Chapter 2

Signal Artifacts and Techniques

for Artifacts and Noise Removal

Md. Kafiul Islam, Amir Rastegarnia, and Saeid Sanei

Abstract Biosignals have quite low signal-to-noise ratio and are often corrupted
by different types of artifacts and noises originated from both external and internal
sources. The presence of such artifacts and noises poses a great challenge in proper
analysis of the recorded signals and thus useful information extraction or classifi-
cation in the subsequent stages becomes erroneous. This eventually results either
in a wrong diagnosis of the diseases or misleading the feedback associated with
such biosignal-based systems. Brain-Computer Interfaces (BCIs) and neural pros-
theses are among the popular ones. There have been many signal processing-based
algorithms proposed in the literature for reliable identification and removal of such
artifacts from the biosignal recordings. The purpose of this chapter is to introduce
different sources of artifacts and noises present in biosignal recordings, such as
EEG, ECG, and EMG, describe how the artifact characteristics are different from
signal-of-interest, and systematically analyze the state-of-the-art signal processing
techniques for reliable identification of these offending artifacts and finally removing
them from the raw recordings without distorting the signal-of-interest. The analysis
of the biosignal recordings in time, frequency and tensor domains is of major interest.
In addition, the impact of artifact and noise removal is examined for BCI and clinical
diagnostic applications. Since most biosignals are recorded in low sampling rate,
the noise removal algorithms can be often applied in real time. In the case of tensor
domain systems, more care has to be taken to comply with real time applications.
Therefore, in the final part of this chapter, both quantitative and qualitative measures

Md. K. Islam (B)

Department of Electrical and Electronic Engineering, Independent University, Bangladesh,
Dhaka, Bangladesh
e-mail: [email protected]
A. Rastegarnia
Department of Electrical Engineering, Malayer University, Malayer 65719-95863, Iran
e-mail: [email protected]
S. Sanei
School of Science and Technology, Nottingham Trent University, Nottingham NG11 8NS, UK
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 23

M. A. R. Ahad and M. U. Ahmed (eds.), Signal Processing Techniques for Computational
Health Informatics, Intelligent Systems Reference Library 192,
https://doi.org/10.1007/978-3-030-54932-9_2
24 Md. K. Islam et al.

are demonstrated in tables and the algorithms are assessed in terms of their computa-
tional complexity and cost. It is also shown that availability of some a priori clinical
or statistical information can boost the algorithm performance in many cases.

Keywords Artifact · Biosignal · ECG · EEG · Neural signal · Noise, etc.

2.1 Introduction

2.1.1 Background and Motivation

Human body is composed of several complex systems including nervous, cardio-

vascular and musculoskeletal systems. Each system has a particular structure and
carries its own physiological, functional and pathological processes. These complex
biological systems are dependent on each other and the processes involved are often
considered as non-linear, nonstationary, and stochastic process. The resultant biosig-
nals generated from these complex biological processes can be recorded in both inva-
sive and non-invasive ways. The signals recorded by non-invasive electrodes often
have extremely low amplitudes (ranging from µV to mV) due to the attenuation
by various body tissues. In addition, due to the non-invasive nature, the recordings
are more prone to many external noise sources such as artifacts and interferences
resulting in low signal-to-noise ratios (SNRs). Thus, the acquired biosignals and
their associated important clinical/biological events are often submerged under noise
and required to be processed properly by removing such artifacts and interferences
before any further analysis and decision can be made. However, often traditional
signal processing techniques, e.g., digital filtering, are not suitable enough to effec-
tively remove such artifacts. Therefore, advanced signal processing techniques have
been proposed in the literature for this purpose. Figure 2.1 illustrates the general
steps involved in a typical biosignal processing system from signal acquisition until
classification.

2.1.2 Objectives

The objectives of this chapter are as follows:

Fig. 2.1 Typical process flow of a biosignal processing system. Only artifact detection and removal
is covered in this chapter
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 25

• Identify biosignal characteristics and their applications,

• Identify artifact sources and their characteristics,
• Compare and identify suitable signal analysis and processing techniques for
artifact detection and removal,
• Demonstrate the influence of reliable artifact detection and removal on the later-
stage detection or classification of clinically significant biological events.

2.2 Biosignals and Artifact/Noise Modeling

and Characterization

2.2.1 Different Types of Biosignals

Four major types of biosignals including their properties is shown in Fig. 2.2.

2.2.1.1 Biosignals Generated by Cardiac Activity

Electrocardiography

Electrocardiography (ECG aka EKG), shows the electrical activity of heart over a
period of time. ECG is recorded via the electrodes that are placed on the chest.
Small changes in electrical potentials are picked up by these electrodes due to the
cardiac muscle’s electrophysiological pattern of depolarizing and repolarizing during
each heartbeat. It generally is composed of QRS complexes, P waves and T waves.
Figure 2.3 shows an ECG signal of a heart in normal sinus or regular rhythm, at a
heart rate of 60–100 beats per minutes.

Fig. 2.2 Typical

characteristics of some
commonly used biosignals in
terms of amplitude and
frequency range
26 Md. K. Islam et al.

Fig. 2.3 A sample ECG recording with its characteristics points (P-wave, QRS complex) during a
cardiac cycle of a normal sinus rhythm

Seismocardiography

Seismocardiography (SCG) is the non-intrusive estimation of cardiovascular vibra-

tions transmitted to the chest divider by the heart and usually recorded by an
accelerometer. SCG can provide information of all the activities during cardiac cycle
which may not be found only from ECG recordings. Simultaneous recordings of
both ECG and SCG may reveal a lot more diagnostic information related to heart
diseases. An example SCG recording is given in Fig. 2.4.

2.2.1.2 Biosignals Generated by Muscle Activity

EMG

Electromyography (EMG) is the process of recording electrical potential generated

from muscle cells during contraction. Electromyogram is the combined action poten-
tials of the muscle cells of muscle tissue. The amplitude of an EMG signal lies
between 0.01 and 1 mV, and its frequency range is 20–1000 Hz. If this signal is
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 27

Fig. 2.4 An example of SCG recording during one complete cardiac cycle with annotation proposed
by Salerno [1]. Where AS, MC, IM, AO, IC, RE, AC, MO, RF refer to atrial systole, mitral value
closure, isovolumic movement, aortic valve opening, isovolumic contraction, rapid systolic ejection,
aortic valve closure, mitral valve opening and rapid diastolic filling respectively. X-axis and Y-axis
referring to the signal amplitude (in mV) and time (in second) respectively

detected from the surface of the skin, it will be the superposition of messages from
all the muscles underneath. A sample raw EMG recording is given in Fig. 2.5.

2.2.1.3 Biosignals Generated by Ocular Activity

Electrooculography

Electrooculogram (EOG) signal is generated due to the potential difference between

retina and cornea of the eye which is modeled as an electrical dipole that moves
with eyeball movement. The dynamic range of EOG signal is typically 0.05–3.5 mV
(peak to peak). The bandwidth of EOG is between 0 and 1000 Hz; however, maximum
usable energy of EOG signal lies between 0.1 and 40 Hz [2]. A sample raw EOG
recording consisting of both horizontal and vertical eye movements is illustrated in
Fig. 2.6.

Eye Blinks

Eye blinks are common ocular artifacts that are found in EEG signals which is due to
the blinking of eyes (both voluntary and involuntary). During eye blinks, movement
28 Md. K. Islam et al.

Fig. 2.5 Sample raw EMG signal recorded from arm (after analog amplification) using Ag-AgCl
disposable electrodes

Fig. 2.6 Sample EOG signal recording with disposable electrodes during horizontal (left plot) and
vertical (right plot) movement of the eyeball [2]. X-axis and Y-axis referring to the normalized
signal amplitude and time (in second) respectively

of eyelid muscles generates such potentials which are counted as artifacts to the
neural signal recordings, such as EEG.

2.2.1.4 Biosignals Generated by the Brain

Different biosignals generated due to neural activities in the brain are mainly clas-
sified based on their location of recording (i.e. placement of recording electrodes)
which is shown in Fig. 2.7 while their characteristics in terms of amplitude and
frequency is shown in the Table 2.1.
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 29

Table 2.1 Different types of

Signal type Frequency range (Hz) Amplitude range (mV pp )
neural signals and their
characteristics EEG 0.05–128 0.02–0.4
ECoG 0.1–64 0.02–0.1
LFP 0.1–100 0.1–1
EAP 100–10 k 0.04–0.2
IAP 100–10 k ~100

Fig. 2.7 Different types of

brain recordings based on the
placement of
electrodes/sensors. Courtesy
http://www.frontiersin.org/
files/Articles/103134/fnsys-
08-00144-HTML/image_t/
fnsys-08-00144-g003.gif

Non-invasive Brain Recordings

The most popular non-invasive brain recording technique is Electroencephalog-

raphy (EEG) that measures the integrated electrical activities produced by billions of
neurons in the brain by placing electrodes on the scalp. It is the most commonly used
brain recording technique for diagnosis of different neurological disorders along with
other applications such as brain-computer interface and basic neuroscience research.
The EEG recordings are described in terms of rhythms and transients while the
rhythmic activity of EEG is composed of non-overlapping bands of frequency. The
most common EEG rhythms are Delta (0–4 Hz), Theta (4–8 Hz), Alpha (8–12 Hz) and
Beta (12–30 Hz) waves. Recently a relatively high frequency Gamma wave (>30 Hz)
is also considered. On the other hand, artifacts are transient events, although epilepsy
seizure events can also be transient but they are more oscillatory than artifacts [3].
To provide a model for recorded raw EEG data, let’s denote the clean EEG back-
ground activity/rhythm as Ec with weight wc ; Artifact event as A Tn with weight wTn
and time delay τTn ; where n = 0, 1, . . . , N denotes the type of artifact. E.g. if it is
a Type-1 artifact [4], then denoted by A T1 with weight wT1 and time delay τT1 . Now
the recorded raw EEG signal is usually modeled as the linear combination of these
two signal components.

N

(2.1)
 
E R (t) = wc E c (t) + wTn A Tn t − τTn
n=1
30 Md. K. Islam et al.

Semi-invasive Brain Recordings

Electro-cortiography (ECoG) or intracranial EEG (iEEG) refers to measuring brain

signals from the surface of the brain after opening the skull. The temporal resolution
of ECoG is better than scalp-EEG but as it requires brain surgery to open up the
skull, often it is discouraged in human subject.

Fully Invasive Brain Recordings

Local Field Potentials

Extracellular local field potentials (LFP) are produced by the collective and
simultaneous activity of many nearby neurons by synaptic transmission.

Neural Action Potentials (Neural Spikes)

An action potential or neural spike is a short-lasting event (usually 2–3 ms) in which
the electrical membrane potential of a neuron rapidly rises (i.e. depolarization) and
falls (repolarization), following a consistent trajectory. Such activity gives rise to a
specific waveform shape known as action potential. The extracellular action poten-
tials are typically about 100 µV–1 mV), smaller than an intracellular action potential.
Microelectrodes (with a tip size of approx. 1 µm) are usually implanted into the brain
of a living animal to detect such electrical activity generated by the nearby neurons
which is known as ’single-unit’ recording. Such recordings of single neurons in living
animals can be used to understand the process of information the brain.

2.2.2 Different Sources of Artifacts and Noises

Artifacts can originate from both external and internal sources. Internal sources of
artifacts are due to different body activities (both voluntary and involuntary activities).
On the other hand, external artifacts arise from coupling due to unwanted external
interferences. In addition to that, artifacts in broad sense can be categorized into two
classes: ‘local’ and ‘global’. Local artifacts are confined in space, i.e. appear only in
a single recording channel while global artifacts are found across multiple channels
at the same temporal window. An example of global artifacts found in all recording
channels of two different neural recordings is shown in Fig. 2.8.
On the other hand, sometimes an artifact is found once in the whole recording
sequence (high entropy) while sometimes can have regular/periodic pattern due to any
periodic activities/motions of the subject. An example of such artifacts is shown in
Fig. 2.9. Table 2.2 summarizes the artifact classification from different perspectives.
Other noise sources are described below:
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 31

Fig. 2.8 Example of global artifacts from two different datasets of invasive neural signals

Fig. 2.9 An example of irregular artifact due to electrode pop (left) and periodic artifact due to
periodic motion activity of the subject (right) found in invasive neural recordings

Table 2.2 Summary of

Perspective Artifact category
artifact classification from
different perspectives Repeatability Irregular/No Periodic/Regular/Yes
Origin Internal External
Appearance Local Global
32 Md. K. Islam et al.

• White Noise: This noise results from thermal electronics noise mainly due to the
resistance and follows a flat frequency spectrum. No digital filtering can remove
the white noise completely as it has constant noise power over all frequency bands.
• Baseline Wandering: It usually results from respiration and has sub-Hz frequency
components.
• 1/ f Noise: This is a colored noise whose PSD follows reciprocal relation with
frequency and thus known as 1/ f α noise where α varies from 1 to 3.
• Power-line noises (50/60 Hz and its harmonics): Interference resulting from power
sources which usually have a very high peak at power line frequency (50/60 Hz)
and its harmonics. Proper grounding of electrodes is often required to minimize
the effect of such interference.
• Electrode Offset: Skin-electrode interface often is modeled as a DC voltage source
known as electrode offset.

2.2.2.1 Different Artifacts in Neural Signals (EEG, ECOG)

Physiological/Internal Artifacts

• Ocular Artifacts: The eyeball acts as an electrical dipole and therefore any move-
ment in eyeball generates large-amplitude artifacts in EEG recordings. Ocular
artifacts include eye blink, both horizontal and vertical eye movement, eye flatter,
eye movement during REM sleep, eye saccade, etc.
• Muscle Artifacts: One of the most prominent physiological artifacts comes from
muscle activity of the subject (EMG). Usually muscle artifacts are of high
frequency range (e.g. from 20 to 40 Hz) and are generated from activities like
chewing, swallowing, clenching, sniffing, talking, scalp contraction, eyebrows
raising, etc.
• Cardiac Artifacts: Cardiac artifacts are due to the electrical activities produced by
heart and are of two types: ECG and pulse artifacts. ECG artifacts are rhythmic
regular activities while the pulsation sometimes can cause slow waves which
might mimic the EEG activity.
• Respiration Artifacts: Respiration artifacts originate from the movement of an
electrode with inhalation or exhalation and can take the form of slow, rhythmic
EEG activity.
• Sweat Artifacts: Electrodermal or sweat artifacts originate from changes in elec-
trolyte concentration of electrode due to sweat secretion on the scalp and take the
shape of a long, slow baseline drift in the spectral band of 0.25–0.5 Hz [5].

Extra-Physiological/External Artifacts

• Motion Artifact: Movement of patient especially in an ambulatory EEG moni-

toring system [6–8], generates a lot of motion artifacts. This artifact often has
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 33

Table 2.3 Different types of artifacts and their sources found in EEG signals
Physiological or internal Extra-physiological or external
Ocular Cardiac Muscle Others Instrumental Interference Movement
Eye blink ECG Swallowing, Glossokinetic, Electrode Electrical, Tremor,
artifacts, artifacts; chewing, respiration, pop-up or sound, movements
eye pulse sniffing, skin, etc. displacement, optical, of head,
movement artifacts clenching, cable motion, magnetic, body, and
artifacts talking, No/poor etc. limbs
(both scalp grounding,
vertical contraction, etc.
and etc.
horizontal
EOG), eye
flatter, etc.

extremely high amplitude such that it can saturate the recordings. Head move-
ment, body movement, limbs movement, tremor, walking, running, browsing PC,
and many other movements in daily activities are responsible for this type of
artifact.
• Environmental Artifact
– Loose electrode: Loose contact between electrode and scalp leads to change
of impedance on the tissue-electrode interface and results in prolonged EEG
spike-like artifact.
– Electrode Pop and Movement: Another common source of artifact is due to
electrode pop which produces sudden change in impedance in the electrode-
tissue interface and results in high amplitude sharp waveform-shaped arti-
facts. Electrode movement occurs when it moves with respect to the scalp and
produces high-amplitude deflection in EEG generally in the low frequency
range of 1–10 Hz.
– EM Interferences: This type of artifacts is due to the interferences coming
from the surrounding electrical/electronic devices/machines that produce EM
waves. Also any sound or optical interference may also be picked up by the
EEG electrodes as artifacts. In addition, one of most common source of artifacts
in any biomedical signal acquisition is the 50/60 Hz main voltage and its
harmonics. A summary of different artifact types and their sources is provided
in Table 2.3.

2.2.2.2 Properties of Artifacts in Neural Recordings

Usually the artifacts have very large magnitude and/or sharper transitions/edges
compared to the biosignals of interest. The frequency range for artifacts may vary
from very low (e.g. motion artifact) to high frequency (e.g. artifacts due to residue
charge on electrodes) range suggesting artifact spectra overlap with biosignal of
interest.
34 Md. K. Islam et al.

Fig. 2.10 Illustration of origins of different artifact during invasive neural recordings

In order to characterize the spectrum statistics of artifacts as described in [9],

different artifact segments have been manually identified and then three artifact
templates (Type-1, 2, and 3) were extracted as shown in Fig. 2.11 (a, b, c). Later,
artifact spectrum is estimated using windowed Fourier Transform. The results are
summarized in Fig. 2.11 (d, e), where it is clearly seen that artifacts tend to display
varied spectrum shape and span over broad frequency band of 0 − 6 kHz. Type-1
and type-2 artifacts are dominating at low frequency range while type-3 has higher

Fig. 2.11 An example of varied artifact characteristics compared with neural signal of interest. a, b,
and c: Three different types of artifact templates: type-1, type-2, and type-3, collected from invasive
neural recordings. d, averaged PSD of type-1 and type-2 artifacts in comparison with LFP. e, type-3
artifact spectrum and neural spike (action potentials) spectrum. Spikes are randomly selected from
two different templates
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 35

frequency bandwidth. Figure 2.10 shows different artifacts’ origin during invasive
neural recordings.
To estimate power spectrum density of LFP only, data recorded from rat’s super-
ficial layer cortex have been analyzed and spectrum are averaged over 8-channels.
Figure 2.11d shows that PSD drops to the level of noise floor at frequency over
150 Hz.
To estimate neural spike spectrum, larger grouped spikes are smoothed and aver-
aged to extract different spike templates. As an example, only two spike templates
are used to estimate PSD as plotted in Fig. 2.11e, and found that PSD of spikes drops
to noise floor at frequency beyond 5 KHz.

2.3 Signal Analysis and Processing Techniques

for Handling Artifacts and Noises

Artifact detection and reduction/removal is one of the most faced challenges for EEG
and other bio-signal processing applications and is an open research problem. Most
of the biosignal recordings are prone to artifacts and interferences. The variety of
artifacts and their overlapping with signal of interest in both spectral and temporal
domain, even sometimes in spatial domain, makes it difficult for simple signal
preprocessing technique such as typical digital filtering or amplitude thresholding
to identify them from desired biosignals. Therefore the use of traditional filters
often results in poor performance both in terms of signal distortion and artifact
removal. Many attempts have been made to develop suitable methods for artifact
detection and removal with the help of recent advancement in signal processing
techniques/algorithms in the past decade and a half. However, there is no universal
complete solution yet and hence still an active area of research. After careful
reviewing almost all the major artifact detection and removal techniques found in
the literature, in this section we present a comparative analysis among these SPTs
considering their brief theoretical background, pros and cons based on their suit-
ability and performance and finally challenges in implementing them in different
biosignal applications.

2.3.1 Pre-processing

2.3.1.1 Epoch-by-Epoch Analysis/Segmentation

The recorded sequence of biosignals is often divided to small duration segments

known as epoch. The duration of epoch is such that it contains at least one cycle
of biological event of interest or such that the signal within the epoch duration
can be considered as stationary since most of the biosignals exhibit non-stationary
36 Md. K. Islam et al.

characteristics. The size of epoch also plays role in determining the computational
complexity of the signal processing algorithm in later stage which can be critical in
real-time application. The epoch duration is a trade-off between accuracy and real-
time computational ability of the SPTs. For EEG, the epoch duration is often decided
as one second since within this one second, it can be considered non-stationary
whereas EEG is typically a non-stationary signal.

2.3.1.2 Re-Referencing

For multi-channel EEG recordings, re-referencing is often used. Any potential

recorded at a particular electrode is with respect to a reference electrode (e.g. in EEG
recordings, the mastoid is often chosen as a reference electrode since it is closest to
the other electrodes as well as least chance to be influenced by neural potentials).
However, still the typical reference electrode may contain some neural information
as it also closer to the brain. Therefore, for high density EEG recordings, average
activities of all the electrodes may be chosen as reference. This re-referencing can be
done offline after the recordings are imported on a software toolbox such as EEGLAB
[10].

2.3.1.3 DC Offset Removal

Usually DC signal remains in the biosignal recordings due to electrode-skin interface

offset voltage which can be reduced by subtracting the average/mean value of the
biosignal from the biosignal itself. If X (t) is a raw biosignal recording and if the
mean of X (t) is µ, then the signal after removing DC offset would be X ′ (t) such that

X ′ (t) = X (t) − µ (2.2)

2.3.1.4 Digital Filtering

Digital filtering, which is nothing but a discrete-time LTI system, is a common part of
preprocessing the recorded biosignals to attenuate out-of-band noises and artifacts.
Both FIR and IIR filters have been found to be used in such preprocessing stage
depending on the application and given specification. The trade-off between FIR and
IIR filter is filter-order and stability, respectively. The transfer function of a digital
filter can be written as
M −l
l=0 bml z
H (z) = N (2.3)
1 + k=1 ak z −k
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 37

where H (z) is the z-transform of the impulse response of the LTI system, h(n) knows
as system/transfer function while ak and bm are the co-efficients of outputs, y(n − k)
and inputs, x(n − l) of a Discrete-time LTI system respectively. The difference equa-
tion based on which the present output, y(n) is related with the present input, x(n),
past inputs, x(n − l) and past outputs, y(n − k) is as follows:

y(n) = bm x(n − l) + ak y(n − k) (2.4)

Low-Pass Filtering

Depending on different biosignals, if desired bandwidth of the recorded signals is

known, then a typical FIR low-pass filter is used to cancel out high frequency out-of-
band noises and artifacts from the raw recordings. This can be done either in analog
domain or in digital domain.

High-Pass Filtering

Sometimes a very steep-slop (IIR or higher order if FIR) high-pass filter with cut-off
frequency as low as 0.05–0.1 Hz is used to cancel out electrode (DC) offset including
slow-wave artifacts (e.g. motion artifacts).

Notch Filtering

In most of the literatures, a 50 or 60-Hz 3rd or 4th order IIR notch filter is used to
remove the 50/60 Hz power line interference (PLI) and its harmonics. Since most of
the biosignals have maximum frequency up to 100 Hz (except EMG or invasive neural
signals), therefore, often it is good enough to remove the fundamental frequency of
the power line interference and removal of higher harmonics (2nd or 3rd harmonic)
are not required. However, the problem with notch filtering is that it not only removes
the PLI at the fundamental frequency but also removes signal component at that notch
frequency. In addition to that, the notch frequency has to be determined in advance
to design the notch filter, but in reality, there might be some fluctuations in the notch
frequency (e.g. it can be 49.8 or 50.1 Hz instead of exactly 50 Hz). In such cases,
the PLI doesn’t get removed properly and instead the desired signal component (e.g.
Gamma rhythm in EEG recordings) may be removed. In some literatures, researchers
proposed the use of adaptive filtering (given that the reference channel can record the
power line noise) to remove the PLI without attenuating the signal of interest in that
particular (i.e. 50 or 60 Hz) frequency. In [11], a regression approach was proposed
to predict the PLI and its harmonics through a mathematical model which doesn’t
require an extra reference channel and which also doesn’t bring any distortion to the
signal of interest. An example of application of notch filtering on ECG recordings is
illustrated in Fig. 2.12 to remove 50-Hz PLI and its harmonics.
38 Md. K. Islam et al.

Fig. 2.12 Effect of notch filtering on ECG signals (time domain on left and frequency domain in
right) to remove 50-Hz power supply noise and its harmonics

2.3.2 Artifact Avoidance

This is a preventive technique to avoid artifacts or minimize the effect of artifacts

by ensuring proper recording environment and protocol such as asking the subject
to have least amount of movements or eye-blinks, proper grounding of the recording
devices, using enough gel for better connectivity of the wet electrodes, etc. However,
artifact avoidance is not the ultimate way of getting rid of artifacts completely, in
some applications such as continuous ambulatory monitoring or BCI/HCI applica-
tions, subject’s movement is inevitable. Moreover, some of the internal or physio-
logical artifacts are involuntary and not possible to avoid, e.g. pulse and eye blink
artifacts found in EEG recordings. In addition, the subject may not limit the move-
ment for more than a specific period of time, especially if it is a child. Therefore,
these unavoidable artifacts should be removed in the later stage, i.e. digital signal
processing (DSP) domain.

2.3.3 Artifact Detection

Reliable identification of artifact contaminated segments of biosignal recording is

the most important step for handling artifacts. If any application requires to separate
or detect artifacts in real-time, therefore having prior knowledge of characteristics or
properties of either the artifact or the signal of interest is really necessary in order to
detect them faster. Artifact detection may refer to detection of a specific epoch or an
independent component (IC) as artifactual. The detection domain (time or frequency
or wavelet) is influenced by the type of artifacts and/or applications. The detection
method also varies depending on whether a reference artifact source is available or
not, whether the no. of channels is enough, whether artifact removal is required after
detection and so on.
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 39

2.3.3.1 Simple Amplitude Threshold

Often simple amplitude threshold based approach is taken to detect certain types
of artifact or artifactual epoch from artifact-free epoch in time domain. Usually the
signal amplitude higher than the set threshold is detected as artifact(s) and lower
than the threshold is assumed to be clean epoch. However, due to the non-stationary
nature of most biosignals as well as due to variety of artifact types, a pre-defined
threshold is not reliable to detect artifact. The following type of threshold value is
often seen to be used in the literature for a time series signal, x:

T hr = 3 ∗ r ms(x) (2.5)

2.3.3.2 Machine Learning

Recently machine learning based methods are being used (mostly supervised
learning) for artifact separation from useful biosignal of interest by extracting impor-
tant dominating features. Identified artifactual epochs are either marked as annotator
of artifacts for clinicians to make decision (e.g. epileptic onset detection) or rejected
before sending for examination to clinician or before sending to automated system
[12]. Machine learning techniques are mainly categorized as: supervised (labeled
training samples) and unsupervised learning (unlabeled samples). Artificial Neural
Network (ANN) [13–17] and Support Vector Machine (SVM) [12, 18–21] are two
widely used classifiers among supervised algorithms for separation between artifact
and brain signals. While popular unsupervised learning algorithms are k-means clus-
tering and outlier detection in artifact detection applications [12]. A typical process
of classifying artifacts from EEG using machine learning is illustrated in Fig. 2.13.

Deep Learning

Deep learning is an advanced version of artificial neural network with representation

learning which uses multiple deep layers of neurons to progressively extract higher
level features from the raw input and recently has been very popular for separation
or identification of artifactual epochs/artifacts from artifact-free epochs [22, 23].

2.3.4 Artifact Rejection

A simple approach to remove the influence of artifacts after detection is to

reject/cancel the artifact contaminated epoch or segment. This process not only
removes artifact but also removes signal of interest since both overlaps in temporal
domain which eventually results in the loss of critical information. This used to
40 Md. K. Islam et al.

Fig. 2.13 Machine learning approach for classifying artifactual epochs from clean epochs

be traditional way of handling artifacts, but recently with advancement of signal

processing techniques, the priority is towards removal or correction of artifactual
segments without distorting signal of interest instead of completely rejecting the
epoch. However, in some applications, this approach may still be useful, such as
offline analysis or training of machine learning classifiers. Artifact rejection can be
of following two types:
(a) Full sequence/channel rejection: Bad channels are completely rejected from
analysis if found too noisy or artifactual.
(b) Particular Epoch(s) rejection: Only the bad/noisy epochs are rejected from
further analysis instead of full channel rejection.

2.3.5 Artifact Removal

Artifact removal refers to cancelling or correcting the artifacts without distorting the
underlying biomedical signal of interest. This is mainly performed in two means:
(i) by filtering or regression and (ii) by separating or decomposing the biosignal
recording to other domains.

2.3.5.1 Regression

In this approach, a multi-modal linear model is assumed between observed artifact-

contaminated EEG channel and a reference channel containing artifact source. Then,
the samples that do not fit with the model are considered as outliers. Physiological
artifacts e.g. ocular and ECG artifacts may be removed using such technique from
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 41

EEG recordings. However, if no reference channel is available, then such regression

will be able to function. In addition, most biomedical signals originate from non-
linear and non-stationary process, which makes the linear regression method not
useful for artifact removal from such biosignals.

Single-Variate Autoregressive

In an autoregressive model the current sample of the signal is estimated from its
previous samples using a set of prediction coefficients optimally calculated using
Yule-Walker equations. Denoting the single biosignal as
p

x(t) = ai x(t − i) + e(t) (2.6)
i=1

where p denotes the prediction order, i.e., the number of previous samples used
in prediction of the current sample, and e(t) is the residual signal and needs to be
minimized and temporally during the prediction coefficient, ai , estimation. Smaller
values of p result in a smoother estimation of the signal by rejecting more redundancy
in the signal considering it as noise. On the other hand, larger p will include more
redundancy or noise within the estimated signal. Akaike Criterion [24] was one of
the first methods in estimating an acceptable value for p. This criterion however, was
improved by the approach proposed by Bengtsson [25] for a more accurate estimation
of p.

Multi-variate Autoregressive

This is an extension of single variate (univariate) autoregressive. assuming the main

multichannel EEGs are both time and space (channel) correlated, then prediction of
one sample from one channel from not only the same channel but samples of other
channels, which exploits the correlations between channels, can lead to rejecting
noise. This has application in brain connectivity, more robust spectrum estimation,
and most importantly denoising.

Variational Bayes

Variational Bayesian methods are some useful techniques to solve the problem of esti-
mating an original signal from degraded observations [26–29]. Unlike the Bayesian
methods (such as the Bayesian estimation methods based on Markov Chain Monte
Carlo (MCMC) algorithms, in VBA methods the intractable true posterior distribu-
tion is approximated using a tractable one from which the posterior mean can be
42 Md. K. Islam et al.

easily calculated. In addition, VBA methods usefully have a lower computational

complexity compared with the sampling-based methods.

2.3.5.2 Blind Source Separation/Subspace Signal Separation

Blind Source Separation (BSS) is known as useful technique for artifact detection
from biosignals. Here, the measured biosignals, X are considered as linear mixture
of the sources, S along with AWGN vector, N in multi-channel recordings

X = AS + N (2.7)

In BSS methods, the ultimate goal is to develop an iterative algorithm which

estimates the linear mixture matrix, A. Then, the estimated matrix (denoted by W)
is used to estimate the source signals, S’ by following formula:

S′ = W X (2.8)

In order to use BSS, it is required that the number of sources must be equal or less
than the number of observed channels. In addition, the sources should be maximally
uncorrelated (for CCA) or independent (for ICA) from each other. A basic BSS based
artifact removal technique is illustrated in Fig. 2.14.

Independent Component Analysis

In general, the Independent Component Analysis (ICA) based algorithms require

that the source signals are linearly independent and non-Gaussian distributed. In
order to apply these algorithms artifact detection and removal, it is also may require
manual intervention to reject independent source components (known as ICs) with
visually identified artifacts. It may be made automated by combining ICA with
another complementary method such as Empirical Mode Decomposition (EMD) or
Wavelet Transform (WT). It also may be used with machine learning classifiers such
as SVM or even with a help of a reference channel [30]. However, artifactual ICs
may also contain few residual neural or biosignals. Therefore, if the artifactual IC is

Fig. 2.14 Illustration of a basic blind source separation technique

2 Signal Artifacts and Techniques for Artifacts and Noise Removal 43

Fig. 2.15 A typical example of application of ICA to identify and remove artifacts (ocular and
muscle) that can be separated as an independent sources from multi-channel EEG recordings.
Adopted from [40, 41]

completely rejected then distortion to the background neural (or bio) signals occurs.
In addition, ICA requires multi-channel recordings to operate which suggests that
it cannot be applicable for single (or few channels) recordings. Another challenge
that restricts the suitability of ICA for artifact removal (especially in real-time appli-
cations) is its high computational complexity. This is because ICA based algorithm
usually requires multiple iterations to converge. Considering these factors, ICA may
be a suitable choice to remove global artifacts, i.e. ocular artifacts [15, 31–34] or
sometimes other physiological artifacts, but not external artifacts. The following
works used modified ICA [35] or constrained ICA [36–39] for making it automated
artifact detection and removal. An example of ICA based ocular and muscle artifact
detection and removal is illustrated in Fig. 2.15.

Canonical Correlation Analysis

A different BSS technique for separating mixed signals is Canonical Correlation

Analysis (CCA). In this method, second-order statistics (SOS) is used in order to
generate components based on their uncorrelated properties. Considering the uncor-
related components, CCA has weaker criteria than statistical independence used by
ICA. CCA addresses temporal correlation unlike ICA, therefore, CCA has maximum
temporal or spatial correlation within each component [42].

Morphological Component Analysis

The idea behind the Morphological Component Analysis (MCA) is to decompose the
recorded signals into components which have different morphological characteristics
and each component is sparsely represented in an over-complete dictionary [43]. It
is only applicable to certain artifact types whose waveform shape or morphological
characteristics are known in advance and stored in a database. MCA-based method’s
performance largely depends on the availability of the artifact-template database. An
example use of MCA is found in [44] for removing ocular artifacts and few EMG
44 Md. K. Islam et al.

artifacts originating from muscle activities during swallowing, jaw clenching, and
eye-brow raising.

Tensor Decomposition

It is an extension of Singular Value Decomposition (SVD) to multi-dimensional

space where in the case of multichannel medical signals can decompose the data into
its constituent components in a multi-dimensional (time, frequency, space- which
is the channel domain-, trial, subject, and even subject groups) or multi-way space.
This method separates the disjoint signals, including noise, and localizes the sources.
Tensor decomposition best exploits the diversity in the data and its variation in any
possible domain. Therefore, generally it is more effective than any other fusion or
decomposition technique such as PCA, ICA, or time-frequency method for detecting
and localizing events [45, 46].

Time-Frequency Representation

Time-frequency analysis is often performed for biomedical signals as suitable for

non-stationary time-series data such as EEG signals. Time and frequency domain
analysis are performed simultaneously since non-stationary biosignals have varied
statistical and spectral properties with time. Therefore, any change in the instan-
taneous frequency of each signal component [e.g. either artifact or seizure [47,
48] can be detected in a particular temporal window. An excellent example of the
use of such analysis is found in [49] where it is observed that frequencies up to
181 Hz can be present in a subject’s EOG signal for certain tasks after simultaneous
time-frequency representation was performed for the recorded ocular artifacts (OA)
including saccades and blinks. This result suggests that if EOG recording is used as a
reference channel for removing ocular artifact from EEG recordings, then sampling
rate of EOG recording must be at least 362 Hz (2 × 181) to avoid aliasing.
The short-time Fourier Transform (STFT) is a widely used time-frequency repre-
sentation. In this method a uniform time-frequency resolution is used for all signal
spectrum (frequency values). The spectrum of most biomedical signals is around
0.5–100 Hz and the spectrum of most artifacts appear in frequency region (< 10 Hz).
This means that having high frequency resolution in the lower frequency region is
required. But, clearly due to uniform frequency resolution, STFT cannot satisfy such
requirement. A nice solution of this issue is to use wavelet transform since it provides
a decent time-frequency resolution for most biosignals.

Short-Time Fourier Transform

Short-Time Fourier transform (STFT) is one of the common time-frequency repre-

sentation techniques which is obtained by segmenting the whole recording sequence
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 45

Fig. 2.16 Real neural signal contaminated with both type-1 and type-2 artifacts (bottom) and its
Spectrogram (top) shows relatively high frequency components at the temporal locations of these
artifacts

into many short-duration epochs by applying window function and then, its frequency
representation is calculated by FFT for each of this epochs:

+∞
Fx (t, f ; h) = ∫ x(u) ∗ h(u − t)e−i2πu f du (2.9)
−∞

Here, h(t) denotes the STFT sliding window. For a finite energy window it can
be represented as:

+∞ +∞
x(t) = E h ∫ ∫ Fx (u, f ; h)h(t − u)ei2πt f dud f (2.10)
−∞ −∞

+∞
where E h = ∫ |h(t)|2 dt. Consequently, STFT is used to determine the energy distri-
−∞
bution of any time-series signal (e.g. biosignals) in simultaneous time-frequency
domain. Figure 2.16 shows how STFT-based spectrogram can be useful in identi-
fying artifacts by plotting the biosignal energy in both temporal and spectral domain
simultaneously.

Wavelet Transform

Wavelets are localized in both temporal and spectral domains compared to the typical
Fourier transform which is localized in only frequency. Although STFT offers time-
frequency representation, but wavelets provide a better signal representation in terms
of higher frequency resolution in lower frequency region and thus more suitable for
46 Md. K. Islam et al.

biosignal time series analysis as most biosignals have dominant frequency compo-
nents in the low frequency region. The wavelet transform decomposes original signal
f(t) into dilated and translated versions of a basis function ψ(t) known as mother
wavelet [50]. Any Wavelet is generated from a mother wavelet via:

ψ j,k (t) = 2 j/2 ψ 2 j/2 t − k (2.11)

 

where k is the translation in time with scaling factor of 2 and j indicates the resolution
level. Wavelet decomposition follows linear expansion expressed as follows:

+∞
 +∞ 
 ∞
[ck ϕ(t − k)] + d j,k ψ 2 j t − k (2.12)
 
f (t) =
k=−∞ k=−∞ j=0

where ϕ(t) is known as the scaling function or father wavelet and ck and d j , k are
the coarse and detail level expansion coefficients, respectively. Theoretically, the
expansion coefficients ck and d j,k are calculated from the inner product of f (t) with
ϕ(t) and ψ(t), respectively. A function may serve as mother wavelet by satisfying
the following condition:

+∞
∫ ψ(t)dt = 0 (2.13)
−∞

There are various types of wavelet methods such as wavelet approximation

and decomposition, wavelet packet decomposition (WPD), discrete and continuous
wavelet transform (CWT), stationary wavelet transform (SWT), and so forth. Among
them, Discrete Wavelet Transform (DWT) is the most commonly used technique. The
relation between input and output of DWT can be expressed as:

N

xa,L [n] = xa−1,L [2n − k]g[k] (2.14)
k=1

N

xa,H [n] = xa−1,L [2n − k]h[k] (2.15)
k=1

where g[n] is a low-pass filter mimicking scaling function and h[n] is a high-pass
filter similar to mother wavelet. Briefly, discrete wavelet transform decomposing a
signal into its low frequency component and high frequency components through
these two filters known as approximate and detailed coefficients. The wavelet filter
decomposition structure is shown in Fig. 2.17.
Once the signal is decomposed, thresholding is applied to denoise the signal from
artifacts. At that point the new sets of detailed and approximate coefficients are
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 47

Fig. 2.17 The decomposition (analysis) and reconstruction (synthesis) structures of wavelet filters

Fig. 2.18 Typical process flow of wavelet-based denoising technique

added up to reconstruct back the artifact-free signal. A typical wavelet denoising

based artifact removal from EEG signal as an example is illustrated in Fig. 2.18.
In recent years, wavelet transform based denoising (e.g. DWT) have gained special
attention in EEG signal processing due to their suitability in non-linear and non-
stationary signal processing [51]. Wavelet transform can be used to separate the
energy of the EEG recording into different frequency bands. When it is applied
to artifactual EEG signal, it yields the wavelet coefficients representing correlation
between the noisy EEG and the wavelet function. Based on the selection of mother
wavelet, larger coefficients correspond to the artifactual segment, while smaller coef-
ficients correspond to the actual EEG. It is important to note that suitable mother
wavelet and thresholds are required to ensure the effective separation of the artifact
coefficients and the EEG signal coefficients. The problem remains how to select
the proper mother wavelet and how to choose the parameters for achieving best
performance in an automatic fashion.

Empirical Mode Decomposition

Empirical mode decomposition (EMD) is an empirical and data-driven technique

applicable for non-stationary, non-linear, stochastic processes, such as EEG signals.
However, the computational burden of EMD is higher doubting its ability to perform
in online applications. EMD algorithm decomposes the original signal, s(n) into a
sum of the band limited functions, dm (n) known as intrinsic mode functions (IMF)
48 Md. K. Islam et al.

with well-defined instantaneous frequencies [52–54]. An IMF should satisfy the

following criteria: (i) at any point, the mean value of the two envelopes defined by
local maxima and local minima is zero [54], (ii) the number of extrema has to be
equal (or at the most differ by one) to that of the number of zero crossings. The
flowchart of EMD algorithm to calculate IMF is shown in Fig. 2.19.
Enhanced Empirical Mode Decomposition (EEM is modified from EMD to make
it robust to noise which will avoid the mode mixing complication arises in EMD. To
achieve this EEMD uses mean value of the number of ensembles (IMFs) as optimal
IMFs allowing it to offer a noise-assisted data analysis technique [42]. An example
of simple EMD-based artifact removal process is shown in Fig. 2.20.

Wiener Filtering

Unlike adaptive filters, in Wiener filter a reference signal is not required. However, it
assumes that both the signal and artifact are stationary linear random processes, where
their spectral characteristics are known and also the signal and artifact are uncorre-
lated. But in reality, most of the biomedical signals exhibit non-stationary charac-
teristics and are believed to be originated from a complicated non-linear stochastic
process. Again, although the spectral characteristics of most biomedical signals are
known, due to the uncertainty of different types of artifact sources, the spectral char-
acteristics of artifacts cannot be determined accurately. In addition, the wiener filter
is unable to be implemented in real-time, thus may not be suitable for applications
where real-time processing is requires such as closed-loop Human-machine inter-
facing (HMI) through which external devices/machines (e.g. wheelchair, computer,
prosthetic limbs) are controlled by biomedical signals (e.g. EEG, EMG, EOG, etc.).

Adaptive Filtering

Adaptive filters have been manifested to prove useful in great deal of biomedical
applications. For example, most biosignals, such as ECG or EEG signal acquisition,
the information-bearing signals may be contaminated by noise and disturbances
caused by the 50/60 Hz power-line, high frequency interference and random body
voltages. In such problems, both required signal and noise occur in an identical
frequency band and so the noise cannot be discriminately filtered out by removing
any specific frequency band. In such cases, filters that can adjust to the changing
noise are required. Adaptive filters, systems with variable as an alternative to fixed
filter coefficients, can overcome these difficulties. This is achieved by employing
adaptive filters such as least-mean-square (LMS) algorithm, recursive-least-square
(RLS) algorithm, and Kalman filter-type algorithms, as the analytical implementation
of Bayesian filtering recursions for linear Normal state-space models. Adaptive filters
are preferably designed as FIR filters, as shown in Fig. 2.21, known for their good
stability properties and ease of implementation.
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 49

Fig. 2.19 The general process flow of EMD algorithm to generate IMFs
50 Md. K. Islam et al.

Fig. 2.20 Typical process flow of EMD-based denoising technique

Fig. 2.21 General adaptive filter structure, with concept of interference cancellation

As mentioned earlier, adaptive filters have also been used for artifact removal
from biosignals. In [55], a hybrid nonlinear adaptive filtering has been reported for
removing motion artifacts from wearable photoplethysmography. In [56] an adaptive
filtering algorithm has been developed for motion artifact removal from capacitive
ECG signals. In this algorithm, the power-line interference (PLI) has been used to
extract the required reference signal. Another adaptive filtering based algorithm for
motion artifact removal from the ECG recordings has been reported in [57]. In this
algorithm, the spectral energy variation during the input process of motion artifacts
is used to develop a cosine transform LMS adaptive cancellation algorithm. In [58],
a motion artifact removal algorithm has been proposed which uses a cascade of LMS
adaptive filters, in conjunction with a reference noise estimation method. Kim et al.
[59] developed a method using ICA and adaptive filtering for MI (motor imagery)-
BCI applications. They showed that this method can remove Ocular Artifacts from the
EEG signals without measuring Electrooculogram (EOG). In [60], a neural network-
enhanced adaptive filtering algorithm has been reported for EEG artifact removal.
In order to remove EOG artifacts from EEG recording, it is usually required to
have multi-channel EEG recording or an additional EOG recording in real-time.
In [61] a new method has been developed which uses a cascade of RLS adaptive
filters and sparse autoencoder (SAE) to remove EOG artifacts from EEG recordings.
A novel time-domain linear filtering algorithm to remove ocular artifacts from EEG
signals has been introduced in [62], where the eye-blink signal is obtained by a small
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 51

Fig. 2.22 A schematic of the Kalman filter structure

number of frontal electrodes (instead of directly estimating the artifact-free signal)

and applying a multichannel Wiener filter (MWF).

Kalman Filtering

Kalman Filter is an estimator based on Minimum Mean Square Error (MMSE) which
is often used to extract or smooth physiological signals. Moreover, Kalman filtering
may be used to denoise, separate signals or fuse sensor data, all three in a single
architecture. The main advantage of Kalman filter, compared to other filtering or
signal separation techniques, is its lower systemic delays in real-time computations.
In order to apply the Kalman Filter, the observations should follow a state-space
model. In this model an equation shows the evolution of the state model, and the
other equation describes the relation of the parameters with the observations:

x i = Ax i−1 + wi (2.16)

yi = B z i + v i (2.17)

In the above model x i are represents the state of model at time i, yi are the noisy
measurements (observations), wi is the state noise, v i is the observation, B is the
observation matrix and A is the state transition matrix. A schematic diagram of the
Kalman Filter is demonstrated in Fig. 2.22.
The Kalman filter consists of two prediction and update steps, which are sequen-
tially executed through time. In the prediction step, the available the data up to time
i +1 is used to estimate x i . The forecast is denoted as x̂ i|i−1 and subject to uncertainty
52 Md. K. Islam et al.

quantified by the prediction error covariance P i|i−1 . When the new measurement yi
is available, the update step is performed. In this step, yi is leveraged to correct x̂ i|i−1
and generate an updated estimate x̂ i|i . At this step P i|i−1 is also updated to obtain
P i|i to quantify the uncertainty imposed on x̂ i|i .

Particle Filtering

Particle filter is a kind of filter based on Bayesian approach which overcomes the
limitation of Kalman filter as it does not require the data follow a linear model or
the distribution to be unimodal. But it still needs a priori user input which may not
be available always in EEG-based applications. And there is very little work has
been done by far to use particle filter to remove artifacts in EEG signals. Hence it
is not guaranteed to be a successful choice, but one can definitely try to observe the
outcome of such filter implementation in removing artifacts.

Spatial Filtering

Principal Component Analysis (PCA) is one kind of spatial filtering that typically
transforms the original dataset from temporal domain to a new domain by rotating
the axes in an N -dimensional space (N is the no. recording channels) where each
dimension in the resultant space having minimum variance as well as axes are being
orthogonal to each other [63]. It reduces dimensionality of the dataset and highlights
dominant features of data that are usually hard to detect in the original domain. In
[64] a robust PCA is used after wavelet-based denoising is done as preprocessing
while in [65], a comparison is made between PCA and ICA for artifact removal
and it is found that ICA outperforms PCA. Since both these articles evaluated the
performance qualitatively; therefore, it is difficult to assess the efficacy of the use of
PCA in artifact removal. One significant drawback of PCA is that it cannot be used
to identify artifacts when signal and artifact amplitudes are comparable to each other
(e.g. ocular artifacts from EEG) as it relies on the higher order statistics of the data
[36].

Hybrid Methods

In recent years, many works are reported that have utilized the advantages of different
SPTs by combining two or more techniques in multiple stages into a single technique
for detecting and removing artifacts from biosignals. Some of such hybrid methods
are described as follows:
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 53

Fig. 2.23 General process flow of wavelet-BSS and EMD-BSS methods

Wavelet-BSS

Wavelet-BSS is formed by combining two commonly used methods: wavelet trans-

form in the first stage followed by blind source separation in the later stage. It is
mainly inspired since BSS based separation of artifactual independent components
(i.e. ICs) is often flawed as the separated IC may also contain residual biosignals
which eventually results in significant distortion in reconstructed signals. To over-
come this issue, multi-channel biosignal recordings are converted to ICs or CCs
by applying BSS and then potential artifactual component is further decomposed by
wavelet decomposition into detail and approximate coefficients of different frequency
bands. After that, wavelet denoising is applied which eventually preserves the residual
bio signals of low amplitude after thresholding is used to remove the higher ampli-
tude artifactual segments. The reference articles are [66–68] for wavelet-ICA, [69,
70] for wavelet-CCA. On the other hand, for single-channel recording, reversing
the order of wavelet transform and BSS i.e. BSS-Wavelet may be used. E.g. [71,
72] reported artifact removal by first decomposing signal into wavelet coefficients;
after that BSS is applied on the artifactual coefficients to separate artifacts from
background neural/bio signal. However, the former method is more popular to the
research community which is known as wavelet enhanced ICA (wICA) or wavelet
enhanced CCA (wCCA).

EMD-BSS

It includes BSS with EMD technique. The initial stage of EMD-BSS is to decompose
the signal into IMFs by applying EMD or EEMD and then BSS (either ICA or
CCA) is applied on the IMFs to identify artifactual components and finally rejecting
the artifactual ICs or CCs. (E)EMD-BSS are reported in [42, 73, 74]. Figure 2.23
illustrates typical steps involved in both wavelet-BSS and EMD-BSS based methods.

BSS-SVM

A hybrid BSS-SVM was reported in [20] for removing eye blink and ECG artifacts
from EEG recordings where features extraction is performed from separated source
components (after BSS is applied) to feed the SVM classifier to separate artifact
components followed by removal of the artifact components. Finally in order to
54 Md. K. Islam et al.

Fig. 2.24 Typical process flow of BSS-SVM method for artifact removal

Fig. 2.25 Process flow of REG-BSS hybrid method for artifact removal

reconstruct artifact-free EEG, the remaining source components are re-projected.

The whole process is illustrated in Fig. 2.24.

REG-BSS

In [31] a hybrid methodology was reported by combining BSS and regression for
removal of ocular artifacts where both vertical EOG and horizontal EOG were used
as reference channels as shown in Fig. 2.25. Similar approach has been proposed by
Guerrero-Mosquera [26] to remove ocular artifacts by involving ICA and adaptive
filtering. Another work [75] proposed to combine ICA and Auto-Regressive eXoge-
nous (ARX) to implement a robust ocular artifact removal where ARX reduced the
negative effect induced due to ICA.

Other Approaches

Nguyen et al. [76] reports removing EOG artifacts from EEG using the combination
of Wavelet decomposition and Artificial Neural Network, i.e. WNN where EOG
reference channel is only required during training of ANN classifier. Another work
[77] proposed a hybrid method combining DWT and ANC (Adaptive noise canceller)
to remove EOG artifacts (the reference signal is estimated from DWT coefficients
required for the adaptive filter). On the other hand, [78] proposed the use of both
EMD and adaptive filter (using RLS algorithm) to remove cardiac artifacts from EEG
recordings. The authors in [16] reported a new hybrid method to remove EOG and
EMG artifacts from EEG recordings by combining functional link neural network
(FLNN) and adaptive neural fuzzy inference system (ANFIS).
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 55

2.3.5.3 Statistical Features

Statistical features are also used in machine learning in feature extraction stage or
during calculation of threshold value used in different SPTs (wavelet, EMD, ICA,
etc.) for identifying artifacts from biosignal of interest. Some of the commonly used
features are discussed below:

Time Domain Features

Among time domain features, most commonly used features are: Entropy, Kurtosis,
Line Length, NEO, Maximum, Minimum, Variance, Mean, etc. [51].

Frequency Domain Features

Among spectral domain features, most commonly used features are mean, maximum,
minimum and variance of the absolute value of FFT or PSD. E.g. EEG rhythms
are non-overlapping frequency bands and therefore, spectral features may be useful
to separate artifacts from a targeted EEG rhythm in consideration for a specific
application.

Spatial Features

Spatial distribution of the recordings can be known from spatial features which
allows to identify the origin of brain signals as well as some artifact types (e.g.
ocular artifacts are mostly found in frontal electrodes as closest to the origin of that
artifact). In addition, global artifacts (e.g. eye blinks) can be differentiated from
local artifacts based on spatial mapping. Therefore spatial features of data with both
temporal and spectral contents can be useful to separate artifacts from EEG signals
[14, 79].

Auto-Regressive Features

Refer to Sect. 3.5.1.1, the AR model parameters, αi can be considered as features

for classification.

2.3.6 Summary and Comparison Between SPTs

See Table 2.4.

Table 2.4 Comparison of different SPTs used in artifact removal from different biosignals

56
Main methods used References Type of Type of artifacts/noise Applications Remarks
bio
signal
Adaptive filtering/ANC [55–61] PPG, Motion/Ocular/EMG + ECG + Cardiac diagnosis Extra channel
ECG, EOG/ (EOG, ECG,
EEG Accelerometer,
Gyroscope, Eye
tracker, motion
captured camera,
contact impedance
measurement, etc.
required to record
reference signal
Kalman Filtering [80–82] ECG Gaussian/Non-Gaussian/muscle Cardiac diagnosis Automated and
artifact/PLI real-time
processing, no
a-priori info
required, not
suitable for
non-linear and
non-stationary
process like EEG
(continued)

Md. K. Islam et al.

Table 2.4 (continued)

2 Signal Artifacts and Techniques for Artifacts and Noise Removal

Main methods used References Type of Type of artifacts/noise Applications Remarks
bio
signal
Wiener Filtering [62] EEG Eye Blink General Proposed for
multi-channel EEG
recording. Also can
work on
single-channel
data. A-priori user
input required and
cannot operate
online in general
(c)ICA [31–35, 37, 39, 83 ] EEG Mostly ocular, few motion and Epilepsy Most of them are
other Physiological artifacts diagnosis, BCI, semi-automated,
general purpose few require
additional sensor(s)
while
CCA [84, 85] EEG Muscle/ocular Ictal EEG source Not
imaging, Epilepsy online/real-time
monitoring processing,
requires
multi-channel
recordings
PCA [35, 57, 64] EEG General SSVEP for Robust PCA
emotion detection
(continued)

57
Table 2.4 (continued)

58
Main methods used References Type of Type of artifacts/noise Applications Remarks
bio
signal
Wavelet transform [51] EEG All types: ocular, muscle Epileptic seizure Can work for both
detection single and
multi-channel
recordings, no extra
reference channel
required
Tensor decomposition [46, 86] EEG Ocular Epileptic Seizure Tensor
detection and decomposition may
localization, involve wavelet
diagnosis in transform, ICA,
Preterm babies etc.
and many other
EEG related
applications
(E)EMD [42, 54] EEG EOG General Extra reference of
Fractional
Gaussian
Noise required and
can work for single
channel recordings
Hybrid methods
(continued)

Md. K. Islam et al.

Table 2.4 (continued)

2 Signal Artifacts and Techniques for Artifacts and Noise Removal

Main methods used References Type of Type of artifacts/noise Applications Remarks
bio
signal
Wavelet-BSS (W-ICA/W-CCA) [66–69, 71] EEG Ocular + muscle + electrical General Mostly applicable
shift for multi-channel
recordings
(E)EMD-BSS (ICA/CCA) [42, 73, 74] EEG Muscle, motion, EOG Ambulatory Can work for single
single-channel channel and mostly
applications, ictal automated
EEG, diagnosis
BSS-SVM [20] EEG Eye blink General; e.g. ERP Applicable only for
analysis multi-channel
recordings
REG-BSS [31, 87, 88] EEG Ocular General vEOG and hEOG
reference channel
required
Others [16, 76–78] EEG EOG, ECG, EMG General, Reference channels
(WNN/DWT-ANC/EMD-ANC/ICA-ANC/FLNN-ANFIS) diagnosis, are required for
portable ANC related
applications methods

59
60 Md. K. Islam et al.

2.3.7 Currently Available Software Plug-Ins

Different biosignal processing and analysis toolboxes are listed in the following
table where it is shown what are the SPTs used for these toolboxes along with type
of biosignal and artifact types handled.

2.4 Discussion

2.4.1 Limitations and Suitability of the SPTs

2.4.1.1 Real-Time/Online Implementation

There are some applications that require real-time or online processing of biosig-
nals such as HMI or real-time seizure detection (i.e. BCI, neural prosthetics, EOG-
controlled wheelchair, etc.). Therefore, choice of artifact removal technique for such
application would be such that it has the required low computational complexity
to be compatible with real-time/online processing. In that case, trade-off between
computational complexity and performance of the artifact removal algorithm is to
be made. On the other hand, there are applications such as diagnosis of neurological
disorders/diseases that may allow offline processing. In such case, one can only focus
on achieving highest performance without much considering about computational
time.

2.4.1.2 Single or Multi-channel

As seen from both Tables 2.4 and 2.5 that some SPTs are suitable for both single and
multi-channel biosignal recordings while few of them (such as BSS based techniques)
require multi-channel recordings to be applied. On the other hand, wavelet or EMD
based techniques can be applied for a single-channel recording. So it is critical to
select SPT for appropriate application considering the no. of channels in mind.

2.4.1.3 Reference Channel

Refer to Table 2.5, some of the available SPTs require an additional dedicated refer-
ence channel to record artifact source for functional. Example of such reference
channels: EOG, ECG, Motion Sensors, and Contact Impedance Measurement chan-
nels to remove ocular artifact, cardiac/pulse artifact, motion artifacts, artifacts due
to electrode popup or movement, respectively. However, having an extra reference
channel is not always feasible in some applications such as portable and contin-
uous monitoring if the reference sensor is not integrated with the original biosignal
Table 2.5 Existing software toolboxes for handling different types of artifacts from biosignals

2 Signal Artifacts and Techniques for Artifacts and Noise Removal

Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
EEGLAB [10] ICA, EEG, MEG, Almost all types of Supports importing MATLAB Y
Time-frequency ECoG Artifacts except of multi-format
Analysis, standard motion artifacts database,
averaging high-density
recordings
scrolling, artifact
removal
(semi-automated),
EVENT and
channel location
handling, forward
and inverse head or
source modeling,
over 100 advanced
plug-ins/extensions
BCILAB [89] Similar to EEGLAB EEG Almost all types of BCI Applications EEGLAB/MATLAB Y
artifacts except and classification
motion artifacts
Bio-SP [90] Signal-specific Peripheral General types of Feature extraction MATLAB (Any Y
algorithms Physiological artifacts and noises and critical events version)
Mainly IIR digital Signals including PLI or parameters Windows, macOS,
filtering (Elliptic, (ECG, EDA, detection Linux
Gaussian, or EMG, BP, or
Butterworth) ICG)
(continued)

61
Table 2.5 (continued)

62
Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
BioSig [91] Linear regression EEG, ECoG, Removal EOG Real-time BCI MATLAB or octave Y
for EOG artifacts ECG, EOG, artifacts and only system and python
removal from EEG, EMG, and detection of EMG implemented in
and time/frequency Respiration artifacts from EEG MATLAB and
domain features signals Simulink
such as slope,
amplitude or
spectral edge
frequency for EMG
artifacts detection
from EEG
ANSLAB [92] Threshold-based, ECG, EDA, General types Analysis of MATLAB Open ANSLAB—Y
digital filtering EMG, eye long-term ANSLAB
blinks, recordings from Professional—Partly
respiration ambulatory settings
signals
Clean [93] ASR and digital EEG Low-frequency default EEGLAB EEGLAB/MATLAB Y
rawdata filtering drifts, flat line, eye method for
blinks, muscle removing artifacts
activity, sensor from EEG
motion
BioSigKit [94] RLS adaptive Filter, ECG, EMG Motion artifacts Fetal-ECG MATLAB Y
variable leaky ALE and EEG extraction from

Md. K. Islam et al.

filter both single and
multi-channel
maternal ECG
(continued)
Table 2.5 (continued)

2 Signal Artifacts and Techniques for Artifacts and Noise Removal

Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
biosppy [95] Digital filtering BVP, ECG, Out of band noise Support for various Python Y
EDA, EEG, and DC offset biosignal
EMG, processing,
Respiration detection of events
and clustering
WFDB [96] Not found ECG, EEG, Not mentioned Allows reading, MATLAB/Octave Y
EMG, fetal writing, and
ECG, PLETH processing of
(PPG), ABP, physiological
respiration, signals and time
etc. series signals from
PhysioBank
databases
BSP [97] Non-linear filtering ECG, ABP, Movement artifact Nonlinear filtering, MATLAB Y
(Rank Filters, ICP, and and quantization detection
Median Filters) SpO2 noise algorithms, spectral
analysis,
multi-signal
analysis,
non-stationary
signal visualization.
EMGLAB [98] Digital filtering EMG Low frequency Decomposing EMG MATLAB Y
artifacts, DC offset, into MUAP trains
out of band noises and averaging
MUAP waveforms
(continued)

63
Table 2.5 (continued)

64
Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
SASICA [99] Reject/select EEG Muscle artifacts Semi-automated EEGLAB/MATLAB Y
independent guided selection of
components based artifactual ICs
various
features/properties
FASTER [100] ICA EEG EOG, EMG Fully automated EEGLAB/MATLAB Y
artifacts, linear and unsupervised;
trend and white supports
noise high-density EEG
recordings
OSET [101] Semi BSS and Any Cardiac and EOG Biological signal MATLAB and C ++ Y
digital filtering Bioelectric artifacts generation,
signals modeling,
processing, and
filtering
AAR [102] Regression/adaptive EEG Ocular and Automated EEGLAB/MATLAB Y
algorithms (LMS, muscular artifact algorithm
RLS), BSS and (ICA-based)
Spatial Filters +
ICA
ADJUST [14] ICA EEG Ocular (blinks, eye Automated; no EEGLAB/MATLAB Y
movements) MATLAB toolbox

Md. K. Islam et al.

is needed to run
FOOBI [103] Underdetermined EEG Ocular Automated, no – –
blind source extra channel
separation required
(continued)
Table 2.5 (continued)

2 Signal Artifacts and Techniques for Artifacts and Noise Removal

Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
LAMIC [104] Temporal extension EEG Ocular, EMG and Automatic – –
of ICA (TDSEP) ECG clustering
followed by algorithm
clustering based on
AMI
FORCe [83] Combination of WT, EEG Eye blink, EOG Fully online and MATLAB Wavelet Y
ICA and and EMG artifacts automated artifact Toolbox/EEGLAB
thresholding Removal for BCI
applications
PRANA [105] Digital filtering + EMG, ECG, Not mentioned GUI for different MATLAB Y
BSS EEG, EOG, (mainly ocular Biosignals
EDA, pulse artifacts) processing and time
oxymetry, series analysis
respiration,
blood
pressure,
temperature,
sound, light
intensity
activity, etc.
PFEIFER [106] Digital filtering ECG Signal noise, drift, Computationally MATLAB Y
etc. efficient GUI
(continued)

65
Table 2.5 (continued)

66
Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
Brainstorm [107] Signal Space MEG, EEG, Eye blinks, ECG Supports wide MATLAB (executable, Y
Projections, ICA, fNIRS, verities of data file OS platform
Digital filtering ECoG, formats and brain independent, doesn’t
invasive recording require to own a
neural modalities with License)
recordings automated
detection of
artifacts, 2D/3D
visualization,
time-series
analysis, head and
source modeling,
functional
connectivity,
decoding and
pattern analysis
FieldTrip [49] Digital filtering and MEG, EEG, EOG, ECG, EMG Both manual and MATLAB Y
ICA iEEG and artifacts and line automatic artifact
NIRS noise rejection. Allows
time-frequency
analysis,
dipoles-based
source

Md. K. Islam et al.

reconstruction, and
non-parametric
statistical testing
(continued)
Table 2.5 (continued)

2 Signal Artifacts and Techniques for Artifacts and Noise Removal

Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
REST [108] Online RLS EEG General noise Real-time source MATLAB/EEGLAB Y
Whitening, Online localization
ICA
BERGEN [109] Digital IIR filters simultaneous Gradient artifacts Head movements EEGLAB/MATLAB Y
(band-pass) and EEG-fMRI from fMRI are taken into
Moving average account for better
correction based on removal of gradient
artifact templates artifacts
CIAC [110] ICA EEG Electrical artifacts Allows to identify EEGLAB/MATLAB Y
(semi-automated) originating from and cluster (Statistics toolbox)
cochlear implants artifactual ICs
across subjects
CORRMAP [111] Template based EEG Ocular and Allows to identify EEGLAB/MATLAB Y
identification heartbeat artifacts and cluster (Statistics toolbox)
(Semi-automatic) artifactual ICs
and clustering of across subjects
independent
components
FMRIB [75] FASTR: FMRI EEG-fMRI fMRI-environment GUI software EEGLAB/MATLAB Y
artifact slice artifacts: e.g.
template removal gradient artifacts,
BCG artifacts and
scanner-related
artifacts from EEG
(continued)

67
Table 2.5 (continued)

68
Name of References Main methods used Type of bio- Type of Extra features or Platform/compatibility Free/open source?
toolbox signal artifacts/noises specs
MARA [112, 113] Linear classifier to EEG Ocular artifacts, Not limited to a EEGLAB/MATLAB Y
classify independent muscular artifacts specific artifact (toolboxes required:
components as and loose type; uses Statistics,
‘accept’ or ‘reject’ electrodes supervised machine Optimization and
learning algorithm Signal Processing)
REGICA [31] Regression EEG EOG artifacts Semi-automated EEGLAB/MATLAB Y
performed on ICA
components
CleanLine [114] ICA + Regression EEG 50/60 Hz line noise Both manual and EEGLAB/MATLAB Y
technique + harmonics automatic artifact
(frequency-domain) rejection, supports
computationally
efficient GUI

Md. K. Islam et al.

2 Signal Artifacts and Techniques for Artifacts and Noise Removal 69

recorder. In addition, these reference channels must synchronize (e.g. in terms of

sampling rate or dynamic range), with relevant biosignal recordings to be able to
apply regression or adaptive filtering technique.

2.4.1.4 Robustness

Robustness is another important issue to decide on the selection of any artifact

removal technique since diverse types of artifacts contaminate and/or affect different
biosignals differently for different recording protocols and for different environ-
ments. In order to evaluate a particular SPT on its ability to detect and remove arti-
facts from a specific biosignal, it is very critical to prove its robustness under different
experimental setups (or different applications or environments) and different set of
subjects.

2.4.2 Future Challenges and Opportunities

With the advancement of technology (e.g. improved sensors, flexible and wear-
able electronics, analog front-ends, wireless data transfer, cloud computing, AI and
machine learning), biomedical signals are going to be recorded and processed real-
time in an ambulatory settings where the subjects can move freely and able to perform
daily activities. The purpose of such continuous ambulatory recording is not only to
monitor patient’s condition or to diagnose diseases, but also to predict future health
condition and utilize continuous biomedical signals for preventive healthcare. In
addition to that, such recording will be used in the area of games and sports and
evaluation of one’s overall fitness. This potential future scope will come up with new
challenges such as handling extreme motion artifacts due to subject’s movements
in daily activities. Another challenge will be the processing of huge amount of data
samples for such continuous recording which will require advanced SPTs to be able
to operate in real-time. In addition to that, transferring the samples wirelessly and
eventually to store this big data for future reference is another potential challenge
to address. In such case, before transferring the samples wirelessly, instead of trans-
ferring raw data, extracted and selected features along with compression may be
important which again requires new improved SPTs to be proposed and tested.

2.4.3 Recommendations

2.4.3.1 Application Specific Models

The choice of SPT should be application specific. Often it is not required to remove
each and every artifact type for a particular application, instead the preference should
70 Md. K. Islam et al.

be given only on those artifacts that affect the application most. For example, in Motor
Imagery based BCI, the frequency bandwidth of EEG signal of interest is 8 − 32 Hz
(i.e. Alpha and Beta rhythms) which means artifacts present in that frequency range
must be handled carefully and removed without distorting targeted signal of interest.
Artifacts due to muscle activities (EMG) lie in that frequency range while EOG/ECG
or motion artifacts belong to less than 8 Hz frequency bandwidth. This implies that
one should not bother about removing EOG/ECG/motion artifacts that much for MI
based BCI applications, rather concentrate on only removing EMG artifacts as it
affect both the Alpha and Beta rhythms most. Therefore, to choose the right SPT for
removing artifacts, one should consider the particular application and required spec-
ifications to be met given a certain computational resources and recording protocol
available. Only those artifacts should be removed which affect the later stage deci-
sion making. If an extra reference channel is available to record artifact source (e.g.
EOG, ECG, motion sensors, etc.), then SPTs such as regression or adaptive filtering
technique may be applied. In case of ambulatory and continuous monitoring appli-
cations, no. of channels is lower and reference channel is not available, then it is
recommended to use computationally efficient methods that are capable of func-
tioning without reference channel and with single or few channels. Examples of
such SPTs are wavelet based methods. In some applications, if prior knowledge
about artifacts is available and some training samples are available, and finally if
the it only requires to identify artifacts without removing them, at that time machine
learning may be a realistic choice. If the biosignals have high-density channels, then
PCA can be applied to reduce the dimensionality of the data before applying SPTs for
artifact removal (e.g. BSS involved techniques). If the application does not require
real-time computation (i.e. involves offline analysis), then computationally expen-
sive techniques with high performance such as cICA or EEMD may be applicable.
An example of different application-specific models is as follows:
• Biosignal Specific: EEG, iEEG/ECoG, ECG, EMG, etc.
• Diagnostic/Clinical Application Specific:
• Artifact Specific: Ocular, Muscle, or Motion artifacts.

2.4.3.2 Standard Performance Evaluation

Lack of standard performance evaluation metrics/criteria for the SPTs is a big

concern. Most works proposing SPTs for artifact removal found in the literature
mentioned some qualitative plots in either time or frequency domain to visually
assess the performance (e.g. assessment by clinical experts). (K. T. Sweeney et al.
[52] proposed a recording protocol for correct assessment and comparison between
different SPTs for physiological signals which may be suitable for some applica-
tions that allow intervention to the recording protocol and an extra artifact reference
channel is available. However, applications such as portable EEG recordings for
ambulatory monitoring may not be compatible with this proposed method. Although
it is highly encouraged to evaluate a particular SPT by the domain experts, however,
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 71

such qualitative assessment varies from one expert to another and therefore diffi-
cult to compare between SPTs for different recording protocols or different biosig-
nals. Therefore, it is high time to develop few standard evaluation criteria for the
SPTs which may consist of both qualitative and quantitative metrics to make it more
realistic and have a fair comparison.

2.4.3.3 Ground Truth Data

The unavailability of ground truth data (i.e. clean reference biosignals) is another
reason for not being able to evaluate performance of a SPT quantitatively. Thus it is
also essential to develop a public database with clean biosignal of all types, especially
EEG. This is because EEG is the most prone to noise and artifacts among all other
non-invasive biosignal recordings and there is no ground truth data of EEG found
to the best of our knowledge. So to record ground truth biosignals, recording proto-
cols and experiment should be designed carefully. In addition to that an acceptable
mathematical model to generate/simulate basic biosignals may also be developed
for quantitative evaluation of any existing or future SPTs. Finally, more research is
required to identify and characterize artifacts as many as possible. Therefore, it will
be feasible to label both ground truth biosignals as well as artifact templates.

2.5 Conclusions

The chapter summarizes different sources of artifacts and their characteristics found
in different biomedical signals and discusses the advances in signal processing tech-
niques and their suitability for handling these artifacts from different perspectives
in a variety of biosignal based applications starting from patient monitoring to
disease diagnosis, basic physiology/neuroscience research to brain-machine inter-
facing, evaluating therapeutic intervention to preventive healthcare, etc. The moti-
vation of deciding to use a particular SPT to remove artifacts must be followed by
enhancing the overall detection or classification performance, e.g. it in terms of reduc-
tion of false alarms for epileptic seizure detection or increasing accuracy for BCI
studies. If it does not add any significant value to the later-stage performance, then it is
better not to use that SPT for removing artifacts or noises since only improving SNR
will not make any sense then. Finally, this chapter discusses the current limitation
of the available SPTs, possible future challenges, and potential recommendations to
overcome those challenges. More research will be required in developing new SPTs
such that they can handle extreme motion artifacts during ambulatory recordings,
enabling online/real-time processing, allowing on-chip implementation (if appli-
cable), being compatible with cloud-computing, storing and utilizing the recorded
Big data for appropriate decision making and more importantly customizable for
different applications and/or different biosignal types.
72 Md. K. Islam et al.

Appendix

Abbreviation Definition Abbreviation Definition

AAR Automatic Artifact Removal LAMIC Lagged Auto-Mutual
Information Clustering
ABP Arterial Blood Pressure LFP Local Field Potential
ADJUST Automatic EEG artifact detector LMS Least Mean Square
based on the joint use of spatial
and temporal features
ALE Adaptve Line Enhancer LTI Linear Time Invariant
ANC Adaptive Noise Canceller MARA Multiple artifact rejection
algorithm
ANFIS Adaptive Neural Fuzzy MCA Morphological Component
Inference System Analysis
ANN Artificial Neural Network MCMC Markov Chain Monte Carlo
ARX Auto-Regressive Exogenous MEG Magnetoencephalography
ASR Artifact Subspace MI Motor Imagery
Reconstruction
AWGN Additive White Gaussian Noise MMSE Minimum Mean Square Error
BCG Ballistocardiography MUAP Muscle Unit Action Potential
BCI Brain-Computer-Interface MWF multichannel Wiener filter
BioSigKit BioSignal Analysis Kit NB Naïve Bayes
Bio-SP Biosignal-Specific Processing NEO Non-linear Energy Operator
(A toolbox)
Biosppy Biosignal Processing in Python OSET Open-source
(toolbox) electrophysiological toolbox
BSP biomedical signal processing PCA Principal Component Analysis
toolbox
BSS Blind Source Separation PCG Phonocardiogram
CCA Canonical Correlation Analysis PFEIFER Preprocessing Framework of
Electrograms Intermittently
Fiducialized from Experimental
Recordings
CWT Continuous Wavelet Transform PLI Power Line Interference
DWT Discrete Wavelet Transform PPG Photoplethysmogram
EAP Extracellular Action Potential PRANA Polygraphic Recording
Analyzer
ECG/EKG Electrocardiography PSD Power Spectral Density
ecg-kit A Matlab toolbox for REM Rapid Eye Movement
cardiovascular (ECG, EKG,
ABP, PPG) signal processing
ECoG Electrocorticography RLS Recursive Least Square
(continued)
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 73

(continued)
Abbreviation Definition Abbreviation Definition
ECoG Electrocorticography RMS Root Mean Squares
EEG Electroencephalography SAE Sparse Autoencoders
EEMD Ensemble Empirical Mode SCG Seismocardiography
Decomposition
EMD Empirical Mode Decomposition SNR Signal to Noise Ratio
EMG Electromyography SOS Second-Order Statistics
EOG Electrooculography SPT Signal Processing Techniques
ERP Event Related Potential SSS Subspace Signal Separation
FASTER Fully Automated Statistical SSVEP Steady State Visual Evoked
Thresholding for EEG Artifact Potential
Rejection
FFT Fast Fourier Transform STFT Short Time Fourier Transform
FIR Finite Impulse Response SVD Singular Value Decomposition
FLNN Functional Link Neural SVM Support Vector Machine
Network
FOOBI Fourth-order Tensor method SWT Stationary Wavelet Transform
FORCe Fully online and automated TDSEP Temporal De-correlation source
artifact removal for BCI SEParation
HCI Human-Computer Interfacing VBA Variational Bayes
Approximation
HMI Human-Machine Interfacing W-CCA Wavelet Enhanced CCA
IAP Intracellular Action Potential WFDB WaveForm DataBase
ICA Independent Component W-ICA Wavelet Enhanced ICA
Analysis
iEEG intracranial WNN Wavelet Neural Network
electroencephalography
IIR Infinite Impulse Response WPD Wavelet Packet Decomposition
k-NN K- Nearest Neighbor WT Wavelet Transform

References

1. Salerno, D.M., Zanetti, J.M., Poliac, L.C., Crow, R.S., Hannan, P.J., Wang, K., Goldenberg,
I.F., Tassel, R.A.V.: Exercise seismocardiography for detection of coronary artery disease.
Am. J. Noninvasive Cardiol. 6(5), 321–330 (1992)
2. Kabir, A.U., Bin Shahin, F., Islam, M.K.: Design and implementation of an EOG-based mouse
cursor control for application in human-computer interaction. In: 4th International Conference
on Data Processing and Robotics (ICDPR 2020). Singapore (2020)
3. Islam, M.K., Rastegarnia, A., Yang, Z.: Methods for artifact detection and removal from scalp
EEG: a review. Neurophysiol. Clin. Clin. Neurophys. 46(4–5), 287–305 (2016)
74 Md. K. Islam et al.

4. Islam, M.K., Rastegarnia, A., Nguyen, A.T., Yang, Z.: Artifact characterization and removal
for in vivo neural recording. J. Neurosci. Methods 15(226), 110–123 (2014)
5. Sweeney, K.: Motion artifact processing techniques for physiological signals. Ph.D. Thesis,
National University of Ireland Maynooth (2013)
6. Chang, B.S.: Atlas of Ambulatory EEG. Elsevier (2005)
7. Holmes, G.L., Jones, H.R., Moshé, S.L.: Clinical Neurophysiology of Infancy, Childhood,
and Adolescence. Elsevier, Amsterdam (2006)
8. Shahrokhi, Farzaneh, Abdelhalim, Karim, Serletis, Demitre, Carlen, Peter L., Genov, Roman:
The 128-channel fully differential digital integrated neural recording and stimulation interface.
IEEE Trans. Biomed. Circuits Syst. 4(3), 149–161 (2010)
9. Islam, M.K., Tuan, N.A., Zhou, Y., Yang, Z.: Analysis and processing of in-vivo neural
signal for artifact detection and removal. In: 5th International Conference on BioMedical
Engineering and Informatics. IEEE. pp. 437–442 (2012)
10. Delorme, Arnaud, Makeig, Scott: EEGLAB: an open source toolbox for. J. Neurosci. Methods
134, 9–21 (2004)
11. Keshtkaran, M.R., Yang, Z.: A fast, robust algorithm for power line interference cancellation
in neural recording. J. Neural Eng. 11(2), 026017 (2014)
12. O’Regan, S., Faul, S., Marnane, W.: Automatic detection of EEG artefacts arising from head
movements using EEG and gyroscope signals. Med. Eng. Phys. 35(7), 867–874 (2013)
13. Jafarifarmand, A., Badamchizadeh, M.A.: Artifacts removal in EEG signal using a new neural
network enhanced adaptive filter. Neurocomputing 1(103), 222–231 (2013)
14. Mognon, A., Jovicich, J., Bruzzone, L., Buiatti, M.: ADJUST: an automatic EEG artifact
detector based on the joint use of spatial and temporal features. Psychophysiology 48(2),
229–240 (2011)
15. Burger, C., van den Heever, D.J.: Removal of EOG artefacts by combining wavelet neural
network and independent component analysis. Biomed. Sign. Process. Control 1(15), 67–79
(2015)
16. Hu, J., Wang, C.S., Wu, M., Du, Y.X., He, Y., She, J.: Removal of EOG and EMG artifacts
from EEG using combination of functional link neural network and adaptive neural fuzzy
inference system. Neurocomputing 3(151), 278–287 (2015)
17. Schetinin, V., Schult, J.: The combined technique for detection of artifacts in clinical elec-
troencephalograms of sleeping newborns. IEEE Trans. Inf Technol. Biomed. 8(1), 28–35
(2004)
18. Bhattacharyya, S., Biswas, A., Mukherjee, J., Majumdar, A.K., Majumdar, B., Mukherjee, S.,
Singh, A.K.: Detection of artifacts from high energy bursts in neonatal EEG. Comput. Biol.
Med. 43(11), 1804–1814 (2013)
19. Shao, S.Y., Shen, K.Q., Ong, C.J., Wilder-Smith, E.P., Li, X.P.: Automatic EEG artifact
removal: a weighted support vector machine approach with error correction. IEEE Trans.
Biomed. Eng. 336–344 (2008)
20. Shoker, L., Sanei, S., Chambers, J.: Artifact removal from electroencephalograms using a
hybrid BSS-SVM algorithm. IEEE Sign. Process. Lett. 721–724 (2005)
21. Lawhern, V., Hairston, W.D., McDowell, K., Westerfield, M., Robbins, K.: Detection and
classification of subject-generated artifacts in EEG signals using autoregressive models. J.
Neurosci. Methods 208(2), 181–189 (2012)
22. Schwabedal, J.T., Sippel, D., Brandt, M.D., Bialonski, S.: Automated classification of sleep
stages and EEG artifacts in mice with deep learning. (2018). arXiv 1809.08443
23. Yang, B., Duan, K., Fan, C., Hu, C., Wang, J.: Automatic ocular artifacts removal in EEG
using deep learning. Biomed. Sign. Process. Control 43, 148–158 (2018)
24. Sakamoto, Y., Ishiguro, M., Kitagawa, G.: Akaike Information Criterion Statistics. Springer,
Dordrecht (1986)
25. Bengtsson, T., Cavanaugh, J.E.: An improved Akaike information criterion for state-space
model selection. Comput. Stat. Data Anal. 50(10), 2635–2654 (2006)
26. Šmídl, V., Anthony, Q.: The Variational Bayes Method in Signal Processing. Springer, New
York (2006)
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 75

27. Chen, Z., Babacan, S.D., Molina, R., Katsaggelos, A.K.: Variational Bayesian methods for
multimedia problems. IEEE Trans. Multimedia 16(4), 1000–1017 (2014)
28. Drémeau, A., Herzet, C., Daudet, L.: Boltzmann machine and mean-field approximation for
structured sparse decompositions. IEEE Trans. Sign. Process. 60(7), 3425–3438 (2012)
29. Zheng, Y., Fraysse, A., Rodet, Thomas: Efficient variational Bayesian approximation method
based on subspace optimization. IEEE Trans. Image Process. 24(2), 681–693 (2014)
30. Zhao, Q., Hu, B., Shi, Y., Li, Y., Moore, P., Sun, M., Peng, H.: Automatic identification
and removal of ocular artifacts in EEG—improved adaptive predictor filtering for portable
applications. IEEE Trans. Nanobiosci. 109–117 (2014)
31. Klados, M.A., Papadelis, C., Braun, C., Bamidis, P.D.: REG-ICA: a hybrid methodology
combining blind source separation and regression techniques for the rejection of ocular
artifacts. Biomed. Sign. Process. Control 6(3), 291–300 (2011)
32. Li, Y., Ma, Z., Lu, W., Li, Y.: Automatic removal of the eye blink artifact from EEG using an
ICA-based template matching approach. Physiol. Measure. 27(4), 425 (2006)
33. Flexer, A., Bauer, H., Pripfl, J., Dorffner, G.: Using ICA for removal of ocular artifacts in
EEG recorded from blind subjects. Neural Netw. 18(7), 998–1005 (2005)
34. Guerrero-Mosquera, C., Navia-Vázquez, A.: Automatic removal of ocular artefacts using
adaptive filtering and independent component analysis for electroencephalogram data. IET
Sign. Process. 6(2), 99–106 (2012)
35. Devuyst, S., Dutoit, T., Stenuit, P., Kerkhofs, M., Stanus, E.: Cancelling ECG artifacts in EEG
using a modified independent component analysis approach. EURASIP J. Adv. Sign. Process.
2008, 1–13 (2008)
36. Roy, V., Shukla, S.: Automatic removal of artifacts from EEG signal based on spatially
constrained ICA using daubechies wavelet. Int. J. Mod. Edu. Comput. Sci. (IJMECS) 6(7),
31–39 (2014)
37. Shoker, L., Sanei, S., Latif, M.A.: Removal of eye blinking artifacts from EEG incorporating a
new constrained BSS algorithm. In: Sensor Array and Multichannel Signal Workshop. IEEE.
pp. 177–181 (2004)
38. Akhtar, M.T., Mitsuhashi, W., James, C.J.: Employing spatially constrained ICA and wavelet
denoising, for automatic removal of artifacts from multichannel EEG data. Sign. Process.
92(2), 401–416 (2012)
39. James, C.J., Gibson, O.J.: Temporally constrained ICA: an application to artifact rejection in
electromagnetic brain signal analysis. IEEE Trans. Biomed. Eng. 50(9), 1108–1116 (2003)
40. Jung, T.P., Makeig, S., Humphries, C., Lee, T.W., Mckeown, M.J., Iragui, V., Sejnowski, T.J.:
Removing electroencephalographic artifacts by blind source separation. Psychophysiology
37(2), 163–178 (2000)
41. Jung, T.P., Makeig, S., Westerfield, M., Townsend, J., Courchesne, E., Sejnowski, T.J.:
Removal of eye activity artifacts from visual event-related potentials in normal and clinical
subjects. Clin. Neurophys. 111(10), 1745–1758 (2000)
42. Sweeney, K.T., McLoone, S.F., Ward, T.E.: The use of ensemble empirical mode decompo-
sition with canonical correlation analysis as a novel artifact removal technique. IEEE Trans.
Biomed. Eng. 60(1), 97–105 (2012)
43. Sweeney, K.T., Ward, T.E., McLoone, S.F.: Artifact removal in physiological signals—prac-
tices and possibilities. IEEE Trans. Inf Technol. Biomed. 16(3), 488–500 (2012)
44. Yong, X., Ward, R.K., Birch, G.E.: Artifact removal in EEG using morphological component
analysis. In: International Conference on Acoustics, Speech and Signal Processing. IEEE.
345–348 (2009)
45. Zheng, Y., Fraysse, A., Rodet, T.: Efficient variational Bayesian approximation method based
on subspace optimization. IEEE Trans. Image Process. 24(2), 681–693 (2015)
46. Hunyadi, Borbála, Van Huffel, Sabine, De Vos, Maarten, Clifton, D.A.: The Power of Tensor
Decompositions in Biomedical Applications. Machine Learning for Healthcare Technologies,
IET (2016)
47. Rankine, L., Stevenson, N., Mesbah, M., Boashash, B.: A nonstationary model of newborn
EEG. IEEE Trans. Biomed. Eng. 54(1), 19–28 (2006)
76 Md. K. Islam et al.

48. Stevenson, N., Rankine, L., Mesbah, M., Boashash, B.: Newborn EEG seizure simulation
using time–frequency signal synthesis. In: APRS Workshop on Digital Image Computing.
145–151 (2005)
49. Oostenveld, R., Fries, P., Maris, E., Schoffelen, J.M.: FieldTrip: open source software for
advanced analysis of MEG, EEG, and invasive electrophysiological data. Comput. Intell.
Neurosci. (2011)
50. Stephane, M.: A Wavelet Tour of Signal Processing: The Sparse Way. Academic Press, London
(1999)
51. Islam, M.K., Rastegarnia, A., Yang, Z.: A wavelet-based artifact reduction from scalp EEG
for epileptic seizure detection. IEEE J. Biomed. Health Inf. 20(5), 1321–1332 (2015)
52. Sweeney, K.T., Ayaz, H., Ward, T.E., Izzetoglu, M., McLoone, S.F., Onaral, B.: A method-
ology for validating artifact removal techniques for physiological signals. IEEE Trans. Inf
Technol. Biomed. 16(5), 918–926 (2012)
53. Rashed-Al-Mahfuz, M., Islam, M.R., Hirose, K., Molla, M.K.: Artifact suppression and anal-
ysis of brain activities with electroencephalography signals. Neural Regeneration Res. 8(16),
1500 (2013)
54. Molla, M.K., Islam, M.R., Tanaka, T., Rutkowski, T.M.: Artifact suppression from EEG
signals using data adaptive time domain filtering. Neurocomputing 15(97), 297–308 (2012)
55. Ye, Y., Cheng, Y., He, W., Hou, M., Zhang, Z.: Combining nonlinear adaptive filtering and
signal decomposition for motion artifact removal in wearable photoplethysmography. IEEE
Sens. J. 7133–7141 (2016)
56. Xu, L., Rabotti, C., Zhang, Y., Ouzounov, S., Harpe, P.J., Mischi, M.: Motion-artifact reduction
in capacitive heart-rate measurements by adaptive filtering. IEEE Trans. Instrument. Measure.
68(10), 4085–4093 (2018)
57. Xiong, F., Chen, D., Chen, Z., Dai, S.: Cancellation of motion artifacts in ambulatory ECG
signals using TD-LMS adaptive filtering techniques. J. Visual Commun. Image Representation
1(58), 606–618 (2019)
58. Saeed F., Gazem N., Patnaik S., Saed Balaid A., Mohammed, F.: Motion artifact reduction
algorithm using sequential adaptive noise filters and estimation methods for mobile ECG. In:
Recent Trends in Information and Communication Technology. IRICT. Springer, New York
(2017)
59. Kim, C.S., Sun, J., Liu, D., Wang, Q., Paek, S.G.: Removal of ocular artifacts using ICA and
adaptive filter for motor imagery-based BCI. IEEE/CAA J. Automatica Sinica (2017)
60. Quazi, M.H., Kahalekar, S.G.: Artifacts removal from EEG signal: FLM optimization-based
learning algorithm for neural network-enhanced adaptive filtering. Biocybernetics Biomed.
Eng. 37(3), 401–411 (2017)
61. Yang, B., Duan, K., Zhang, T.: Removal of EOG artifacts from EEG using a cascade of sparse
autoencoder and recursive least squares adaptive filter. Neurocomputing 214, 1053–1060
(2016)
62. Borowicz, A.: Using a multichannel Wiener filter to remove eye-blink artifacts from EEG
data. Biomed. Sign. Process. Control 1(45), 246–255 (2018)
63. Cohen, M.X.: Analyzing Neural Time Series Data: Theory and Practice. MIT Press,
Cambridge (2014)
64. Turnip, A.: Automatic artifacts removal of EEG signals using robust principal component anal-
ysis. In: 2nd International Conference on Technology, Informatics, Management, Engineering
& Environment. IEEE. pp. 331–334 (2014)
65. Turnip, A., Junaidi, E.: Removal artifacts from EEG signal using independent component
analysis and principal component analysis. In: 2nd International Conference on Technology,
Informatics, Management, Engineering and Environment. pp. 296–302 (2014)
66. Hamaneh, M.B., Chitravas, N., Kaiboriboon, K., Lhatoo, S.D., Loparo, K.A.: Automated
removal of EKG artifact from EEG data using independent component analysis and continuous
wavelet transformation. IEEE Trans. Biomed. Eng. 61(6), 1634–1641 (2013)
67. Castellanos, N.P., Makarov, V.A.: Recovering EEG brain signals: artifact suppression with
wavelet enhanced independent component analysis. J. Neurosci. Methods 158(2), 300–312
(2006)
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 77

68. Mahajan, R., Morshed, B.I.: Unsupervised eye blink artifact denoising of EEG data with
modified multiscale sample entropy, kurtosis, and wavelet-ICA. IEEE J. Biomed. Health Inf.
19(1), 158–165 (2014)
69. Zhao, C., Qiu, T.: An automatic ocular artifacts removal method based on wavelet-enhanced
canonical correlation analysis. In: Annual International Conference of the IEEE Engineering
in Medicine and Biology Society. IEEE. pp. 4191–4194 (2011)
70. Raghavendra, B.S., Dutt, D.N.: Wavelet enhanced CCA for minimization of ocular and muscle
artifacts in EEG. World Acad. Sci. Eng. Technol. 57(6), 1027–1032 (2011)
71. Mammone, N., Morabito, F.C.: Enhanced automatic wavelet independent component analysis
for electroencephalographic artifact removal. Entropy 16(12), 6553–6572 (2014)
72. Calcagno, S., La Foresta, F., Versaci, M.: Independent component analysis and discrete wavelet
transform for artifact removal in biomedical signal processing. Am. J. Appl. Sci. 11(1), 57
(2014)
73. Chen, X., He, C., Peng, H.: Removal of muscle artifacts from single-channel EEG based on
ensemble empirical mode decomposition and multiset canonical correlation analysis. J. Appl.
Math (2014)
74. Zeng, H., Song, A., Yan, R., Qin, H.: EOG artifact correction from EEG recording using
stationary subspace analysis and empirical mode decomposition. Sensors 14839–14859
(2013)
75. Niazy, R.K., Beckmann, C.F., Iannetti, G.D., Brady, J.M., Smith, S.M.: Removal of FMRI
environment artifacts from EEG data using optimal basis sets. Neuroimage 28(3), 720–737
(2005)
76. Nguyen, H.A., Musson, J., Li, F., Wang, W., Zhang, G., Xu, R., Richey, C., Schnell, T.,
McKenzie, F.D., Li, J.: EOG artifact removal using a wavelet neural network. Neurocomputing
15(97), 374–389 (2012)
77. Peng, H., Hu, B., Shi, Q., Ratcliffe, M., Zhao, Q., Qi, Y., Gao, G.: Removal of ocular artifacts
in EEG—an improved approach combining DWT and ANC for portable applications. IEEE
J. Biomed. Health Inf. 17(3), 600–607 (2013)
78. Navarro, X., Porée, F., Carrault, G.: ECG removal in preterm EEG combining empirical mode
decomposition and adaptive filtering. In: International Conference on Acoustics, Speech and
Signal Processing (ICASSP). IEEE. pp. 661–664 (2012)
79. Skupch, A.M., Dollfuss, P., Fürbaß, F., Gritsch, G., Hartmann, M.M., Perko, H., Pataraia,
E., Lindinger, G., Kluge, T.: Spatial correlation based artifact detection for automatic seizure
detection in EEG. In: 35th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC). IEEE. pp. 1972–1975 (2013)
80. Hesar, H.D., Mohebbi, M.: An adaptive particle weighting strategy for ECG denoising using
marginalized particle extended Kalman filter: An evaluation in arrhythmia contexts. IEEE J.
Biomed. Health Inf. 21(6), 1581–1592 (2017)
81. Hesar, H.D., Mohebbi, M.: ECG denoising using marginalized particle extended kalman filter
with an automatic particle weighting strategy. IEEE J. Biomed. Health Inf. 21(3), 635–644
(2016)
82. Warmerdam, G.J., Vullings, R., Schmitt, L., Van Laar, J.O., Bergmans, J.W.: A fixed-lag
Kalman smoother to filter power line interference in electrocardiogram recordings. IEEE
Trans. Biomed. Eng. 64(8), 1852–1861 (2016)
83. Daly, I., Scherer, R., Billinger, M., Müller-Putz, G.: FORCe: fully online and automated
artifact removal for brain-computer interfacing. IEEE Trans. Neural Syst. Rehabil. Eng. 23(5),
725–736 (2014)
84. Hallez, H., De Vos, M., Vanrumste, B., Van Hese, P., Assecondi, S., Van Laere, K.: Removing
muscle and eye artifacts using blind source separation techniques in ictal EEG source imaging.
Clin. Neurophys. 120(7), 1262–1272 (2009)
85. De Clercq, W., Vergult, A., Vanrumste, B., Van Paesschen, W., Van Huffel, S.: Canonical
correlation analysis applied to remove muscle artifacts from the electroencephalogram. IEEE
Trans. Biomed. Eng. 53(12), 2583–2587 (2006)
78 Md. K. Islam et al.

86. Kolda, T.G., Bader, B.W.: Tensor decompositions and applications. SIAM Rev. 51(3), 455–500
(2009)
87. Wang, Z., Xu, P., Liu, T., Tian, Y., Lei, X., Yao, D.: Robust removal of ocular artifacts by
combining independent component analysis and system identification. Biomed. Sign. Process.
Cont. 1(10), 250–259 (2014)
88. De Vos, M., Deburchgraeve, W., Cherian, P.J., Matic, V., Swarte, R.M., Govaert, P., Visser,
G.H., Van Huffel, S.: Automated artifact removal as preprocessing refines neonatal seizure
detection. Clin. Neurophys. 122(12), 2345–2354 (2011)
89. Kothe, C.A., Makeig, S.: BCILAB: a platform for brain–computer interface development. J.
Neural Eng. 10(5), 056014 (2013)
90. Nabian, M., Yin, Y., Wormwood, J., Quigley, K.S., Barrett, L.F., Ostadabbas, S.: An open-
source feature extraction tool for the analysis of peripheral physiological data. IEEE J. Transl.
Eng. Health Med. 6, 1–11 (2018)
91. Vidaurre, C., Sander, T.H., Schlögl, A.: BioSig: the free and open source software library for
biomedical signal processing. Comput. Intell. Neurosci. (2011)
92. Blechert, J., Peyk, P., Liedlgruber, M., Wilhelm, F.H.: ANSLAB: Integrated multichannel
peripheral biosignal processing in psychophysiological science. Behav. Res. Methods 48(4),
1528–1545 (2016)
93. n.d. Clean_rawdata EEGLAB plugin. Accessed 09 Feb 2020. https://github.com/sccn/clean_
rawdata
94. Sedghamiz, H.: BioSigKit: a matlab toolbox and interface for analysis of biosignals. J. Open
Sour. Softw. 3(30), 671 (2018)
95. n.d. BioSPPy. Accessed 09 Feb 2020. https://biosppy.readthedocs.io/en/stable/index.html
96. Silva, I., Moody, G.B.: An open-source toolbox for analysing and processing physionet
databases in matlab and octave. J. Open Res. Softw. (2014)
97. Aboy, M., Crespo, C., McNames, J., Bassale, J., Jenkins, L., Goldsteins, B.: A biomedical
signal processing toolbox. Proc. Biosignal 2002(26), 49–52 (2002)
98. McGill, K.C., Lateva, Z.C., Marateb, H.R.: EMGLAB: an interactive EMG decomposition
program. J. Neurosci. Methods 149(2), 121–133 (2005)
99. Chaumon, M., Bishop, D.V., Busch, N.A.: A practical guide to the selection of indepen-
dent components of the electroencephalogram for artifact correction. J. Neurosci. Methods
30(250), 47–63 (2015)
100. Nolan, H., Whelan, R., Reilly, R.B.: FASTER: fully automated statistical thresholding for
EEG artifact rejection. J. Neurosci. Methods 192(1), 152–162 (2010)
101. Sameni, R.: OSET: The Open-Source Electrophysiological Toolbox. Accessed 09 Feb 2020.
https://gitlab.com/rsameni/OSET. (2012)
102. Gómez-Herrero, G.: Automatic artifact removal (AAR) toolbox v1. 3 (Release 09.12. 2007)
for MATLAB. Technical Manual, Tampere University of Technology (2007)
103. Ge, S., Han, M., Hong, X.: A fully automatic ocular artifact removal from EEG based on
fourth-order tensor method. Biomed. Eng. Lett. 4(1), 55–63 (2014)
104. Nicolaou, N., Nasuto, S.J.: Automatic artefact removal from event-related potentials via
clustering. J. VLSI Sign. Process. Syst. Sign. Image Video Technol. 48(1–2), 173–183 (2007)
105. n.d. PRANA® Software Suite. Accessed 09 Feb 2020. https://www.phitools.com/prana.php
106. Rodenhauser, A., Good, W.W., Zenger, B., Tate, J., Aras, K., Burton, B., MacLeod,
R.S.: Pfeifer: preprocessing framework for electrograms intermittently fiducialized from
experimental recordings. J. Open Sour. Softw. 3(21), 472 (2018)
107. François, T., Baillet, S., Mosher, J.C., Pantazis, D., Leahy, R.M.: Brainstorm: a user-friendly
application for MEG/EEG analysis. Comput. Intell. Neurosci. (2011)
108. Pion-Tonachini, L., Hsu, S.H., Makeig, S., Jung, T.P., Cauwenberghs, G.: Real-time EEG
source-mapping toolbox (REST): online ICA and source localization. In: 37th Annual Inter-
national Conference of the IEEE Engineering in Medicine and Biology Society (EMBC).
pp. 4114–4117 (2015)
109. n.d. Bergen fMRI Toolbox. Accessed 09 Feb 2020. http://fmri.uib.no/index.php?option=com_
content&view=article&id=53&Itemid=78
2 Signal Artifacts and Techniques for Artifacts and Noise Removal 79

110. Viola, F.C., De Vos, M., Hine, J., Sandmann, P., Bleeck, S., Eyles, J., Debener, S.: Semi-
automatic attenuation of cochlear implant artifacts for the evaluation of late auditory evoked
potentials. Hearing Res. 284(1–2), 6–15 (2012)
111. Viola, F.C., Thorne, J., Edmonds, B., Schneider, T., Eichele, T., Debener, S.: Semi-automatic
identification of independent components representing EEG artifact. Clin. Neurophysiol.
120(5), 868–877 (2009)
112. Winkler, I., Haufe, S., Tangermann, M.: Automatic classification of artifactual ICA-
components for artifact removal in EEG signals. Behav. Brain Funct. 7(1), 30 (2011)
113. Winkler, I., Brandl, S., Horn, F., Waldburger, E., Allefeld, C., Tangermann, M.: Robust artifac-
tual independent component classification for BCI practitioners. J. Neural Eng. 11(3), 035013
(2014)
114. n.d. NITRC CleanLine Tool. Accessed 09 Feb 2020. https://www.nitrc.org/projects/cleanline
115. Mohseni, H.R., Ghaderi, F., Wilding, E.L., Sanei, S.: Variational Bayes for spatiotemporal
identification of event-related potential subcomponents. IEEE Trans. Biomed. Eng. 57(10),
2413–2428 (2010)
116. Hubel, D.H., Wiesel, T.N.: Receptive fields, binocular interaction and functional architecture
in the cat’s visual cortex. J. Physiol. 160(1), 106 (1962)
117. Fisher, R.S., Vickrey, B.G., Gibson, P., Hermann, B., Penovich, P., Scherer, A., Walker, S.: The
impact of epilepsy from the patient’s perspective I. Descriptions and subjective perceptions.
Epilepsy Res. 41(1), 39–51 (2000)
118. n.d. Plexon Systems. Accessed 08 Feb 2020. http://www.plexon.com/products/
119. Csicsvari, J., Henze, D.A., Jamieson, B., Harris, K.D., Sirota, A., Barth, P., Wise, K.D.,
Buzsaki, G.: Massively parallel recording of unit and local field potentials with silicon-based
electrodes. J. Neurophysiol. 90(2), 1314–1323 (2003)
120. O’Regan, S.H.: Artefact detection and removal algorithms for EEG diagnostic systems (2013)
121. Savelainen, A.: Movement artifact detection from electroencephalogram utilizing accelerom-
eter. Helsinki, Finland: M.S. thesis, School of Sci. and Tech., Aalto Univ (2011)
122. Savelainen, A.: An Introduction to EEG Artifacts. School of Sci. Aalto Univ, Finland (2010)
123. Babacan, S.D., Molina, R., Katsaggelos, A.K.: Variational Bayesian super resolution. IEEE
Trans. Image Process. 20(4), 984–999 (2011)
124. Hubel, D.H., Wiesel, T.N.: Receptive fields of single neurones in the cat’s striate cortex. J.
Physiol. 148(3), 574 (1959)