, S. Sivathamboo, P. Kwan, L. Kuhlmann,
T. O’Brien, A. Razi, Machine learning for predicting epileptic seizures using EEG
AAR toolbox Adaptive methods combining EOG as a regressor channel, and signals: a review, IEEE Rev. Biomed. Eng. 14 (2021) 139–155, https://doi.org/
[109] BSS methods 10.1109/RBME.2020.3008792.
ADJUST [110] ICA methods with objective criterion for selection of artifact- [4] D. Kapgate, Future of EEG based hybrid visual brain computer interface systems
related components in rehabilitation of people with neurological disorders, Int. Res. J. Adv. Sci. Hub.
DETECT [111] Employ Machine learning classification based on AR features that 2 (2020) 15–20, https://doi.org/10.47392/irjash.2020.31.
can discriminate the real EEG and from artifacts. [5] H. Raza, A. Chowdhury, S. Bhattacharyya, Deep learning based prediction of EEG
FASTER [112] such as the variance, mean correlation and spatial kurtosis to motor imagery of stroke patients’ for neuro-rehabilitation application, in: 2020
detect bad channels in the data. FASTER than applies Int. Jt. Conf. Neural Netw, IJCNN, IEEE, Glasgow, United Kingdom, 2020,
independent component analysis (ICA) pp. 1–8, https://doi.org/10.1109/IJCNN48605.2020.9206884.
CORRMAP the correlation of ICA inverse weights, and finds independent [6] M. Fatourechi, A. Bashashati, R.K. Ward, B. Ge, EMG and EOG artifacts in brain
computer interface systems: a survey, Clin. Neurophysiol. 118 (2007) 480–494.
[113] components that are like a user-defined template
[7] M.M.N. Mannan, M.A. Kamran, M.Y. Jeong, Identification and removal of
SASICA [114] SASICA is also a didactic tool that allows users to quickly
physiological artifacts from electroencephalogram signals: a review, IEEE Access
understand what signal features captured by ICs make them
6 (2018) 30630–30652, https://doi.org/10.1109/ACCESS.2018.2842082.
likely to reflect artifacts. [8] X. Gu, Z. Cao, A. Jolfaei, P. Xu, D. Wu, T.-P. Jung, C.-T. Lin, EEG-based Brain-
FORCe [115] Combines wavelet transform, ICA and thresholding to implement Computer Interfaces (BCIs): A Survey of Recent Studies on Signal Sensing
fully automated and online EEG artifact removal algorithm for Technologies and Computational Intelligence Approaches and Their Applications,
BCI applications 2020, pp. 1–22.
HEAR [116] Removes high variance electrode pop and drift artifacts from EEG [9] C.Q. Lai, H. Ibrahim, M.Z. Abdullah, J.M. Abdullah, S.A. Suandi, A. Azman,
data Literature Survey on Applications of Electroencephalography (EEG), 2018,
https://doi.org/10.1063/1.5055472. Penang, Malaysia, p. 020070.
[10] M. Rashid, N. Sulaiman, A.P.P. Abdul Majeed, R.M. Musa, A.F. Ahmad, B.S. Bari,
S. Khatun, Current Status, Challenges, and possible solutions of EEG-based brain-
libraries that are optimized for scientific computation. ‘Numpy’, ‘Scipy’,
computer interface: a comprehensive review, Front. Neurorobot. 14 (2020) 1–36,
and Matplotlib are backbones of the scientific community in Python. For https://doi.org/10.3389/fnbot.2020.00025.
machine learning and deep learning-based models, Python is the go-to [11] M.K. Islam, A. Rastegarnia, Z. Yang, Methods for artifact detection and removal
option with ‘PyTorch’ and ‘Tensorflow’ as the most popular choices. from scalp EEG: a review, Neurophysiol. Clin. Neurophysiol. 46 (2016) 287–305,
https://doi.org/10.1016/j.neucli.2016.07.002.
If we talk about EEG artifact removal techniques, then typically these [12] X. Jiang, G.-B. Bian, Z. Tian, Removal of artifacts from EEG signals: a review,
algorithms come as a submodule of a larger library. MNE-Python [91] is Sensors 19 (2019) 987, https://doi.org/10.3390/s19050987.
probably the most widely used library for EEG and MEG analysis. This [13] J.A. Urigüen, B. Garcia-Zapirain, EEG artifact removal—state-of-the-art and
guidelines, J. Neural Eng. 12 (2015), 031001, https://doi.org/10.1088/1741-
library has built-in methods for Regression, Independent Component 2560/12/3/031001.
Analysis (ICA) and Signal Space Projection (SSP) based EEG artifact [14] S. Kotte, J.R.K. Kumar Dabbakuti, Methods for removal of artifacts from EEG
removal. Similarly, ‘Scipy’ has the standard implementation of Principal signal: a review, J. Phys. Conf. Ser. 1706 (2020), 012093, https://doi.org/
10.1088/1742-6596/1706/1/012093.
Component Analysis (PCA). For deep learning-based models, the [15] A. Bisht, C. Kaur, P. Singh, Recent advances in artifact removal techniques for
braindecode library is also a very good choice that builds upon EEG signal processing, in: S. Choudhury, R. Mishra, R.G. Mishra, A. Kumar (Eds.),
MNE-Python so these two libraries can easily be integrated into a Intell. Commun. Control Devices, Springer, Singapore, Singapore, 2020,
pp. 385–392, https://doi.org/10.1007/978-981-13-8618-3_41.
project.
[16] N.P. Castellanos, V.A. Makarov, Recovering EEG brain signals: artifact
Although this may not be as exciting as MATLAB toolboxes, however, suppression with wavelet enhanced independent component analysis,
the most important thing is that a small Google/GitHub search with J. Neurosci. Methods 158 (2006) 300–312.
[17] D. Gajic, Z. Djurovic, J. Gligorijevic, S. Di Gennaro, I. Savic-Gajic, Detection of
“[artifact name] code in Python” would lead us to open-source imple­
epileptiform activity in EEG signals based on time-frequency and non-linear
mentations of these algorithms that can easily be used in our own pro­ analysis, Front. Comput. Neurosci. 9 (2015), https://doi.org/10.3389/
jects. Moreover, Python Package Index (PyPI) is a great online resource fncom.2015.00038.
for exploring different python libraries and using a keyword like “EEG” [18] D. Gajic, Z. Djurovic, S. Di Gennaro, F. Gustafsson, Classification of EEG signals
for detection of epileptic seizures based on wavelets and statistical pattern
or “BCI” would give us all the relevant libraries that were uploaded in recognition, Biomed. Eng. Appl. Basis Commun. 26 (2014), 1450021, https://doi.
the database. There are many other libraries as well related to EEG but org/10.4015/S1016237214500215.
here we have only mentioned the most popular ones. One may explore [19] J.S. Barlow, A. Remond, Eye movement artifact nulling in EEGs by multichannel
on-line EOG subtraction, Electroencephalogr. Clin. Neurophysiol. 52 (1981)
the ‘PyPI’ website https://pypi.org/ by himself to explore more open- 418–423.
source packages/libraries. [20] J.C. Corby, K. Bs, Differential contributions of blinks and vertical eye movements
It is also important to note that some of the MATLAB toolboxes also as artifacts in EEG recording, Psychophysiology 9 (1972) 640–644.
[21] C. Fortgens, M.P.D. Bruin, Removal of eye movement and ECG artifacts from the
have their equivalents in Python. Moreover, MNE also has a MATLAB non-cephalic reference EEG, Electroencephalogr. Clin. Neurophysiol. 56 (1983)
and C language interface. Which makes it useful to switch between these 90–96.
libraries. [22] D. Girton, J. Kamiya, A simple on-line technique for removing eye movement
artifacts from the EEG, Electroencephalogr. Clin. Neurophysiol. 34 (1973)
The ‘Matlab-based artifact removal toolboxes’ may help in expediting
212–216.
the implementation of various artifact reduction methods for new re­ [23] W. McCallum, W. Wg, The effects of attention and distraction on the contingent
searchers. Table 6 provides a summary table of different artifact removal negative variation in normal and neurotic subjects, Electroencephalogr. Clin.
Neurophysiol. 25 (1968) 319–329.
methods.
[24] T.W. Picton, S. Bentin, P. Berg, E. Donchin, S.A. Hillyard, R. Johnson, G.A. Miller,
W. Ritter, D.S. Ruchkin, M.D. Rugg, M.J. Taylor, Guidelines for using human
event-related potentials to study cognition: recording standards and publication
Declaration of Competing Interest criteria, Psychophysiology 37 (2000) 127–152.
[25] R.J. Croft, R.J. Barry, EOG correction of blinks with saccade coefficients: a test
and revision of the aligned-artefact average solution, Clin. Neurophysiol. 111
The authors report no declarations of interest.
(2000) 444–451.
[26] T. Gasser, L. Sroka, J. Möcks, The correction of EOG artifacts by frequency
References dependent and frequency independent methods, Psychophysiology 23 (1986)
704–712.
[27] D. Hagemann, E. Naumann, The effects of ocular artifacts on (lateralized)
[1] Y. Park, S.-H. Han, W. Byun, J.-H. Kim, H.-C. Lee, S.-J. Kim, A real-time depth of
broadband power in the EEG, Clin. Neurophysiol. 112 (2001) 215–231.
anesthesia monitoring system based on deep neural network with large EDO
[28] S.A. Hillyard, R. Galambos, Eye movement artifact in the CNV,
tolerant EEG analog front-end, IEEE Trans. Biomed. Circuits Syst. 14 (2020)
Electroencephalogr. Clin. Neurophysiol. 28 (1970) 173–182.
825–837, https://doi.org/10.1109/TBCAS.2020.2998172.
[2] J. Thomas, P. Thangavel, W.Y. Peh, J. Jing, R. Yuvaraj, S.S. Cash, R. Chaudhari,
S. Karia, R. Rathakrishnan, V. Saini, N. Shah, R. Srivastava, Y.-L. Tan,

11
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

[29] J.L. Kenemans, P. Molenaar, M.N. Verbaten, S. Jl, Removal of the ocular artifact [56] M.R. Mowla, S.-C. Ng, M.S. Zilany, R. Paramesran, Artifacts-matched blind source
from the EEG: a comparison of time and frequency domain methods with separation and wavelet transform for multichannel EEG denoising, Biomed.
simulated and real data, Psychophysiology 28 (1991) 114–121. Signal Process. Control 22 (2015) 111–118.
[30] P.M. Quilter, B. MacGillivray, D. Wadbrook, The removal of eye movement [57] V. Kamath, Y.-C. Lai, L. Zhu, S. Urval, Empirical mode decomposition and blind
artefact from EEG signals using correlation techniques, Random Signal Anal. IEEE source separation methods for antijamming with GPS signals, Proc. IEEEION
Conf. Publ. 159 (1977) 93–100. PLANS 2006 (2006) 335–341.
[31] A. Schlögl, C. Keinrath, D. Zimmermann, R. Scherer, R. Leeb, G. Pfurtscheller, [58] S. Halder, M. Bensch, J. Mellinger, M. Bogdan, A. Kübler, N. Birbaumer,
A fully automated correction method of EOG artifacts in EEG recordings, Clin. W. Rosenstiel, Online artifact removal for brain-computer interfaces using
Neurophysiol. 118 (2007) 98–104. support vector machines and blind source separation, Comput. Intell. Neurosci.
[32] J.L. Whitton, F. Lue, H. Moldofsky, A spectral method for removing eye 2007 (2007).
movement artifacts from the EEG, Electroencephalogr. Clin. Neurophysiol. 44 [59] G. Bartels, L.-C. Shi, B.-L. Lu, Automatic artifact removal from EEG-a mixed
(1978) 735–741. approach based on double blind source separation and support vector machine,
[33] J. Woestenburg, M. Verbaten, J. Slangen, The removal of the eye-movement in: 2010 Annu. Int. Conf. IEEE Eng. Med. Biol., IEEE, 2010, pp. 5383–5386.
artifact from the EEG by regression analysis in the frequency domain, Biol. [60] G.L. Wallstrom, R.E. Kass, A. Miller, J.F. Cohn, F. Na, Automatic correction of
Psychol. 16 (1983) 127–147. ocular artifacts in the EEG: a comparison of regression-based and component-
[34] C. Beach, M. Li, E. Balaban, A. Casson, Motion Artefact Removal in based methods, Int. J. Psychophysiol. 53 (2004) 105–119.
Electroencephalography and Electrocardiography by Using Multichannel Inertial [61] M. Klados, A Semi-simulated EEG/EOG Dataset for the Comparison of EOG
Measurement Units and Adaptive Filtering, 2021. Artifact Rejection Techniques, 2019, https://doi.org/10.17632/WB6YVR725D.4.
[35] S. Selvan, R. Srinivasan, Removal of ocular artifacts from EEG using an efficient [62] Reinmar Kobler, Andreea Sburlea, Catarina Dias, Andreas Schwarz,
neural network based adaptive filtering technique, IEEE Signal Process. Lett. 6 Valeria Mondini, Gernot Müller-Putz, EEG Eye Artifact Dataset, 2020, https://doi.
(1999) 330–332. org/10.17605/OSF.IO/2QGRD.
[36] P. He, G. Wilson, C. Russell, Removal of ocular artifacts from electro- [63] A. Hamid, K. Gagliano, S. Rahman, N. Tulin, V. Tchiong, I. Obeid, J. Picone, The
encephalogram by adaptive filtering, Med. Biol. Eng. Comput. 42 (2004) Temple University Artifact Corpus: An Annotated Corpus of EEG Artifacts, IEEE
407–412. Signal Process. Med. Biol. Symp. SPMB. 1 (n.d.). https://par.nsf.gov/biblio
[37] T.-P. Jung, S. Makeig, A.J. Bell, T.J. Sejnowski, Independent component analysis /10199675.
of electroencephalographic and event-related potential data. Cent. Audit. Process. [64] N. Carlini, C. Liu, Ú. Erlingsson, J. Kos, D. Song, The Secret Sharer: Evaluating
Neural Model, Springer, 1998, pp. 189–197. and Testing Unintended Memorization in Neural Networks, 2019.
[38] S. Makeig, A.J. Bell, T.-P. Jung, T.J. Sejnowski, Independent component analysis ArXiv180208232 Cs. (Accessed 19 November 2020), http://arxiv.org/abs/1
of electroencephalographic data, Adv. Neural Inf. Process. Syst. (1996) 145–151. 802.08232.
[39] S. Makeig, T.-P. Jung, D. Ghahremani, T.J. Sejnowski, Independent Component [65] C. Song, T. Ristenpart, V. Shmatikov, Machine learning models that remember too
Analysis of Simulated ERP Data, Institute for Neural Computation, University of Much, in: Proc. 2017 ACM SIGSAC Conf. Comput. Commun. Secur., ACM, Dallas
California, 1996. Texas USA, 2017, pp. 587–601, https://doi.org/10.1145/3133956.3134077.
[40] W. De Clercq, A. Vergult, B. Vanrumste, W. Van Paesschen, S. Van Huffel, [66] J. Zhang, C. Li, Adversarial examples: opportunities and challenges, IEEE Trans.
Canonical correlation analysis applied to remove muscle artifacts from the Neural Netw. Learn. Syst. (2019) 1–16, https://doi.org/10.1109/
electroencephalogram, IEEE Trans. Biomed. Eng. 53 (2006) 2583–2587. TNNLS.2019.2933524.
[41] W. De Clercq, A. Vergult, B. Vanrumste, J. Van Hees, A. Palmini, W. Van [67] B. Abdi-Sargezeh, R. Foodeh, V. Shalchyan, M.R. Daliri, EEG artifact rejection by
Paesschen, S.A. Van Huffel, New muscle artifact removal technique to improve extracting spatial and spatio-spectral common components, J. Neurosci. Methods
the interpretation of the ictal scalp electroencephalogram, 2005 IEEE Eng. Med. (2021), 109182, https://doi.org/10.1016/j.jneumeth.2021.109182.
Biol. 27th Annu. Conf. (2006) 944–947. [68] Q. Chen, Y. Li, X. Yuan, A hybrid method for muscle artifact removal from EEG
[42] J. Gao, C. Zheng, P. Wang, Online removal of muscle artifact from signals, J. Neurosci. Methods 353 (2021), 109104, https://doi.org/10.1016/j.
electroencephalogram signals based on canonical correlation analysis, Clin. EEG jneumeth.2021.109104.
Neurosci. 41 (2010) 53–59. [69] N. Bajaj, J. Requena Carrión, F. Bellotti, R. Berta, A. De Gloria, Automatic and
[43] D.M. Vos, S. Riès, K. Vanderperren, B. Vanrumste, F.-X. Alario, V.S. Huffel, tunable algorithm for EEG artifact removal using wavelet decomposition with
B. Burle, Removal of muscle artifacts from EEG recordings of spoken language applications in predictive modeling during auditory tasks, Biomed. Signal
production, Neuroinformatics 8 (2010) 135–150. Process. Control 55 (2020), 101624, https://doi.org/10.1016/j.
[44] A. Vergult, W.D. Clercq, A. Palmini, B. Vanrumste, P. Dupont, S.V. Huffel, W. bspc.2019.101624.
V. Paesschen, Improving the interpretation of ictal scalp EEG: BSS–CCA algorithm [70] M. Val-Calvo, J.R. Álvarez-Sánchez, J.M. Ferrández-Vicente, E. Fernández,
for muscle artifact removal, Epilepsia 48 (2007) 950–958. Optimization of real-time EEG artifact removal and emotion estimation for
[45] O.G. Lins, T.W. Picton, P. Berg, M. Scherg, Ocular artifacts in recording EEGs and human-robot interaction applications, Front. Comput. Neurosci. 13 (2019) 80,
event-related potentials II: source dipoles and source components, Brain Topogr. https://doi.org/10.3389/fncom.2019.00080.
6 (1993) 65–78. [71] S. Blum, N.S.J. Jacobsen, M.G. Bleichner, S. Debener, A riemannian modification
[46] S. Casarotto, A.M. Bianchi, S. Cerutti, C. Ga, Principal component analysis for of artifact subspace reconstruction for EEG artifact handling, Front. Hum.
reduction of ocular artefacts in event-related potentials of normal and dyslexic Neurosci. 13 (2019) 141, https://doi.org/10.3389/fnhum.2019.00141.
children, Clin. Neurophysiol. 115 (2004) 609–619. [72] M. Plechawska-Wojcik, M. Kaczorowska, D. Zapala, The Artifact Subspace
[47] T. Zikov, S. Bibian, G.A. Dumont, M. Huzmezan, C.R. Ries, A wavelet based de- Reconstruction (ASR) for EEG signal correction. A comparative study, in:
noising technique for ocular artifact correction of the electroencephalogram, in: J. Świątek, L. Borzemski, Z. Wilimowska (Eds.), Inf. Syst. Archit. Technol. Proc.
IEEEProc. Second Jt. 24th Annu. Conf. Annu. Fall Meet. Biomed. Eng. Soc. Med. 39th Int. Conf. Inf. Syst. Archit. Technol. – ISAT 2018, Springer International
Biol., vol. 1, 2002, pp. 98–105. Publishing, Cham, 2019, pp. 125–135, https://doi.org/10.1007/978-3-319-
[48] V. Krishnaveni, S. Jayaraman, N. Malmurugan, A. Kandasamy, D. Ramadoss, Non 99996-8_12.
adaptive thresholding methods for correcting ocular artifacts in EEG, Acad. Open [73] M. Chavez, F. Grosselin, A. Bussalb, F.D.V. Fallani, X. Navarro-Sune, Surrogate-
Internet J. 13 (2004). based artifact removal from single-channel EEG, IEEE Trans. Neural Syst. Rehabil.
[49] V. Krishnaveni, S. Jayaraman, L. Anitha, K. Ramadoss, Removal of ocular artifacts Eng. 26 (2018) 540–550.
from EEG using adaptive thresholding of wavelet coefficients, J. Neural Eng. 3 [74] R. Guarnieri, M. Marino, F. Barban, M. Ganzetti, D. Mantini, Online EEG artifact
(2006) 338. removal for BCI applications by adaptive spatial filtering, J. Neural Eng. 15
[50] P.S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, Removal of ocular (2018), 056009, https://doi.org/10.1088/1741-2552/aacfdf.
artifacts in the EEG through wavelet transform without using an EOG reference [75] B. Somers, T. Francart, A. Bertrand, A generic EEG artifact removal algorithm
channel, Int. J. Open Probl. Comput. Math. 1 (2008) 188–200. based on the multi-channel Wiener filter, J. Neural Eng. 15 (2018), 036007,
[51] P.A. Bizopoulos, T. Al-Ani, D.G. Tsalikakis, A.T. Tzallas, D.D. Koutsouris, D. https://doi.org/10.1088/1741-2552/aaac92.
I. Fotiadis, An automatic electroencephalography blinking artefact detection and [76] L. Frølich, I. Dowding, Removal of muscular artifacts in EEG signals: a
removal method based on template matching and ensemble empirical mode comparison of linear decomposition methods, Brain Inform. 5 (2018) 13–22,
decomposition, in: 2013 35th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBC, https://doi.org/10.1007/s40708-017-0074-6.
IEEE, 2013, pp. 5853–5856. [77] Kobler, Reinmar J., Sburlea, Andreea I., Mller-Putz, Gernot R., A Comparison Of
[52] S. Romero, M.A. Mañanas, M.J. Barbanoj, Ocular reduction in EEG signals based Ocular Artifact Removal Methods For Block Design Based
on adaptive filtering, regression and blind source separation, Ann. Biomed. Eng. Electroencephalography Experiments, (n.d.). https://doi.org/10.3217/978-3-85
37 (2009) 176–191. 125-533-1-44.
[53] P.S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, Removal of artifacts [78] M. Mohammadpour, V. Rahmani, A hidden markov model-based approach to
from EEG signals using adaptive filter through wavelet transform, in: 2008 9th removing EEG artifact, in: 2017 5th Iran. Jt. Congr. Fuzzy Intell. Syst. CFIS, IEEE,
Int. Conf. Signal Process., IEEE, 2008, pp. 2138–2141. Qazvin, Iran, 2017, pp. 46–49, https://doi.org/10.1109/CFIS.2017.8003655.
[54] P.S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, An adaptive method to [79] S. Zhang, J. McIntosh, S.M. Shadli, P.S.-H. Neo, Z. Huang, N. McNaughton,
remove ocular artifacts from EEG signals using wavelet transform, J. Appl. Sci. Removing eye blink artefacts from EEG—a single-channel physiology-based
Res. 5 (2009) 711–745. method, J. Neurosci. Methods 291 (2017) 213–220, https://doi.org/10.1016/j.
[55] X. Navarro, F. Porée, G. Carrault, ECG removal in preterm EEG combining jneumeth.2017.08.031.
empirical mode decomposition and adaptive filtering, in: 2012 IEEE Int. Conf. [80] A. Kilicarslan, R.G. Grossman, C.-V. Jl, A robust adaptive denoising framework
Acoust. Speech Signal Process, ICASSP, IEEE, 2012, pp. 661–664. for real-time artifact removal in scalp EEG measurements, J. Neural Eng. 13
(2016), 026013.

12
W. Mumtaz et al. Biomedical Signal Processing and Control 68 (2021) 102741

[81] N.T. Haumann, L. Parkkonen, M. Kliuchko, P. Vuust, E. Brattico, Comparing the [98] S. Roy, Machine Learning for Removing EEG Artifacts: Setting the Benchmark,
performance of popular MEG/EEG artifact correction methods in an evoked- 2019, pp. 5–6.
response study, Comput. Intell. Neurosci. 2016 (2016) 1–10, https://doi.org/ [99] W.-L. Zheng, W. Liu, Y. Lu, B.-L. Lu, A. Cichocki, EmotionMeter: A Multimodal
10.1155/2016/7489108. Framework for Recognizing Human Emotions, IEEE Trans. Cybern. 49 (2019)
[82] C. Damon, A. Liutkus, A. Gramfort, S. Essid, Non-negative matrix factorization for 1110–1122, https://doi.org/10.1109/TCYB.2018.2797176.
single-channel EEG artifact rejection, in: 2013 IEEE Int. Conf. Acoust. Speech [100] V. Jayaram, A. Barachant, MOABB: trustworthy algorithm benchmarking for
Signal Process., IEEE, Vancouver, BC, Canada, 2013, pp. 1177–1181, https://doi. BCIs, J. Neural Eng. 15 (2018), 066011, https://doi.org/10.1088/1741-2552/
org/10.1109/ICASSP.2013.6637836. aadea0.
[83] D. Safieddine, A. Kachenoura, L. Albera, G. Birot, A. Karfoul, A. Pasnicu, [101] K.J. Gorgolewski, T. Auer, V.D. Calhoun, R.C. Craddock, S. Das, E.P. Duff,
A. Biraben, F. Wendling, L. Senhadji, I. Merlet, Removal of muscle artifact from G. Flandin, S.S. Ghosh, T. Glatard, Y.O. Halchenko, D.A. Handwerker, M. Hanke,
EEG data: comparison between stochastic (ICA and CCA) and deterministic (EMD D. Keator, X. Li, Z. Michael, C. Maumet, B.N. Nichols, T.E. Nichols, J. Pellman, J.-
and wavelet-based) approaches, EURASIP J. Adv. Signal Process. 2012 (2012) B. Poline, A. Rokem, G. Schaefer, V. Sochat, W. Triplett, J.A. Turner,
127, https://doi.org/10.1186/1687-6180-2012-127. G. Varoquaux, R.A. Poldrack, The brain imaging data structure, a format for
[84] E. Kroupi, A. Yazdani, J.-M. Vesin, T. Ebrahimi, Ocular artifact removal from organizing and describing outputs of neuroimaging experiments, Sci. Data 3
EEG: a comparison of subspace projection and adaptive filtering methods, in: (2016), 160044, https://doi.org/10.1038/sdata.2016.44.
2011 19th Eur. Signal Process. Conf., IEEE, 2011, pp. 1395–1399. [102] C.R. Pernet, S. Appelhoff, K.J. Gorgolewski, G. Flandin, C. Phillips, A. Delorme,
[85] M.A. Klados, C. Papadelis, C.D. Lithari, P.D. Bamidis, The removal of ocular R. Oostenveld, EEG-BIDS, an extension to the brain imaging data structure for
artifacts from EEG signals: a comparison of performances for different methods, electroencephalography, Sci. Data 6 (2019) 103, https://doi.org/10.1038/
in: J. Vander Sloten, P. Verdonck, M. Nyssen, J. Haueisen (Eds.), 4th Eur. Conf. s41597-019-0104-8.
Int. Fed. Med. Biol. Eng., Springer, Berlin Heidelberg, Berlin, Heidelberg, 2009, [103] R. Karkon, S.M. Reza Shahshahani, H.R. Mahdiani, A Custom Hardware CCA
pp. 1259–1263, https://doi.org/10.1007/978-3-540-89208-3_300. Engine for Real-time SSVEP-based BCI Applications, 2020, pp. 1–6, https://doi.
[86] M. Crespo-Garcia, M. Atienza, Jose L. Cantero, Muscle artifact removal from org/10.1109/cads50570.2020.9211863.
human sleep EEG by using independent component analysis, Ann. Biomed. Eng. [104] S. Gul, M.F. Siddiqui, N. Ur Rehman, FPGA based real-time implementation of
36 (2008) 467–475. online EMD with fixed point architecture, IEEE Access 7 (2019) 176565–176577,
[87] P. He, G. Wilson, C. Russell, M. Gerschutz, Removal of ocular artifacts from the https://doi.org/10.1109/ACCESS.2019.2957819.
EEG: a comparison between time-domain regression method and adaptive [105] E. Lashgari, D. Liang, U. Maoz, Data augmentation for deep-learning-based
filtering method using simulated data, Med. Biol. Eng. Comput. 45 (2007) electroencephalography, J. Neurosci. Methods 346 (2020), 108885, https://doi.
495–503, https://doi.org/10.1007/s11517-007-0179-9. org/10.1016/j.jneumeth.2020.108885.
[88] T.-P. Jung, C. Humphries, T.-W. Lee, S. Makeig, M.J. McKeown, V. Iragui, T. [106] A. Mur, R. Dormido, N. Duro, An unsupervised method for artefact removal in
J. Sejnowski, Removing electroencephalographic artifacts: comparison between EEG signals, Sensors 19 (2019) 2302, https://doi.org/10.3390/s19102302.
ICA and PCA, Neural Netw. Signal Process. VIII (1998) 63–72. [107] M.R. Azghadi, C. Lammie, J.K. Eshraghian, M. Payvand, E. Donati, B. Linares-
[89] T. Gasser, P. Ziegler, W.F. Gattaz, The deleterious effect of ocular artefacts on the Barranco, G. Indiveri, Hardware Implementation of Deep Network Accelerators
quantitative EEG, and a remedy, Eur. Arch. Psychiatry Clin. Neurosci. 241 (1992) Towards Healthcare and Biomedical Applications, 1, 2020, pp. 1–21.
352–356. [108] F. Minhas, A. Asif, A. Ben-Hur, Ten Ways to Fool the Masses With Machine
[90] G.L. Wallstrom, R.E. Kass, A. Miller, J.F. Cohn, N.A. Fox, Correction of ocular Learning, ArXiv190101686 Cs Stat, 2019. http://arxiv.org/abs/1901.01686.
artifacts in the EEG using bayesian adaptive regression splines, in: C. Gatsonis, R. [109] G. Gómez-Herrero, Automatic Artifact Removal (AAR) Toolbox v1. 3 (Release
E. Kass, A. Carriquiry, A. Gelman, D. Higdon, D.K. Pauler, I. Verdinelli (Eds.), 09.12. 2007) for MATLAB, Tampere University of Technology, 2007.
Case Stud. Bayesian Stat., Springer, New York, New York, NY, 2002, pp. 351–365, [110] A. Mognon, J. Jovicich, L. Bruzzone, M. Buiatti, ADJUST: an automatic EEG
https://doi.org/10.1007/978-1-4612-2078-7_19. artifact detector based on the joint use of spatial and temporal features,
[91] A. Gramfort, MEG and EEG data analysis with MNE-Python, Front. Neurosci. 7 Psychophysiology 48 (2011) 229–240.
(2013), https://doi.org/10.3389/fnins.2013.00267. [111] V. Lawhern, W.D. Hairston, K. Robbins, DETECT: a MATLAB toolbox for event
[92] S. Halder, M. Bensch, J. Mellinger, M. Bogdan, A. Kübler, N. Birbaumer, detection and identification in time series, with applications to artifact detection
W. Rosenstiel, Online artifact removal for brain-computer interfaces using in EEG signals, PLoS One 8 (2013) e62944.
support vector machines and blind source separation, Comput. Intell. Neurosci. [112] H. Nolan, R. Whelan, R. Reilly, FASTER: fully automated statistical thresholding
2007 (2007). for EEG artifact rejection, J. Neurosci. Methods 192 (2010) 152–162.
[93] T. Radüntz, J. Scouten, O. Hochmuth, B. Meffert, Automated EEG artifact [113] F.C. Viola, J. Thorne, B. Edmonds, T. Schneider, T. Eichele, S. Debener, Semi-
elimination by applying machine learning algorithms to ICA-based features, automatic identification of independent components representing EEG artifact,
J. Neural Eng. 14 (2017), 046004, https://doi.org/10.1088/1741-2552/aa69d1. Clin. Neurophysiol. 120 (2009) 868–877.
[94] N. Mammone, F. La Foresta, F.C. Morabito, Automatic artifact rejection from [114] M. Chaumon, D.V. Bishop, B. Na, A practical guide to the selection of independent
multichannel scalp EEG by wavelet ICA, IEEE Sens. J. 12 (2012) 533–542, components of the electroencephalogram for artifact correction, J. Neurosci.
https://doi.org/10.1109/JSEN.2011.2115236. Methods 250 (2015) 47–63.
[95] C.Y. Sai, N. Mokhtar, H. Arof, P. Cumming, M. Iwahashi, Automated classification [115] I. Daly, R. Scherer, M. Billinger, G. Muller-Putz, FORCe: fully online and
and removal of EEG artifacts with SVM and Wavelet-ICA, IEEE J. Biomed. Health automated artifact removal for brain-computer interfacing, IEEE Trans. Neural
Inform. 22 (2018) 664–670, https://doi.org/10.1109/JBHI.2017.2723420. Syst. Rehabil. Eng. 23 (2015) 725–736, https://doi.org/10.1109/
[96] S. López, I. Obeid, J. Picone, Automated Interpretation of Abnormal Adult TNSRE.2014.2346621.
Electroencephalograms, Temple Univ, 2017. MS Thesis. [116] R.J. Kobler, A.I. Sburlea, V. Mondini, G.R. Müller-Putz, HEAR To Remove Pops
[97] A. Jafari, S. Gandhi, S.H. Konuru, W. David Hairston, T. Oates, T. Mohsenin, An and Drifts: the High-variance Electrode Artifact Removal (HEAR) Algorithm,
EEG artifact identification embedded system using ICA and multi-instance ArXiv190712354 Eess, 2019. http://arxiv.org/abs/1907.12354.
learning, Proc. - IEEE Int. Symp. Circuits Syst. (2017) 1–4, https://doi.org/
10.1109/ISCAS.2017.8050346.

13