International Journal of Computer Applications (0975 – 8887)

Volume 169 – No.3, July 2017

Brain Computer Interface: EEG Signal Preprocessing

Issues and Solutions

Nelly Elsayed Zaghloul Saad Zaghloul Magdy Bayoumi

Center for Advanced Computer Center for Advanced Computer Center for Advanced Computer
Studies (CACS) Studies (CACS) Studies (CACS)
University of Louisiana at University of Louisiana at University of Louisiana at
Lafayette, Lafayette, LA, USA Lafayette, Lafayette, LA, USA Lafayette, Lafayette, LA, USA

ABSTRACT Brain cells communicate with each other through electrical

Brain Computer Interface (BCI) is often directed at mapping, impulses that structure the brain signal. There are several
assisting, or repairing human cognitive or sensory-motor techniques for brain signal acquisition: invasive, semi-
functions. Electroencephalogram (EEG) is a non-invasive invasive, and non-invasive [10, 19]. EEG technique (which
method of acquisition brain electrical activities. Noises are this paper concern) is a non-invasive technique for brain
impure the EEG recorded signal due to the physiologic and signal acquisition. It records brain signal along the scalp
extra-physiologic artifacts. There are several techniques are through measuring voltage fluctuations accompanying
intended to manipulate the EEG recorded signal during the neurotransmission activity within the brain. The main
BCI preprocessing stage of to achieve preferable results at the advantages of EEG are that it is non-invasive so that there is
learning stage. This paper aims to present an overview on BCI no surgical intervention needed to set the electrodes on the
different EEG brain signal recording artifacts and the brain, relative ease of use, portable, excellent time resolution
methodologies to remove these artifacts from the signal (it gives the ability to see brain activity as it unfolds in real
focusing on different novel trends at BCI research areas. time, at the level of milliseconds (thousandths of a second)),
low-cost comparing to the other brain signal acquisition
General Terms techniques. The disadvantages of EEG that its very small
Brain Computer interface, EEG Signal Filtering, Machine signal to noise ratio and spatial localization resolution
learning. represent a limitation compared to other brain signal
acquisition methods. According to the very small signal to
Keywords noise at the EEG signal recorded, it cannot be directly used at
Brain Computer interface (BCI), EEG, artifact removal, processing stage such as feature extraction and pattern
preprocessing, EMG, EOG, filtering. recognition directly at the BCI system. The EEG recorded
signal has to pass through filtering and artifact removing at
1. INTRODUCTION the preprocessing stage at BCI system. Figure 2 illustrates the
The BCI research has arisen at the last decade. BCI research BCI system stages and its components [20].
and development has focused primarily on neuro-prosthetics
This paper organized as follows: Section 2 explains different
applications that aim to restore damaged hearing, sight and
kinds of artifacts that contaminated the brain signal acquired
movement [1]. Nowadays, BCI research reached remarkable
using EEG. Section 3 shows the different methodologies at
results at robot control [6, 8, 9, 12], motor disabilities
novel BCI applications that targeting these artifacts removal,
recovery using BCI chip implants [1, 4, 5, 7, 11], medical
finally a brief conclusion at section 4.
diagnoses and prediction [15, 18], security and authentications
[14] and game controlling [10, 16, 17]. Figure 1 illustrates
different BCI research.

BCI
research
areas

Vision and
Robot Motor Medical Security and Game
hearing
Control disabilities diagnouses authentications controlling
recovery

Paralysis
Artificial Early
functionality Detection
limbs prediction
recovery

Fig 1: BCI research areas

12
 International Journal of Computer Applications (0975 – 8887)
Volume 169 – No.3, July 2017

Fig 2: BCI Stages

2. EEG ARTIFACTS 2.2.4 Interference of other Equipment

EEG recorded signal artifacts are all the signals that This kind of artifacts is also named as powerline noise. High-
contaminated the brain signal during the process of its frequency radiation from different devices such as radio,
acquisition. There are two types of artifacts: extra- television, cell phone, hospital paging systems, light
physiologic and physiologic artifacts. Extra-physiologic florescent, and equipment which is captures by the electrode
artifacts are occurred due to external environment rather than interference can overload EEG amplifiers. In this case, the
the human body during brain signal acquisition such as EEG signal cannot be recorded because pens may fully be
(environment, equipment). Physiologic artifacts occurred due deflected by swerve upward or downward. The optimal
to the human body, and arose from sources other than the solution to solve this problem is turn off these different
brain such as (eye blink, eyeball movement, breath, heart devices during EEG machine recording the brain signal.
beats, muscles movement) [20].
2.2 Physiologic Artifacts
2.1 Extra-Physiologic Artifacts
2.2.1 Electrooculography (EOG)
2.1.1 (60-Hz) Alternating Current Artifact EOG is the signal artifact occurs from eye movement
The 60-Hz alternating is one EEG acquisition technical functionality during EEG recording. EOG overlapping the
preparation problems. It happened when there is no sufficient EEG brain recorded signal. It is useful only in the medical
grounding of the EEG cap electrodes on patient’s head. The case of identifying sleep stages at the EEG signal. Bell
impedance between the ground of the amplifier and the active phenomena is observed by the vertical eye movements (i.e.
electrodes of the EEG cap becomes considerably large. At this eye blinks) which producing symmetric downward deflections
case, the ground acts like an electrode which produces 60-Hz for the front polar electrodes recording while the lateral eye
artifact and causes indistinct signal recording. movements affect lateral frontal electrodes recording [21].

2.1.2 Electrode Artifacts 2.2.2 Electrocardiogram (EKG or ECG)

Electrode popping and incorrect location of EEG electrode ECG (or EKG) artifacts are considered as the largest artifacts
placement are the most common electrode artifacts at EEG that could be found in the EEG recorded signal especially for
brain signal acquisition. Incorrect location of EEG electrode those patients having wide and short necks. ECG is caused by
placement is considered as technical preparation problem. the interference of the heart beats and moves potentials over
Incorrect placement of the electrode on patient’s head patient’s scalp surface [24]. Due to the regularity
generates an incorrect acquisition of EEG signal. Electrode (rhythmicity) and coincidence of the ECG artifact, it could be
popping artifact happened due to abrupt impedance change at easily detected and recognized through the EEG signal
the acquiesced signal. It could be identified visually as it recording. The problem of detecting this artifact occurs at the
causes multiple sharp waveforms that bind to the EEG signal case when the cerebral abnormal activity overlapped with
[25]. other EEG signal artifacts and its appearance could be
overlooked [25].
2.1.3 Movements in recording environment
Any movement of other persons around the patient can 2.2.3 Pulse Artifact
generate artifacts at signal recording, usually this artifact of There is a direct relationship between the pulse waves and the
capacitive or electrostatic origin. This should be avoided as ECG. Pulse waves artifact occurs when the EEG electrode in
much as possible, if avoidance is not possible as in the places over a pulsating vessel causes slow waves that may
operating room and the intensive care unit (ICU), proper simulate EEG activity [25]. The easiest solution to reduce this
notation (remarks) must be placed on the records of the signal artifact is by replacing the electrode location on the scalp.
recorded.

13
 International Journal of Computer Applications (0975 – 8887)
Volume 169 – No.3, July 2017

2.2.4 Skin artifacts between either both EEG and EOG (or EMG) signals or both
Skin sweat (perspiration or diaphoresis) and not clean hairy EMG and EOG signals, the filter fails to filter the EEG signal
leather of the patient are the major skin artifacts. Sweat is a and could not remove the EOG and EMG artifacts from the
fluid that secreted by sweat glands. Sweat major components EEG signal.
are lactic acid and sodium chloride that causing enormous
slow baseline due to their reaction with electrodes metals [24]. 3.3 Linear Adaptive Filter
The hairy leather artifact increases the impedance between the Linear-adaptive filter is introduced at [2] is mainly based on
electrode and the skin which effect on the recorded EEG the optimization theory and how the filter design can modify
signal correctness. its propertied according to the selected feature from the signal
being analyzed. The basic idea of the adaptive filter is to
2.2.5 Glossokinetic artifact adapt the frequency response (coefficient) of the linear filter
This artifact is generated by the movement of the tongue. to generate a signal similar to the noise existing in the signal
Glossokinetic artifact is similar to the EOG artifact caused by to be filtered. This linear-adaptive filter algorithm as
the eye movement although it is less steep than that produced considers that there are two signals primary signal and
by eye movement artifacts [26]. Glossokinetic artifact has a secondary signal, d(n) and x(n) respectively. H(z) is a linear
broad potential field that drops from frontal to occipital areas. filter which produces an output y(n). To compute the error
In this artifact the amplitude of the parasagittal is smaller e(n) the output value y(n) is subtracted from the primary
inferiorly than in potentials regions. signal d(n) [2, 23]. Figure 3 illustrates the linear-adaptive
basic structure. The two major processes of this filtering
2.2.6 Electromyogram (EMG) algorithm are: filtering process and adaptive process. At the
EMG is the signal generated by muscles movement. The most filtering process the output signal is the response of a digital
common artifact effect on EEG signal is the myogenic filter. In this process, FIR filters are commonly used because
potentials that are caused by clenching of jaw muscles they are simple, linear and stable filters. At the adaptive
(frontalis and temporalis muscles). The potentials generated process, The adjustment of transfer function H(z) is according
by the brain are longer duration than those that are generated to an optimizing algorithm. This adaptation is directed by the
by muscles. These artifacts are identified on the frequency, primary signal and the filter output difference (error) value.
duration, and morphology of the signal [13]. Frontalis and
temporalis muscles (e.g., clenching of jaw muscles) are
common causes. Generally, the potentials generated in the
muscles are of shorter duration than those generated in the
brain and are identified easily on the basis of duration,
morphology, and rate of firing (i.e., frequency) [26].

3. ARTIFACTS REMOVAL
At the BCI system, EEG signals filtering done at the
preprocessing stage. There are several approaches that are
designed to eliminate the noise that generated by various
artifacts from the EEG signal. This section demonstrates
different type of most remarkable filters approaches that are
used in at BCI applications. Fig 3: The linear-adaptive filter structure

3.1 Artifacts Avoidance 3.4 Quantum Neural Network-Based Filter

The idea here is to remove the extra-physiologic artifacts by Due to Moore’s law, the interest of finding new hardware
correctly assigning the EEG cap on the patient and preparing design rather than the classical computers led to various
the place of the recording of other devices that can effect on innovations at quantum computations and quantum inspired
the EEG signal acquisition from the patient. It is not an algorithms. Quantum inspired algorithms achieved
algorithm to be applied but it is a technical preparation for remarkable results at various algorithms [3]. One example of
acquisition a high quality record before signal filtering from quantum inspired algorithms for EEG signal filtering is the
physiologic artifacts. recurrent quantum neural network (RQNN). RQNN filter
improves the acquired EEG brain signal quality and
3.2 Linear Filtering segregation of signal artifacts. RQNN filter mechanism is
Linear filtering is commonly used for removing artifacts combining between both the recurrent neural networks (RNN)
located in certain signals that do not have any overlapped and quantum computer architecture mechanisms and
frequency bands with the frequency of the brain signals. Low- concepts. RQNN quantum mechanism which based on
pass filtering and high-pass filtering are the most noticeable Schrodinger wave equation trigged the recurrent neural
examples for linear filtering algorithms [23]. High-pass filter network to efficient dealing with the EEG signal filtering, in
is commonly used to remove the EOG artifacts as the EOG addition that the RNN the is using an unsupervised learning
artifacts consist of low frequency components. Low-pass filter algorithm, it enables the filter to be a powerful tool for
may remove most of EMG artifacts as most of EMG capturing the EEG brain signal statistical behaviors. The
components have high-frequency values. RQNN filter shows a significant result for noise to signal
ration for EEG signal filtering comparing to the RNN filter
Ease implementation from software and hardware
approach [3]. The major disadvantage of the neural network-
perspectives is one of the major advantages of linear filter. In
based filters that they are hard to be implemented on a
addition, it does not need any extra information about the wearable hardware chip with size implementation limits due
artifact itself, for example, to remove the EOG or the EMG to the huge requirement of wires and circuits to implement the
noises from the EEG signal the linear filter does not need
neural network, so that is it not suitable for multiple medical
predefining of their signal information [2]. The major
applications such as seizure prediction chips [22, 15] and
drawback of linear filter that in case of any overlapping

14
 International Journal of Computer Applications (0975 – 8887)
Volume 169 – No.3, July 2017

Alzheimer's disease detection [18]. However it can be applied all the hardware constraints should be determined to
at the BCI game applications where there are no restrictions manipulate to select the convenient technique for EEG brain
on the hardware implementation size. signal filtering and preprocessing. The BCI applications have
specific needs for several artifact removals from EEG brain
3.5 Spatial filtering signal recording. This paper show the most remarkable filter
Spatial filtering concept is to use a small number of new approaches for different BCI applications. The medical
channels that are a linear combination of the original channels applications especially that are related to on-line disease
of EEG brain signal reading [20]. Generally, the spatial filter detection such as epilepsy seizure have limited restrictions on
is defined as the following: the hardware implementation. In such case a good filtering
technique with light-weight hardware implementation is
needed. The game application and security applications that
have no limits on the hardware implementation can use more
powerful filtering techniques such as neural networks.
Where the spatially filtered signal, xi the EEG signal from
channel i, the weight given to that channel in the spatial 5. REFERENCES
filter and X is the original matrix of EEG brain signals from [1] E. A. Curran and S. J. Maria, S. J, "Learning to control
all channels. brain activity: A review of the production and control of
EEG components for driving brain–computer interface
Spatial filtering has a neurophysiological meaning because it (BCI) systems", Brain and Cognition, vol. 51, no. 3,
helps to recover brain original signal by gathering the relevant April 2003, pp. 326–336.
information that is spread over different EEG channels. In
addition, it reduces the dimensions of the EEG channels to [2] M. Honing and D. Messerschmitt, "Adaptive Filters:
less number of spatially filtered signals. So that spatial Structures, Algorithms and Applications", ISBN-13: 978-
filtering can be used for space reduction, filtering the signal 0898381634, 1984, Kluwer Academic Publishers.
and recover the original brain signal simultaneously.
[3] V. Gandhi, V. Arora, G. Prasad, D. Coyle and T. M.
3.6 Cauchy-Based Filter McGinnity, “A novel EEG signal enhancement approach
The Cauchy-based filter technique presented at [15] is used at using a recurrent quantum neural network for a brain–
seizure detection chip. This filter hardware implementation computer interface,” in Proc. 3rd Eur. Conf., Tech.
can be applied under the size limitation of BCI chips for brain Assist. Rehabil., Mar. 2011, pp. 1–8.
medical detection chips. The proposed Cauchy distribution- [4] Z. S. Zaghloul and M. Bayoumi, "Implementable Spike
based filter targets the impulsive observation noises since it is Sorting techniques for VLSI wireless BCI/BMI implants:
a heavy-tailed distribution which is more powerful than the A survey," 5th International Conference on Energy
Gaussian distribution to capture abrupt changes in noise. Aware Computing Systems & Applications, Cairo, 2015,
Cauchy distribution characterizes the impulsive artifacts in pp. 1-4
EEG observation such as EMG and EOG. The Cauchy
distribution probability density function is: [5] P. Horki, T. Solis-Escalante, C. Neuper, "Combined
motor imagery and SSVEP based BCI control of a 2 DoF
artificial upper limb" , Medical & Biological Engineering
& Computing, May 2011, vol. 49, no. 5, pp 567–577
[6] A. Jackson, C. T. Moritz, J. Mavoori, T. H. Lucas and E.
Where μ is the location parameter denoting the location of the E. Fetz, "The neurochip BCI: towards a neural prosthesis
peak of Cauchy distribution and s is the scale parameter that for upper limb function," in IEEE Transactions on Neural
specifies the half-width at half-maximum. Systems and Rehabilitation Engineering, vol. 14, no. 2,
June 2006, pp. 187-190.
The noise observation function is defined at the following:
[7] Z. S. Zaghloul and M. Bayoumi, "Toward fast low power
adaptive spike sorting VLSI chip design for wireless BCI
implants," 2015 IEEE 58th International Midwest
Where zt is a n-dimensional feature vector standing for an Symposium on Circuits and Systems (MWSCAS), Fort
observation at time t, xt is defined as: Collins, CO, 2015, pp. 1-4.
[8] Y. Chae, J. Jeong and S. Jo, "Toward Brain-Actuated
Humanoid Robots: Asynchronous Direct Control Using
Where denotes xt to the power i, W is an n×(q +1) matrix an EEG-Based BCI," in IEEE Transactions on Robotics,
containing coefficients of the polynomial function, and ut is vol. 28, no. 5, Oct. 2012, pp. 1131-1144.
the observation noise of the signal.
[9] C. Guger, S. Daban, , E.Sellers, C. Holzner, G. Krausz,
EOG and EMG effect with abrupt changes on the EEG brain R. Carabalona, F. Gramatica, G. Edlinger, "How many
signal such as large amplitudes and high frequencies. Cauchy people are able to control a P300-based brain–computer
filter able efficiently to filter these artifacts especially at interface (BCI)?, " Neuroscience Letters, vol. 462, no.1,
medical applications such as Epilepsy Seizures. 18 September 2009, pp. 94-98.
4. CONCLUSION [10] S. Abdulkader, A. Atia , M. M.Mostafa, "Brain computer
EEG brain signal filtering is a significant step at BCI interfacing: Applications and challenges," Egyptian
preprocessing stage to prepare an efficient signal for detection Informatics Journal, vol. 16, no. 2, July 2015, pp. 213–
and classification learning stages. In order to effectively 230.
choose and use methods for dealing with artifact removal of
EEG brain signal, the BCI application should be decided and

15
 International Journal of Computer Applications (0975 – 8887)
Volume 169 – No.3, July 2017

[11] Z. Saad Zaghloul and M. Bayoumi, "Adaptive neural using spontaneous EEG recordings, "Physiological
matching online spike sorting VLSI chip design for Measurement, vol. 28, 2007, pp. 335–347.
wireless BCI implants," 2015 IEEE International
Conference on Acoustics, Speech and Signal Processing [19] R. Oostenveld, P. Fries, E. Maris, and J. Schoffelen,
(ICASSP), South Brisbane, QLD, 2015, pp. 977-981. "FieldTrip: open source software for advanced analysis
of MEG, EEG, and invasive electrophysiological data,"
[12] D. Nurseitov, A. Serekov, A. Shintemirov and B. Neuroscience, article 1, January 2011, 9 pages.
Abibullaev, "Design and evaluation of a P300-ERP based
BCI system for real-time control of a mobile robot," [20] J. R. Wolpawa,b, N. Birbaumerc, D. J. McFarlanda, G.
2017 5th International Winter Conference on Brain- Pfurtschellere, T. M. Vaughan, "Brain–computer
Computer Interface (BCI), Sabuk, 2017, pp. 115-120. interfaces for communication and control," Clinical
Neurophysiology, vol. 113, March 2002, pp. 767–791.
[13] M. van de Velde, G. van Erp, PJ. Cluitmans PJ.,
"Detection of muscle artifact in the normal human awake [21] T. W. Picton, P. van Roon, M. L. Armilio, "The
EEG," Electroencephalography and Clinical correction of ocular artifacts: a topographic perspective,"
Neurophysiology, vol. 107, no. 2, April 1998, pp. 149– Clinical Neurophysiology, vol. 111, no. 1, January
158. 2000,pp. 53–65.

[14] M. K. Bashar, I. Chiaki and H. Yoshida, "Human [22] N. Elsayed, Z. S. Zaghloul, M. Bayoumi, "BCI/AIS Low
identification from brain EEG signals using advanced Power Adaptive Architecture for Early Prediction of
machine learning method EEG-based biometrics," 2016 Epilepsy Seizures," 15th IEEE International New
IEEE EMBS Conference on Biomedical Engineering and Circuits and Systems Conference (NEWCAS),
Sciences (IECBES), Kuala Lumpur, 2016, pp. 475-479. Strasbourg, France, June 2017, in press.

[15] Y. Wang et al., "A Cauchy-Based State-Space Model for [23] A. GuruvaReddy, S. Narava, "Artifact Removal from
Seizure Detection in EEG Monitoring Systems," in IEEE EEG Signals," International Journal of Computer
Intelligent Systems, vol. 30, no. 1, pp. 6-12, Jan.-Feb. Applications, vol. 77, no. 13, September 2013, pp. 17-19.
2015. [24] D. Craven, B. McGinley, L. Kilmartin, M. Glavin and E.
[16] D. Coyle, J. Principe, F. Lotte and A. Nijholt, "Guest Jones, "Adaptive Dictionary Reconstruction for
Editorial: Brain/neuronal - Computer game interfaces Compressed Sensing of ECG Signals," in IEEE Journal
and interaction," in IEEE Transactions on Computational of Biomedical and Health Informatics, vol. 21, no. 3,
Intelligence and AI in Games, vol. 5, no. 2, pp. 77-81, May 2017, pp. 645-654.
June 2013. [25] S. R. Benbadis and K. Lin, "Errors in EEG Interpretation
[17] L. Liao, C. Chen, I. Wang, S. Chen, S. Li, B. Chen, J. and Misdiagnosis of Epilepsy," European Neurology,
Chang and C. Lin, "Gaming control using a wearable and vol. 59, 2008, pp. 267–271.
wireless EEG-based brain-computer interface device [26] D. V. Morettia, F. Babilonia, F. Carduccia, F. Cincottia,
with novel dry foam-based sensors," Journal of E. Remondinia, P.M. Rossinib, S. Salinarie, C. Babiloni,
NeuroEngineering and Rehabilitation, vol. 9, no. 5, "Computerized processing of EEG–EOG–EMG artifacts
January 2012, pp. 1-11. for multi-centric studies in EEG oscillations and event-
[18] W. L. Woon, A. Cichocki, F. Vialatte and T, Musha, related potentials," International Journal of
"Techniques for early detection of Alzheimer's disease Psychophysiology, vol. 47, no. 3, March 2003, pp. 199–
216.

IJCATM : www.ijcaonline.org 16