Ubc 2012 Fall Adib Mani

A New Method for Removing
Electrical Vestibular Stimulation

Induced Artifacts from
Electro-Encephalographic Recordings
by
Mani Adib
B.Sc., Sharif University of Technology
M.Sc., Sharif University of Technology
A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF
THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF APPLIED SCIENCE
in
The Faculty of Graduate Studies
(Electrical and Computer Engineering)
THE UNIVERSITY OF BRITISH COLUMBIA

(Vancouver)
August 2012

c Mani Adib 2012
Abstract
In this work, we present a new method for removing artifacts from scalp
Electro-Encephalography (EEG) signals recorded during Electrical Vestibu-
lar Stimulation (EVS). Using EVS, we stimulate the vestibular nerves, which
can affect different regions in the brain via the interconnection of the vestibu-
lar system with some regions in the brain. As a result, some of the brain
functions can be altered during the EVS application.
Throughout its long history, EVS has been found as an interesting research
tool in physiology and neurology. Various applications of EVS have been
implemented in the health-care industry and also in other industries such as
entertainment. Although there have been many advances in the EVS appli-
cations, it still remains a challenging problem to understand how the EVS
stimulus acts as an input to the brain and how the brain responds. In this
study, we monitored and recorded the brain activities during the application
of EVS, using EEG. The recorded EEG data during EVS application, con-
tain the information that elicit the EVS induced responses. However, the
distribution of the EVS current throughout the scalp generates an artifact
on the EEG signals. To analyze the EEG and study the brain functions
ii
Abstract
during EVS, we have to eliminate this artifact.
We developed a method to remove this artifact by estimating the contri-
bution of the EVS current in the EEG signals at each electrode. The pro-
posed method is a hybrid method, which combines time series regression
and wavelet decomposition methods to estimate the artifact and remove
it. Wavelet transform was employed to project the recorded EEG signal
into various frequency bands and then the regression method was used to
estimate the EVS current distribution in each frequency band separately.
We optimized the proposed method using simulated data. Then we assessed
the performance of our method and compared it to the other well accepted
artifact removal methods, using both simulated and real data. The results
show that the proposed method has better performance compared to the
others, in terms of achieving higher signal to artifact ratio and introducing
less distortion to the original EEG signals.
iii
Preface
• In this project, all the experimental EVS/EEG data was collect by
our collaborators in the Pacific Parkinson’s Research Centre and the
Sensorimotor Physiology Laboratory.
• The novel method proposed in this study, was submitted as a journal
paper.
• All experimental procedures in this work were reviewed and approved
by the University of British Columbia Clinical Research Ethics Board
under Certificate Number H09-02016.
iv
Table of Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Motivation and Contribution . . . . . . . . . . . . 3
1.3 Future Research Goals and Potential Innovations . . . . . . . 5
2 EVS Brain Interaction . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3 EVS Studies Problem . . . . . . . . . . . . . . . . . . . . . . . 17
3.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . 17
v
Table of Contents
3.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4 Experimental Set-up and Data Acquisition . . . . . . . . . 23
4.1 EVS Instrumentation and Set-up . . . . . . . . . . . . . . . . 23
4.2 EEG Acquisition Set-up . . . . . . . . . . . . . . . . . . . . . 26
5 A Review of Artifact Removal Methods and Their Mathe-
matical Principles . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 ICA-based Artifact Removal Methods . . . . . . . . . . . . . 33
5.3 Regression-based Artifact Removal Methods . . . . . . . . . 36
5.4 Adaptive Filter-based Artifact Removal Methods . . . . . . . 39
5.5 Wavelet-based Artifact Removal Methods . . . . . . . . . . . 42
6 A Novel Hybrid Artifact Removal Method for EVS Arti-
facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.1 Design a Hybrid Wavelet-based Artifact Removal Method . . 46
6.2 Impedance Analysis and Simulation Study for Method Opti-
mization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.3 Optimization of The Proposed Method Using Simulated Data 58
7 Study Results and Assessment . . . . . . . . . . . . . . . . . 63
7.1 Assessment of The Artifact Removal Methods in Stochastic
EVS Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.2 Assessment of The Proposed Artifact Removal Method in
Pulsed GVS Studies . . . . . . . . . . . . . . . . . . . . . . . 79
vi
Table of Contents
8 Conclusion and Future Work . . . . . . . . . . . . . . . . . . 85
8.1 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . 85
8.2 Areas of Future Studies . . . . . . . . . . . . . . . . . . . . . 88
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Appendices
A DS5 Isolated Bipolar Constant Current Stimulator . . . . 97
B Model 2200 Analog Stimulus Isolator . . . . . . . . . . . . . 100
C Matlab CNT Converter . . . . . . . . . . . . . . . . . . . . . . 102
D Extended-Infomax ICA Algorithm . . . . . . . . . . . . . . . 122
vii
List of Tables
5.1 Frequency bands for approximation and details components
of the measured EEG when sampling rate is 1KHz . . . . . . 45
6.1 Correlation and normalized residual sum of squares between
the artifact-removed signals and the original artifact-free EEG
signals for simulated data using different wavelets . . . . . . . 60
6.2 Correlation and normalized residual sum of squares between
the artifact-removed signals and the original artifact-free EEG
signals for simulated data using different regression models . 61
7.1 Correlation between the EVS signals and the estimated EVS
artifact extracted from EEG signals for real data using dif-
ferent methods . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.2 Correlation between the EVS signal and the estimated EVS
artifact reconstructed from different frequency components
for real data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
viii
List of Figures
2.1 The inner ear in human (the left ear), including the Cochlea
and the vestibular system — Courtesy of NIH Medical Arts . 8
2.2 Vestibular nerve system and the EVS mechanism: Stimula-
tion is applied to the mastoids (the arrow) and activates the
relay stations upstream [31]. . . . . . . . . . . . . . . . . . . . 11
2.3 Fitzpatricks semicircular canal model which shows the yaw,
pitch, and roll sensations evoked by bilateral bipolar vestibu-
lar stimulations with the anode connected to the right side
of the head and the cathode connected to the left side of the
head [9]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 3-dimensional fMRI brain images during EVS application in
which the areas with significant activation were colored. From
left to right: right hemisphere view, superior view, left hemi-
sphere view — Courtesy of Elie Lobel et al. [17] . . . . . . . 18
ix
List of Figures
3.2 3-dimensional fMRI brain images during EVS application in
which the areas with significant deactivation were colored.
From left to right: anterior view, paramedian sagittal view,
inferior view — Courtesy of Elie Lobel et al. [17] . . . . . . . 19
3.3 Measured EEG data 60 seconds before, 72 seconds during
the EVS stimulation (pink noise with 100uA RMS) and 60
seconds after applying the EVS . . . . . . . . . . . . . . . . . 22
4.1 The experimental set-up for SVS studies . . . . . . . . . . . . 26
4.2 20 and 64-channel EEG electrodes placement (20-channel sys-
tem is represented by electrodes in black) — Courtesy of For-
rest S. Bao . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 Measured EEG data 60 seconds before, during the SVS stim-
ulation using 72 seconds Pink noise with 100uA RMS and 60
seconds after the SVS application . . . . . . . . . . . . . . . . 31
4.4 Measured EEG data 30 seconds before and during the GVS
stimulation using 5 square pulses with the cycle of 10 seconds
On, 30 seconds Off and amplitude of 500uA . . . . . . . . . . 31
5.1 Power spectrum of a EVS with pink noise waveform in the
frequency range of 0.1-10Hz . . . . . . . . . . . . . . . . . . . 43
6.1 Flowchart of the proposed method . . . . . . . . . . . . . . . 48
6.2 Electrical equivalent circuit for the electrode-skin interface
and the underlying skin [33] . . . . . . . . . . . . . . . . . . 50
x
List of Figures
6.3 Fit percentage between the simulation output and the mea-
sured EEG at each channel . . . . . . . . . . . . . . . . . . . 52
6.4 The transfer function of the modeled impedance at the occip-
ital channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.5 The fit percentage for the impedance model at channel 18 for
time intervals (a) 1sec and (b) 2sec . . . . . . . . . . . . . . . 56
6.6 The fit percentage for the impedance model at channel 18 for
time intervals (a) 5sec, (b) 7 sec, (c) 10sec, and (d) 14sec . . 57
artifact using the RLS filters . . . . . . . . . . . . . . . . . . 66
artifact using the RLS filters (top) and the LMS filters (bottom) 67
artifact using the LMS filters . . . . . . . . . . . . . . . . . . 68
7.4 Fit percentage of the estimated EVS artifacts with simulated
EVS contaminated signals using the OE order 2 in 12 fre-
quency bands . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.5 Correlation between the estimated EVS signals and simulated
EVS contaminated signals using the OE order 2 in 12 fre-
quency bands . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.6 The Occipital EEG channel data after applying the proposed
artifact removal method using the frequency components be-
tween 1-16Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
xi
List of Figures
artifact removal method using the frequency components be-
tween 1-32Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
artifact removal method using the frequency components lower
than 64Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.9 The Occipital EEG channel data before removing EVS artifacts 76
7.10 Power spectrum of the EEG data before, during and after
EVS application (top) and Power spectrum of the EVS (bottom) 77
7.11 Correlation between the EVS signals and the estimated EVS
artifacts using the proposed method (red) and the ICA method
(blue) for different stimulation amplitudes . . . . . . . . . . . 79
7.12 EEG signal during the GVS pulses with the amplitude of 0.5mA 82
7.13 The cleaned up EEG signal during the third GVS pulse using
the proposed method, reconstructed from all the frequency
components . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.14 The cleaned up EEG signal reconstructed from the frequency
components between 4-32Hz (top) and GVS artifact compo-
nent (bottom) . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
8.1 Schematic of a closed-loop EVS/EEG system . . . . . . . . . 90
xii
Chapter 1
Introduction
1.1 Overview
We used Electrical Vestibular Stimulation (EVS) to deliver stimulation sig-

nals to the brain through the vestibular system. This sensory system, which
is one part of the inner ear, is responsible for balance, sense of spatial orien-
tation and perception of velocity and acceleration. Manipulating vestibular
system using EVS, has been introduced over a century ago, when Alessan-
dro Volta applied electrical stimulation to his head using the battery that
he had invented, and experienced the feeling of dizziness. At the same time,
Luigi Galvani conducted many experiments to demonstrate electrical prop-
erties of the nerves. These two scientists made a lot of contributions in
the search of the effects of electrical or galvanic stimulation on the nervous
system. Subsequently, EVS has been used in many neurological and phys-
iological researches and many different applications for the EVS have been
introduced in different fields, such as clinical-therapeutic, navigation sys-
tems, virtual-reality simulations and games. For example, in the navigation
system, which was designed for blind people, the orientation of the subject
1
1.1. Overview
is controlled by EVS, while the EVS device is connected to a Global Po-

sitioning System (GPS), so that the subject is guided in a given direction.
In virtual-reality simulations or in some games, EVS is used to convey the
sense of motion and acceleration.
In spite of the fact that many applications have been developed and imple-
mented using EVS, it is yet to completely understand how the EVS stimulus
acts as an input to the brain system and how the vestibular responses are in-
terpreted by the brain. Although many scientific endeavors have been made
in the field, researchers are still striving to learn more about the vestibular
contributions to brain dynamics and whole-body control.
To study the effects of EVS on the brain functions, and also to develop
a model that explains the EVS responses, we need to monitor and record
the brain activities during the EVS application. Simultaneous acquisition
of EEG/EVS or fMRI/EVS [17] can provide us with data, which shows
how the vestibular system is connected to different regions in the brain
and how the signals sent from vestibular system affect different functions
of the brain. Collecting brain signals using either MRI or EEG during
EVS is very challenging, due to the different types of artifacts that EVS
causes in the MRI images and EEG signals. Some researchers had suggested
to use different type of vestibular stimulations, such as caloric vestibular
stimulation (CVS), to avoid the electrical artifacts. However, they realized
that the other vestibular stimulations may not provide the same effects as
EVS does. EVS can activate the entire vestibular system and stimulate a
wide range of central vestibular neurons, including the semicircular canals
2
1.2. Research Motivation and Contribution
and the otolith organs neurons, as opposed to other stimulations such as

CVS, which mainly excites the semicircular canals [17].
Since EEG is available in almost any physiology/neurology research cen-

ter, and also it has been extensively used to study the brain functions, we
used EEG simultaneously with EVS to collect the brain signals during the
stimulation. But analyzing the EEG/EVS data is very challenging, due to
different artifacts, such as myogenic, ocular and stimulation artifacts. Re-
searchers have been looking for methods to remove these artifacts to be able
to analyze the original brain signals. In this study, we introduce a new
method to remove the EVS induced artifacts from EEG recordings.
1.2 Research Motivation and Contribution
One of the main motivation of this study is to enable researchers to study

the brain functions during EVS and characterize the central nervous sys-
tem responses to EVS. This can provide an insight into how the dynamics
of the human body, specifically brain, can be altered by EVS. The impor-
tance of understanding the EVS mechanism lies in its contribution to the
development of applications, such as diagnostic or therapeutic solutions for
neurological disorders. Already it has been shown that the EVS can be
employed in such applications [23, 35, 36], but as researchers have not had
access to the brain signals during the EVS, they have not been able to adjust
the EVS signals in accordance with the specific needs for each subject.
3
1.2. Research Motivation and Contribution
However, exploring the EVS effects and studying the stimulus interactions
with the brain using EEG is very challenging, as the EVS current is dis-
tributed throughout the scalp, and generates an artifact on EEG signals.
The thesis tackles this problem to help researches in the field of neuro-
sciences and physiology analyze the EEG signals during EVS, thus utilize
EVS in a more efficient manner. The main contribution of this study is to
a novel method to eliminate the EVS induced artifacts, which have much
higher amplitude compared to that of the brain signals. This new method
can estimate the contribution of the EVS current in the EEG signals at each
electrode and remove it in a way that the least undesired changes are in-
troduced to the original brain signals. This method is less computationally
demanding and faster than most of the other methods. In this method, as
opposed to ICA-based methods, we do not need to collect EEG data from
many different channels and analyze all the data, in order to remove the
artifact. It is very useful for the applications that only a limited number of
EEG channels are to be analyzed, as the artifact can be removed without
additional data processing from any extra channels.
We design this method in a way that, it can be used in future real-time

closed-loop applications. It can be implemented in FPGA-based embedded
systems to remove the EVS-induced artifacts as the system acquires the
EEG and delivers the EVS signals, and thus provides the ability of changing
the EVS stimuli based on the acquired EEG signals.
4
1.3. Future Research Goals and Potential Innovations
1.3 Future Research Goals and Potential
Innovations
The EVS studies show that it can be potentially employed for diagnostic and
therapeutic procedures, not only for vestibular disorders but also for some
of the neurological diseases. However, implementing these applications is
very challenging due to lack of a complete understanding of the function
of the EVS stimuli and the ability to precisely trace the vestibular signals
path in different brain regions. In this project, our primary goal is to pro-
vide researchers with clean EEG signals acquired from any individual EEG
channel during EVS, independent from the other EEG channels. This novel
method can be employed in many EEG/EVS studies to remove the artifacts
induced by EVS stimuli with different waveforms, power spectra, duration
and frequency range.
After removing the EVS artifact, researchers will be able to pursue long-term
research goals such as finding the correlation between EVS and brain signals
and identifying the effects of the EVS on the brain dynamics. This can lead
to better characterization of the vestibular system and the central nervous
system responses to the EVS stimuli. Identifying a model for the vestibular
system can benefit in detecting the EVS effects on different regions of the
brain, and ultimately it can help researcher find a way to alter the brain
dynamics using EVS in future.
To optimize the EVS effects on the brain dynamics in different applications,

researchers need to find the best operational EVS parameters and waveforms
5
1.3. Future Research Goals and Potential Innovations
for each specific application. Thus it is very important that the artifact
removal methods can be run in a real time closed-loop system. In this study,
we developed and optimized the method in a way that it can be implemented
in embedded systems, with low latency and low computational cost.
A potential innovation lies in the implementation of a closed loop system for

delivery of alternative treatments to suppress or control the symptoms’ of
neurological diseases (e.g. Parkinson’s, Epilepsy, etc), or for early diagnosis
of vestibular and neuro-physiological disorders, as well as quantification of
the diseases severity.
6
Chapter 2
EVS Brain Interaction
2.1 Background
Electrical Vestibular Stimulation (EVS) is the process of delivering electric

signals to the brain through the vestibular system. This process is some-
times called Galvanic Vestibular Stimulation (GVS) as well. However, in the
literature, this term has been extensively used for a prolonged square-wave
current stimulation, while EVS refers to vestibular stimulation using elec-
trical current with any waveforms. In this report, EVS is used as a general
term for any types of electrical vestibular stimulation.
The vestibular system is one part of the inner ear and along with the other
part of the auditory system, the Cochlea, constitutes the labyrinth of the
inner ear in most mammals (Fig. 2.1). The vestibular system, has two parts:
the semi-circular canal system, which senses rotational movements, and the
Otoliths, which senses linear accelerations.
The two main groups of receptors in the vestibular system, are located in
the three semi-circular canals (to sense the rotations along the pitch, roll,
7
2.1. Background
and yaw axes), and in the two Otolith organs (the Utricle for horizontal and
the Saccule for vertical sensory information).
Figure 2.1: The inner ear in human (the left ear), including the Cochlea and
the vestibular system — Courtesy of NIH Medical Arts
The three semi-circular canals are interconnected tubes, which are approx-
imately orthogonal to each other, and aligned in different planes. These
canals are: the horizontal semicircular canal (or the lateral semicircular
canal), superior semicircular canal (or the anterior semicircular canal), and
the posterior semicircular canal. As these canals are oriented along the
pitch, roll, and yaw axes, they can provide sensory inputs for rotary move-
ments. The anterior and posterior semicircular ducts are oriented vertically
(in the sagittal plane and the frontal plane) and can detect vertical head
movements such as bending forward the head towards the ground, and the
8
2.1. Background
lateral semicircular duct, which has a 30-degree angle with the horizontal
plane can detect horizontal head movements, such as spinning the head in
a rotating chair. The movement of a fluid inside the semicircular canals
pushes on a structure which is called the Cupula. Cupula contains hair cells
that convert the mechanical movement to electrical signals. The canals in
right and left sides are oriented in a way that each canal on each side has a
parallel counterpart on the other side. These counterparts work in a push-
pull mode, which means when one canal is stimulated, the counterpart on
the other side is inhibited.
The two organs in the Otoliths, Utricle and Saccule, contain small crystals
of calcium carbonate called Otoconia. Otoconia crystals are placed on a
viscous gel layer. Since they are heavier than their surroundings, they get
displaced during linear movements and deflect the hair cells, which leads to
polarization of the cells and produces a sensory signal. Utricle and Saccule
provide both static and dynamic sensory information. The static infor-
mation is based on the orientation of the head relative to gravity, while
the dynamic information provided by the Utricle and Saccule indicates the
changes of the velocity of the person (for example, in case of accelerating or
decelerating in a car or elevator).
When a change in orientation or velocity occurs, the organs in the vestibular

system either increase or decrease their neurons (afferent neurons) firing
rates, and the brain interprets these variation in the firing rates into sensory
information. The brain uses this sensory information to properly react to
changes in orientation and velocity through ocular and muscle systems.
9
2.1. Background
The vestibular system signals are projected to the Cerebellum and to differ-
ent areas in the Cortex. The projections to the Cerebellum are relayed back
as eye movements, head muscle movements and postures. It also affect the
body motor control through the Thalamus and balance through the Spinal
Cord. The projections to the Cortex, spread out over different areas, have
different effects that are not completely understood yet.
Electrical stimuli can affect a part of primary afferents of the Otolith and
the semicircular canals in the vestibular system. Afferent neurons, called
as well receptor neurons, carry signals from receptors (the Otolith and the
semicircular canals) towards the central nervous system. EVS stimulation
changes the afferent neurons discharge patterns through polarization effects.
For example, EVS can increase the firing rate of the responsive afferents of
the semicircular canals in the cathode side [9, 31].
The vestibular system encodes the EVS signals (analogous to frequency

modulation) and sends them to the brain. The vestibular nerve, behind the
mastoids, goes from the inner ear to vestibular brain stem nuclei and thala-
mic nuclei. The vestibular fiber pathways reach different cortical vestibular
areas including the somato-sensory cortex, parietal area, and the Parieto-
insular vestibular-cortex(PIVC) (Fig. 2.2). Functional imaging studies of
EVS show that EVS can activate different regions of brain such as different
areas in the Cortex, the Cerebellum, and the Thalamus [31].
10
2.1. Background
Figure 2.2: Vestibular nerve system and the EVS mechanism: Stimulation is
applied to the mastoids (the arrow) and activates the relay stations upstream
[31].
11
2.1. Background
There are three major responses to the EVS stimulation applied to the mas-
toids in human subjects: ocular, postural, and perceptual responses. As
previous mentioned, different studies showed that during EVS, vestibular
afferents from both otoliths and semicircular canals can be stimulated. A
canal signal may cause a rotation of head and an otolith signal may cause a
tilt or induced a translational acceleration of the body. In healthy subjects,
a GVS stimulus (square pulse EVS) over 0.5 mA produces a roll of the trunk
and head and GVS stimuli around 1mA (applied for few seconds), produce
perceptions of body rotation and may cause the subjects to sway if they are
standing or perceive illusory movements if they are not standing. GVS stim-
uli at 2 mA produce ocular torsion [9, 31]. Some of these responses can be
modulated by the EVS frequency. For instance, low frequency components
of the EVS current seems to be responsible for the body sway, thus when the
frequency components lower than 2Hz are removed from a vestibular stimuli
with a 0-25 Hz bandwidth, the trunk sway responses to the EVS stimulus
are drastically suppressed [5]. These responses are the major behavioral
EVS effects that are usually observed in healthy subjects. Fitzpatrick et al.
illustrated a model of semicircular canals’ responses to EVS stimulations in
Fig. 2.3.
EVS has some effects on cognitive functions as well. The connections be-
tween vestibular brainstem nuclei and other regions in brain (e.g. medial
temporal lobe, and lateral temporal–inferior parietal lobe) are thought to
be the reason of these effects. As an example, applying EVS in healthy sub-
ject can enhance the visual memory (e.g. increase the speed of the visual
12
2.1. Background
Figure 2.3: Fitzpatricks semicircular canal model which shows the yaw,
pitch, and roll sensations evoked by bilateral bipolar vestibular stimulations
with the anode connected to the right side of the head and the cathode
connected to the left side of the head [9].
13
2.2. Applications
memory recall) and increase the response time of the mental transformation
tasks [34]. In patient with parietal lobe lesions, EVS can improve a variety
of multi-modal spatial cognition tasks. For example, in patient with visual
neglect resulting from parietal lesions, applying EVS can cause corrections
of the spatial reference frame distortion [31]. Vesibular system interactions
with sensory-motor and cognitive mechanisms, such as spatial attention,
spatial cognition, non-spatial attention (neglect), body cognition and visual
memory [5, 9, 31, 34], have been studied using EVS.
In some research studies, the effects of EVS on the basal ganglia and the
limbic system have been investigated. The vestibular nerves send signals to
the cerebellar vermis which has influences on the turnover of dopamine and
noradrenaline in the basal ganglia. This indicates the possible vestibular
effects on the basal ganglia and the limbic system. It has been shown that
EVS can be employed to suppress the symptoms of the degenerative neuronal
diseases such as multiple system atrophy and Parkinson’s disease [23, 35, 36].
2.2 Applications
Many research institutions around the world have been studying EVS from
different aspects and developing a variety of applications. Since EVS can
provide a cost effective and safe way to stimulate the human body bal-
ance, motion and acceleration sensors, it can be used to create the illusion
of motion and spatial disorientation, induce body movements and possibly
improve the motor performance.
14
2.2. Applications
EVS has been used in flight simulation systems to train pilots through a vir-
tual environment. One of the main problems in flights is that there are situ-
ations, where the vestibular sensors cannot provide reliable signals. Hence,
the sensory signals are interpreted incorrectly, leading to spatial disorienta-
tion (false perception related to the position in three dimensions). These
sensory illusions occur because of the unnatural environment in the flights.
For example, when a pilot makes a banking left turn, the vestibular system
will detect the roll into the turn. As the turn is continued, the vestibular
system will compensate and does not react to the constant turn. Conse-
quently the brain perceive a false return to level flight. To train pilots and
familiarize them with these situations, EVS can be employed to provide
these sensory illusions in virtual reality simulation systems.
Not only can EVS be useful in the flight trainings, but also it could be
applicable for pre-flight tests, to check the pilots vestibular system prior to
aircraft operation. Recently, EVS has been also used to simulate post-flight
sensorimotor effects, since it can safely induce the symptoms of post-flight
sensorimotor disruptions, such as postural, gaze, locomotor and perceptual
deficits. These symptoms are usually observed in pilots and astronauts
following long duration flights [7].
In gaming industry, EVS could be used to enhance the gamers’ experience

through a virtual reality environment. For example, in car-racing games,
EVS can be used to simulate the actual feeling of the acceleration and
movements in driving such as gravity shift when the car goes through a
sharp turn. Creating the illusion of motion and acceleration along with the
15
2.2. Applications
audio-visual components can make computer games feel more realistic.
Furthermore, as EVS can cause or induce body movements, it can be used to

remotely control a person. Researchers have utilized this EVS property for
guiding and steering a person. Such a system has been developed in the NTT
Atsugi Research and Development Center. This system can be advantageous
for blind people, when it is connected to a GPS and induce walking or
turning in accordance with the position data. Also, it could be applicable
in military missions, when the personnels must be silently steered through
the battlefield and hazardous locations, using the information provided by
different types of detectors.
In the biomedical field, EVS can be employed to help patients with an im-
paired sense of balance. There have been several trials to introduce novel
treatments using EVS for various neuro-psychological disorders such as mul-
tiple system atrophy (MSA) and Parkinson’s disease (PD) [22, 23, 35, 36].
Studies of the EVS evoked responses in patients with symptoms of neuro-

physiological or vestibular disorders, show promising outcomes from the ap-
plication of EVS in the field of neuro-psychology, which can lead to both
non-drug treatments for affected patients, and ideal diagnostic tools to ex-
amine the state of the diseases.
16
Chapter 3
EVS Studies Problem
3.1 Problem Statement
One of the main problems that researchers have faced in advancing the EVS
applications and exploring new applications, lies in the fact that it is not pre-
cisely understood how the stimulus interacts with the human body system.
To tackle this problem, researchers have used devices, such as functional
Magnetic Resonance Imaging (fMRI) to study the brain activities during
EVS application. For example, some areas of the cerebral cortex involved
in the perceptual and motor functions were studied during the stimulation
of the vestibular organs, using fMRI [17]. The MRI images acquired in the
study conducted by Elie Lobel et al. show the brain areas that are activated
and inactivated during EVS, where the stimulus intensity was set at 0.5 mA
less than the subject pain threshold, ranged from 3.0 to 4.2 mA (Fig. 3.1
and Fig. 3.2).
MRI images are precise and provide very useful information about brain
activities. However, MRI scanners are very expensive and in many research
17
3.1. Problem Statement
centers, they are not available or there is very limited availability to the
MRI scanner for non-urgent studies and research purposes. On the other
hand, the condition of the subject inside the MRI tunnel, e.g. lying still in
an enclosed space and surrounded by loud MRI noises, may affect the results
of the EVS studies. Also, some of the EVS studies cannot be conducted in
MRI due to the induced body movements during the application of EVS,
which affect the MRI images.
More importantly, we can not use MRI to monitor the brain activities of the
subject in the real applications, while the subjects are involved with their
other daily activities.
Figure 3.1: 3-dimensional fMRI brain images during EVS application in

which the areas with significant activation were colored. From left to right:
right hemisphere view, superior view, left hemisphere view — Courtesy of
Elie Lobel et al. [17]
18
3.1. Problem Statement
Figure 3.2: 3-dimensional fMRI brain images during EVS application in

which the areas with significant deactivation were colored. From left to
right: anterior view, paramedian sagittal view, inferior view — Courtesy of
Elie Lobel et al. [17]
Another research tool extensively used for the brain studies is EEG. In
EEG recordings, the brain’s spontaneous electrical activities are recorded
from multiple electrodes, placed on the scalp. EEG has been used in dif-
ferent diagnostic procedures, such as diagnosis of epilepsy. In most of the
diagnostic procedures, the general focus is on the spectral content of EEG,
which illustrates the type of oscillations in neural activities.
The brain’s EEG signals are formed by the electrical activities of billions of
neurons. The neurons which are electrically charged to propagate potential
signals create a wave of ions. The ions’ wave reaches the EEG electrodes
on the scalp and generates a potential difference between the scalp and the
EEG electrodes, recorded over time. The EEG recordings, therefore, are
the reflection of the summation of the activity of millions of synchronized
19
3.2. Limitations
neurons with similar spatial orientation. This makes the EEG recordings
complicated to interpret into the brain activities. Especially the electrical
activity from deep sources in the brain is more difficult to detect, not only
because their signals are mixed with currents from different regions, but also
the voltage falls off when the current passes through the skull.
In spite of the complexity of analyzing the EEG signals due to its poor signal
to noise ratio and low spatial resolution , EEG is still an essential tool in
studying the brain activities, because it can provide very useful data with
high temporal resolution close to 1 ms, which can reflect neural activities
almost in a real-time manner. In addition, EEG studies can be performed
in most of conditions at very low cost and it allows some degree of flexibility
in different experiments. Therefore, in this project, we used EEG during
application of EVS to monitor the brain activities.
3.2 Limitations
To obtain the information on how the EVS can influence neural activities, it
is essential for us to acquiring clean EEG signals during the EVS stimulation.
One of the limitations in regard to the EEG/EVS studies is the occurrence
of different artifacts in the EEG signals. Especially, the ability to study
the ongoing neural activities during the EVS stimulation is very limited
because of the artifact which appears in the EEG recordings due to the
EVS stimulation.
20
3.2. Limitations
The EVS artifact appears because of the leakage of the EVS current mainly
through the scalp. The EVS current flowing through the scalp generates an
additional potential, with much higher amplitude than that of the neural
activities, especially in the area close to the stimulation site. This additional
potential, superimposed on the EEG signals, is observed as a large artifact.
The EVS stimulation artifact is the major difficulty in understanding the
physiological mechanism of the EVS and monitoring the EVS effects on
different regions of the brain and the neural circuitries.
Hence, removing the EVS artifact is the first step in the EEG/EVS studies.
Dealing with the EVS artifact is very challenging because usually the ampli-
tude of the EVS artifacts is much higher than that of the EEG signals, even
when the EVS signals are very small (e.g. 100uA). As a result, in EEG/EVS
studies, we have very poor signal to artifact ratio. As an example, an ex-
perimentally measured EEG signal contaminated with the EVS artifacts is
illustrated in Fig. 3.3. Although the stochastic EVS applied in this example
is a small pink noise with 100uA RMS, the signal to artifact ratio (SAR; the
RMS value of the true EEG signal to that of the EVS artifact in decibels)
is -15.171dB .
Considering that the frequency spectra of the neural signals and EVS arti-
facts overlap very often, filtering the frequency components of EVS artifacts
results in loss of the original neural signals. The major focus on the EEG
frequency bands is on delta, theta, alpha and beta bands. Delta is the fre-
quency band up to 4 Hz, theta is the frequency band between 4 Hz to 8 Hz,
alpha is the frequency band between 8 to 12Hz and beta is the frequency
21
3.2. Limitations
2500
2000
1500
EEG voltage (uV)

1000
500
−500
−1000
−1500
−2000
−2500
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

5
Time (msec) x 10
Figure 3.3: Measured EEG data 60 seconds before, 72 seconds during the
EVS stimulation (pink noise with 100uA RMS) and 60 seconds after applying
the EVS
band between 12 to 30 Hz. To remove the EVS artifact, the artifact removal
method must be able to remove the artifacts from the frequency band of
interest in such a way that the original neural signals remain intact and the
physiological effects of EVS on EEG signals can be analyzed.
Another limitation in EEG/EVS studies and implementing applications such

as diagnosis or control the symptoms of physiological and neurological dis-
orders using EVS, is that we need to adjust the stimulation signals depends
on the brain responses during the EVS. To be able to deliver the EVS stim-
ulation in a closed loop mode, and modify the EVS signals adaptively in
response to the neural activities, it is necessary to remove the EVS artifacts
from neural activities signals in a real-time implementation. So the suitable
artifact removal methods in these applications are based on algorithms with
lower computational complexity and faster running time.
22
Chapter 4
Experimental Set-up and

Data Acquisition
4.1 EVS Instrumentation and Set-up
In the process of EVS, an electrical signal generator is used to deliver elec-

trical stimulation to the vestibular system. The electrical current stimulus is
generated by a current source and delivered to the vestibular system through
the electrodes placed over the skin behind ears. Nowadays the EVS elec-
trodes are placed on the mastoid bones as apposed to the former set-up
where the electrodes were placed in the ears.
There are different EVS electrodes configurations. The most common con-
figuration is bilateral bipolar, where the anodal and cathodal electrodes are
placed behind each ear. There are other configurations as well, such as bi-
lateral mono-polar or unilateral mono-polar. In bilateral mono-polar EVS,
the electrodes with same polarity are placed behind each ear, with a distant
reference electrode. In unilateral configurations, the stimulation is delivered
23
4.1. EVS Instrumentation and Set-up
to one ear. In each of these configurations, the physiological mechanism of

the EVS is different. In our studies, we delivered the stimuli to the mastoid
bones in a bipolar-bilateral configuration. We used carbon rubber electrodes
(9cm2 ), and attached them by sticky electrode gel over the mastoid bones
and secured them with extra tape and elastic headbands.
For EVS experiments, in a basic set-up, a battery over 6 V is enough to

drive the EVS current through the head and stimulate the vestibular sys-
tem. However, in most of EVS studies, stimulation is usually delivered by
means of isolated constant current supplies at the levels ranging from a few
hundreds of micro-ampere to a few milli-ampere.
In our experimental set-up, we used isolated bipolar constant current stimu-

lators, DS5-Digitimer (Digitimer Ltd, UK) and Model 2200 Analog Stimulus
Isolator (A-M Systems, WA, U.S.A.). For more information about these de-
vices, please refer to Appendix A and B.
In both of these stimulators, the waveform of the stimulation current is

selectable and can be proportional to the input voltage. The command
input voltage is fed into the stimulator input and the stimulator follows the
shape of the command voltage waveform. As long as the output does not
exceed any of the limits that the stimulator imposed for safety matters, the
stimulation waveform is precisely the same as the input voltage waveform.
We created the desirable stimulation waveforms on a PC computer, using

the LabVIEW software (National Instruments, TX). The software-generated
waveform is then sent to the stimulator through a Data Acquisition (DAQ)
24
4.1. EVS Instrumentation and Set-up
board (USB-6221, National Instruments, TX), shown in Fig. 4.1.
We conducted different studies using different types of EVS, such as stochas-

tic and galvanic stimulations. In the Stochastic Vestibular Stimulation
(SVS) studies, we applied zero-mean pink noise currents, with a 1/f-type
power spectrum, within a frequency range of 0.1 to 10Hz and duration of
72 seconds. We kept the amplitude of the delivered stimuli lower than the
feeling threshold, in the range of 100uA to 800uA, with root mean square
(RMS) values between 60uA to 450uA.
In the Galvanic Vestibular Stimulation (GVS) experiments, we applied single-

polarity and dual-polarity square pulses (e.g. 10 pulses) with longer duration
of 10s ON, 30s OFF and shorter duration of 500ms ON and 10s OFF. The
amplitude levels of GVS pulses were set between 0.5mA to 3mA.
In all studies, the delivered stimulation signals are recorded as the reference
signals, which allows us to identify the EVS artifacts (this will be discussed
in the next chapter). To record the stimulation signals, we used the same
DAQ board that we used to send the control signals. Fig. 4.1 shows the
experimental set-up for our SVS studies, where we used the DS5 stimulator,
connected to a PC computer via a DAQ board. This set-up enables us to
deliver the desirable stimuli (generated by the PC computer) and simulta-
neously record the output of the stimulator, and save it on the computer.
25
4.2. EEG Acquisition Set-up

Figure 4.1: The experimental set-up for SVS studies
4.2 EEG Acquisition Set-up
We monitored the brain activities and recorded them using an Electroen-

cephalography (EEG) system. In EEG, the electrical signals produced by
brain neurons are recorded from multiple electrodes, placed on the scalp.
When a large array of electrodes are needed, usually the electrodes are em-
bedded into a cap or a net, which can contain up to 256 electrodes around the
scalp. Each electrode, along with one common reference electrode are con-
nected to a differential amplifier (one amplifier per pair of electrodes). The
voltage between electrodes and the reference electrode are generally ampli-
fied 1,000 to 100,000 times (or 60 –100 dB). The amplified signals are passed
through an anti-aliasing filter, converted to digital data, and recorded using
a computer. The sampling frequency is usually 256Hz or 512 Hz in clinical
26
applications and for research studies, it is usually over 512 Hz (sometimes

the sampling rate up to 20 kHz is used).
Before collecting EEG/EVS data, we need to look into the electrical prop-
erties of the body, specifically the scalp impedance. The most important
reason for the impedance check, is the quality of the EEG and the EVS
stimulation. If the impedance at the electrode-skin interface is not opti-
mized, neither can we acquire high quality EEG signals nor deliver effective
EVS stimulations.
The skin impedance measurements can be affected by different factors, such

as the amount of dirt on the surface of the skin and dead cells, which can
act as a protective layer against electrical shocks and the subsequent stim-
ulations. Because of these factors, which make the impedance of the body
highly variable, before the experiment, we need to wash the head, clean the
mastoid by alcohol prep pads, and add the low impedance highly conductive
gel between electrodes and the skin.
The electrical properties of the skin influence the EVS current threshold,
(the EVS current which is large enough to change the firing pattern of the
nerves at the stimulation site), the pain threshold and muscle stimulation.
The muscle stimulation may occur indirectly from the EVS stimulation.
Muscle contractions can be triggered at a current level which varies with the
electrical properties of the body. In the EEG/EVS experiments, twitches of
ear and skin during the EVS stimulation, is not desired since it adds some
more artifacts to the EEG data. In some of our studies, we tried to keep
27
the EVS current lower than the muscle stimulation threshold, to reduce or
eliminate the muscle contraction.
In our SVS studies, we carried out the EEG recording with a Neuroscan
Synamps2 system (NeuroScan, NC, U.S.A.). This EEG device uses a low
noise 24 bit Analog-to-Digital converter which allows a broad dynamic range.
During our EVS studies no saturation occurred. We acquired the EEG
signals from 20 electrodes located on the scalp according to the international
10-20 EEG system (Fig. 4.2). We set the EEG sampling frequency to 1 kHz
and recorded the EEG signals before, during and after the stimulation, (e.g.
60 seconds before and 60 seconds after the stimulation and 72 seconds during
the stimulation). The EEG data were stored in files with cnt format, which
is a standard format containing binary data, used by many EEG system
manufacturers. To load these data in the MATLAB software, we used a
software converter, presented in Appendix C.
All the EEG data for the SVS experiments, were collected by our collabo-
rator in the Pacific Parkinson’s Research Centre. In these experiments, we
applied vestibular stimulation to nine healthy subjects (6 male, 3 female)
between the ages of 21 and 53 yr, with no known history of neurological
disease or injury.
In our GVS studies, we used the ASALab system (ANT-Neuro) to acquire

EEG signals from 64 electrodes, located on the scalp according to the inter-
national 10-10 EEG system (Fig. 4.2).
28
Figure 4.2: 20 and 64-channel EEG electrodes placement (20-channel system

is represented by electrodes in black) — Courtesy of Forrest S. Bao
In these GVS experiments, the EEG sampling frequency was set to 4 kHz
29
and we recorded the EEG signals during the stimulation, before and after
the stimulation (e.g. 30 seconds before and 30 seconds after the stimulation).
This ASALab EEG device uses a low noise 24 bit Analog-to-Digital converter
which allows an acquisition of 29.8 nV per bit and it provides a very broad
dynamic range, so that no saturation occured during our EEG/GVS data
acquisition.
All the EEG data for the GVS experiments, were collected by our collabo-
rator in the Sensorimotor Physiology Laboratory. In these experiments, we
applied vestibular stimulation to four healthy subjects (3 males, 1 female)
between the ages of 22 and 31, with no known history of neurological disease
or injury.
In all of our EVS studies, subjects were asked to relax, remain still and
concentrate on a focal point on the screen in front of them, so that myogenic
and ocular artifacts occur as least as possible. Under resting conditions,
there are also less variations in the head impedance [29], which is important
for data acquisition and dealing with artifacts in EEG/EVS studies.
Fig. 4.3 and Fig. 4.4 present the acquired EEG during the SVS and GVS
studies.
30

Figure 4.3: Measured EEG data 60 seconds before, during the SVS stimu-
lation using 72 seconds Pink noise with 100uA RMS and 60 seconds after
the SVS application

"

!

! !!

Figure 4.4: Measured EEG data 30 seconds before and during the GVS
stimulation using 5 square pulses with the cycle of 10 seconds On, 30 seconds
Off and amplitude of 500uA
31
Chapter 5
A Review of Artifact
Removal Methods and Their
Mathematical Principles
5.1 Background
There are various methods to remove different types of artifacts: myogenic

artifacts [4, 19, 20, 27], ocular artifacts [10, 13, 18, 25, 26, 32], or extrin-
sic artifacts, such as MRI induced artifacts during simultaneous EEG/fMRI
studies [11], and stimulations artifacts, such as Transcranial Magnetic Stim-
ulation (TMS) and Deep Brain Stimulation (DBS) [1, 8, 12, 21].
The major classes of artifact removal methods are component-based and

regression-based methods. The most commonly used component-based meth-
ods to remove artifacts from EEG signals are the Independent Component
Analysis (ICA) and Principal Component Analysis (PCA) methods. Gener-
ally, in the component-based methods, the EEG signals are decomposed into
32
5.2. ICA-based Artifact Removal Methods
statistically independent or uncorrelated components. The artifact compo-

nents are then identified and filtered out, thus the EEG signals are recon-
structed without the artifact components, from the remaining terms. How-
ever, applying ICA/PCA to remove EVS stimulation artifacts from EEG
signals is challenging, particularly when the amplitude of the EVS artifacts
is much higher than that of the EEG signals and we have a poor signal to
artifact ratio.
Another class of the artifact removal methods are regression based methods.
In these methods, in order to identify the artifacts, the propagation factors or
transmission coefficients must be calculated. Then the estimated proportion
of the EVS is subtracted from the EEG. There are also other artifact removal
methods, such as wavelet denoising methods and adaptive filtering methods.
In the following, some of the most common artifact removal methods are
discussed, and we explained how we applied them to clean the EEG signals,
contaminated by the EVS artifacts.
5.2 ICA-based Artifact Removal Methods
Among various component-based methods, we employed the Independent

Component Analysis (ICA) to remove the EVS artifacts. In some studies
it has been showed that the extended Infomax ICA method has better per-
formance in removing different types of artifacts than the other component-
based methods, such as PCA or other ICA algorithms such as Jade, Second
33
Order Blind Identification (SOBI), etc. [4, 11, 14, 18].
ICA is a statistical method to extract independent components from a set

of measured signals. This method is a special case of the Blind Source Sep-
aration methods where the K channel of the recorded EEG signals (E(t) =
e1 (t), ..., eK (t)) are assumed to be a linear combination of N (N ≤ K) un-
known statistically independent sources (S(t) = s1 (t), ..., sN (t)) :
E(t) = M S(t) (5.1)
where M is the unknown mixing matrix defining weights for each source
contributions to the EEG signals recorded at each channel. In ICA, the
measured K channel EEG signals are taken into a N dimensional space and
projected onto a coordinate frame, where the data projections are minimally
overlapped and maximally independent of each other. There are different
approaches to find the independent components. The main ICA approaches
are minimizing the mutual information and maximizing the non-Gaussianity
(using kurtosis or joint entropy) among the data projections.
We use two of the most popular ICA algorithms in this study: Fast-ICA
and extended Infomax ICA algorithms. Please refer to Appendix D and
Appendix ?? for the MATLAB codes, we used for the extended Infomax
ICA and Fast-ICA algorithms.
Fast-ICA is an efficient and popular algorithm for the independent compo-

nent analysis. This algorithm maximizes non-Gaussianity as a measure of
34
statistical independence, based on an iteration method such as fixed-point

or approximative Newton method. The other ICA algorithm that we used,
is the extended Infomax algorithm. This algorithm is a modified version
of the Infomax algorithm proposed by Bell and Sejnowski. Infomax is a
learning algorithm using a neural network to maximize the joint entropy
of the output of a neural processor which leads to a minimization of the
mutual information among the output components [2]. The extended In-
fomax algorithm uses a learning rule that switches between different types
of distributions, such as Subgaussian and Supergaussian sources [15]. The
extended Infomax algorithm is implemented in EEGLAB MatLab toolbox
[6] and it is widely used in EEG studies. We observed that the extended
Infomax ICA has better performance than Fast-ICA in dealing with EVS
artifacts.
We applied ICA to the measured EEG sets to identify the EVS artifacts
components. To remove the EVS artifact, we find all components that
are attributed to the EVS reference signals using the correlation coefficient
between the ICA components and the EVS signals. After identifying the
artifact components, we applied two approaches to remove the artifact.
In the first approach, we set the components that account for artifacts to
zero, Sartf (t) = 0, and obtained a new source matrix Ŝ(t) in which the
artifact sources were filtered out.
In the second approach, we applied a threshold on the artifact components

to extract the artifact spikes and set them to zero. The threshold was
35
5.3. Regression-based Artifact Removal Methods
set at four standard deviations above the mean of the normal EEG signal
without the artifacts (e.g. the signals before applying the EVS) and all
data points with amplitude over the threshold were set to zero. Thus a new
source matrix Ŝ(t) was obtained with the modified components. We set the
threshold at a level, chosen in order to keep a part of the original neural
activities which are not contaminated by EVS.
Finally, we reconstruct ICA-corrected EEG signals as:
Ê(t) = M Ŝ(t) (5.2)
Where Ê(t) is the new data set which represents the estimated artifact-
removed data. In the chapter “Study Results and Assessment”, we compared
the ICA-based artifact removal methods with the other methods, that we
introduced here.
5.3 Regression-based Artifact Removal Methods
The injected EVS current signals and the EEG signals are recorded concur-
rently, while the EVS current distribution through the scalp contaminates
the recorded EEG signals. We used the recorded EVS current to estimate
the EVS artifacts in the measured EEG signals using regression analysis. In
this method, the dependency of the EVS artifacts and the EVS reference
signals is modeled and as a result we can estimate the EVS artifacts in the
contaminated EEG signals. The recorded EVS current is taken as the input
36
reference signal and the contaminated EEG data is modeled as a mixture of

true EEG and EVS artifact components.
We applied time-series regression methods using different model structures,

to optimize the estimation of the EVS artifacts. The model structures that
we used to optimize the estimation of the EVS artifacts, are linear discrete-
time polynomial model, nonlinear Hammerstein-Wiener model and State-
Space model.
The regression-based methods use different iterative algorithms to minimize

the estimation error. We applied the weighted least squares and the maxi-
mum likelihood algorithms. In the iterative weighted least square estimation
algorithm, the cost function is the weighted sum of the squares of the model
residuals and can be obtained from the trace of the matrix of E T ∗ E ∗ W ,
where E represents the estimation error matrix and W is the weighting ma-
trix. When W is an identity matrix, minimizing the trace of the matrix of
E T ∗ E ∗ W is equal to the traditional least-sum-of-squared-errors algorithm.
In the frequency weighting method, the weighting factors are determined in
a way that they favor the frequency range, where the input spectrum has the
most power. This means we have higher weight factors at specific frequen-
cies where the input spectrum has higher power. If the measurement noise
has a normal distribution, and the inverse of the estimated noise variance is
used as the weighting function W , then minimizing the determinant of the
matrix of E T ∗ E ∗ W , leads to the maximum likelihood estimates [16].
On the other hand, different model structures can be used in regression anal-
37
ysis. One class of model structures is the discrete-time polynomial models

described by the following general equation (5.3):
B(q) C(q)
A(q)y(t) = u(t) + e(t) (5.3)
F (q) D(q)
Here, u(t) is the reference input (the recorded GVS current), y(t) is the
output (the measured EEG) and e(t) is a white noise (mean = 0, variance
= σ 2 ) which represents the stochastic part of the model. A(q), B(q), C(q),
D(q) and F(q) are polynomials in terms of the time-shift operator q which
describe the influence of the GVS current and measurement noise on the
EEG data. Model structure such as ARMAX, Box-Jenkins and Output-
Error (OE) are the subsets of the above general polynomial equation. For
example, in ARMAX model F(q) and D(q) are equal to 1, in Box-Jenkins
A(q) is equal to 1 and in Output-Error model A(q), C(q) and D(q) are equal
to 1.
Another class of model structures is Hammerstein-Wiener models. These

models use one or two static nonlinear blocks in series with a linear block,
which enable them to capture some of the nonlinear behavior of the system.
The linear block is a discrete transfer function representing the dynamic
component of the model and parameterized using an Output-Error model.
The nonlinear block can be a nonlinear function such as dead-zone, satura-
tion or piecewise linear functions. As we did not observed any dead-zone
or saturation type of nonlinearity in our data, we chose the piecewise-linear
function, which breaks down a system into a number of linear systems.
38
5.4. Adaptive Filter-based Artifact Removal Methods
We also used State-Space models, where the relation between the EVS sig-
nals, the EVS artifacts and noise are described by a system of first order
differential equations. The state equation relates the state variables, noise
and the EVS signals to the first derivatives of the state variables. The
Output equation relates the state variables and the EVS signal to the EVS
artifact.
Among the regression methods using different model structures, we obtain

the best artifact removal results, when the weighted maximum likelihood
algorithm with frequency weighting factor was employed in the estimation
method. In the section “Optimization of the proposed method using sim-
ulated data”, we compared the different model structures with different
orders.
5.4 Adaptive Filter-based Artifact Removal
Methods
Adaptive filtering is another approach to remove artifacts. This method

is specifically suitable for real time applications. In adaptive filters, the
received input data points are used to refine the properties of the filter and
match the changing parameters at every time instant. Adaptive filters have
been employed to remove different types of EEG artifacts [3].
In our application, the primary input of the adaptive filter system is the
measured contaminated EEG signal, Em (n), which is taken as a mixture
39
of a true EEG, Et (n), and an artifact component z(n). The adaptive filter
block takes the GVS current, iGV S (n), as the reference input, and estimates
the artifact component. The filter coefficients, hm , are adjusted recursively
in an optimization algorithm driven by an error signal 5.4:
e(n) = Em (n) − ÊEV S (n) = Et (n) − [z(n) − ÊEV S (n)] (5.4)
where:

M
ÊEV S (n) = hm .iEV S (n + 1 − m) (5.5)
m=1
Because of the function of vestibular system, the stimulation signals are

modulated and then sent to the brain [9]. As a result, there is no direct
correlation between the true EEG, E(n), and the EVS signal. On the other
hand, there is a strong correlation between the EVS artifact, z(n), and the
EVS signal. Therefore, we can calculate the expected value of e2 as following:
E[e2 (n)] = E[(Em (n) − ÊEV S (n))2 ] (5.6)
= E[Et 2 (n)] − E[(z(n) − ÊEV S (n))2 ] (5.7)
As the adjustment of the filter coefficients does not affect the E[Et 2 (n)], min-
imizing the term E[(z(n) − ÊEV S (n))2 ] is equivalent to minimizing E[e2 (n)].
Among the various optimization techniques, we chose the Recursive Least-

Squares (RLS), and the Least Mean Squares (LMS) for our application.
40
In the LMS algorithm, (also known as gradient adaptation algorithm), the

filter coefficients are calculated by :
hm (n + 1) = hm (n) + μ(−Δm ) (5.8)
where, μ is the adaptation gain and Δm is the gradient defined by :
∂E[e2 (n)]
Δm = (5.9)
∂hm (n)
The equation 5.8 can be written as :
∂e(n)
hm (n+1) = hm (n)−2μE[e(n) ] = hm (n)−2μE[e(n)iEV S (n)] (5.10)
∂hm (n)
We also employed RLS, where the algorithm recursively finds the filter co-
efficients that minimize the following weighted least squares cost function:

n
ε= λn−i e2 (i) (5.11)
i=M
where, λ is the forgetting factor and M is the length, or the number of

coefficients of the filter.
In the chapter “Study Results and Assessment”, we compared the results of

adaptive filters with that of the other methods.
41
5.5. Wavelet-based Artifact Removal Methods
5.5 Wavelet-based Artifact Removal Methods
We used wavelet denoising algorithms to remove the EVS artifacts. How-

ever as both artifacts and original EEG signals are in the same frequency
range, wavelet-based methods are not very efficient and a large portion of
the original EEG may be removed along with the artifact. For example,
when a EVS current with a pink noise waveform in the frequency range of
0.1-10Hz is delivered and the EEG/EVS data are acquired from scalp at a
sampling rate of 1000Hz, artifacts and EEG signals have the overlapping fre-
quency components, which cannot be separated by wavelet analysis. Using
the Welch’s method, we calculated the power spectrum of a EVS current
with a pink noise waveform in the frequency range of 0.1-10Hz, and it is
shown in Fig.5.1. In this spectrum, there are also two smaller peaks at
120Hz and 420Hz, which are related to electrical interference of the other
devices around (such as fluorescent lamps, dimmers, etc.), which generate a
mixture of harmonics at multiples of 60Hz. These peaks are usually much
smaller (more than 10dB lower than the main peak) and can be eliminated
easily. As shown in Fig.5.1, the main EVS frequency components overlap
with the EEG frequency bands: delta (up to 4 Hz), theta (4 Hz to 8 Hz)
and alpha (8Hz to 12Hz).
Another application of the wavelet transform is signal decomposition. Using

wavelet transform, we can decompose the measured signals into various fre-
quency bands. Wavelet transform provides a high resolution time-frequency
representation of non-stationary signals such as EEG signals.
42
Figure 5.1: Power spectrum of a EVS with pink noise waveform in the
frequency range of 0.1-10Hz
In wavelet transform, the signal is decomposed onto a set of basis functions,

obtained by dilations and shifts from a unique function, called the mother
or the prototype wavelet. Therefore, we can employ different basis functions
in the wavelet transform, as opposed to the Fourier Transform, where the
basis function is a sine wave.
For the discrete signals, the discrete wavelet transform (DWT) algorithm
can be applied, where the set of basis functions are defined on a dyadic grid
in the time-scale plane as:
43
ψj,k (t) = 2−j/2 ψ(2−j t − k) (5.12)
where, 2j governs the amount of scaling and k2j governs the amount of
translation or time shifting. The wavelet transform is the inner-product
of the basis wavelet functions and the signal in the time domain. In the
DWT algorithm, the discrete time-domain signal is decomposed into high
frequency or details components and low frequency or approximation compo-
nents through successive high pass and low pass filters. For multi-resolution
analysis, the original signal is decomposed into an approximation and de-
tails parts and the approximation part is recursively decomposed again by
iterating this process. Thus, one signal can be decomposed into many com-
ponents.
However, as the basic DWT algorithm does not preserve translation invari-
ance, a translation of wavelet coefficients does not necessarily correspond
to the same translation of the original signal. This non-stationary property
originates from the down-sampling operations in the pyramidal algorithm.
The algorithm can be modified by inserting 2j − 1 zeros between filters co-
efficients of the layer j, instead of down-sampling. This modified version
of the DWT algorithm is called stationary wavelet transform (SWT) and it
can preserve the translation invariance property.
We applied both DWT and SWT, to decompose the EEG signals using
different mother wavelets, such as Symlet and Daubechies of different orders.
Both the EVS current and the simulated EEG signals were decomposed into
12 levels, as a result, for the signals acquired at the sampling rate of 1000Hz,
44
the frequency bands for approximation and detail components are tabulated
in Table 5.1.
Table 5.1: Frequency bands for approximation and details components of

the measured EEG when sampling rate is 1KHz
L1 L2 L3 L4 L5 L6
Approximation 0-250 0-125 0-62.5 0-31.25 0-15.75 0-7.87
Details 250-500 125-250 62.5-125 31.25-62.5 15.75-31.25 7.87-15.75
L7 L8 L9 L10 L11 L12
Approximation 0-3.93 0-1.96 0-0.98 0-0.49 0-0.24 0-0.12
Details 3.93-7.87 1.96-3.93 0.98-1.96 0.49-0.98 0.24-0.49 0.12-0.24
In the section “Optimization of the proposed method using simulated data”,

we compared the result of the artifact removal method using DWT and SWT
algorithms with different mother wavelets.
45
Chapter 6
A Novel Hybrid Artifact

Removal Method for EVS
Artifacts
6.1 Design a Hybrid Wavelet-based Artifact
Removal Method
None of the current artifact removal methods are able to remove the EVS
artifacts in a way that the resulting cleaned up EEG signals are usable for
physiological and neurological studies. In EVS/EEG studies, the artifact
to signal ratio is so large that, even if a small amount of the EVS artifacts
remained, the original brain signals are not detectable. We concluded that
there is a need to design a specific artifact removal method, which can deal
with the EVS artifacts. We proposed a novel method for EVS artifact re-
moval by combining regression analysis and wavelet decomposition 1 . In
1
This work was submitted to the IEEE TRANSACTIONS ON BIOMEDICAL ENGI-
NEERING and currently is under revision
46
6.1. Design a Hybrid Wavelet-based Artifact Removal Method
this method, we used wavelet transform to decompose the EEG and EVS
signals into a number of frequency bands (e.g. 12 frequency bands, when the
sampling frequency is 1kHz). Then, we employed the regression methods to
estimate the EVS artifacts in each frequency band, using the decomposed
reference components within the same frequency band. In this method, the
wavelet decomposition helps to obtain the best estimate of the EVS con-
tribution in each EEG channel, since the regression model deals with the
frequency components of each band once at a time. In other words, the
recorded EVS signals and the contaminated EEG data are broken down
into various frequency bands by means of wavelet analysis, then the EVS
artifacts are estimated separately in each frequency band through the regres-
sion analysis. After estimating the EVS artifacts in each frequency band,
the estimated artifacts are subtracted from the contaminated EEG in that
frequency band. The cleaned up EEG signals can be reconstructed from the
artifact-removed components in the frequency range of interest (e.g. 1Hz to
32Hz).
The wavelet decomposition enables us to focus on the frequency bands of

interest and enhances the estimation of the EVS artifact in each frequency
band using dedicated regression models. Fig. 6.1 presents the flowchart of
the proposed method.
Using the wavelet decomposition, we can separate the EVS artifacts fre-
quency components, thus the regression-based estimation method can deal
better with the nonlinear behavior of the skin conduction. This wavelet
based time-frequency analysis approach enhances the performance of the
47
6.1. Design a Hybrid Wavelet-based Artifact Removal Method
regression analysis, and the hybrid proposed method outperforms the other
methods. The comparison results are presented in the chapter “Study Re-
sults and Assessment”.
Figure 6.1: Flowchart of the proposed method
48
6.2. Impedance Analysis and Simulation Study for Method Optimization
6.2 Impedance Analysis and Simulation Study
for Method Optimization
To optimize the performance of the proposed method, and to quantify the ac-
curacy of the method relative to alternative approaches, a simulation study
was carried out. We simulated the contaminated EEG signals with EVS ar-
tifact in a way that we exactly know what is the true EEG signal and what
is the artifact. For simulating the contaminated data, we combine some
artifact-free EEG data with the simulated EVS contamination. To be able
to simulate the actual process of the EVS contamination in EEG signals, we
analyzed the skin-electrode interface and simulated the physical structures
of EVS and EEG electrodes contacts and skin. The simulation of the elec-
trical interactions between EVS and body is based on a physiological model,
where the EVS and EEG electrode-skin interfaces and skin were replaced
with their electrical equivalent. Then, the EVS signals were applied to the
model to generated the simulated EVS artifacts.
In this simulation study, we assumed that the EVS current mainly dis-
tributes through the scalp and the deeper layers of the head have negli-
gible effects, considering the skull impedance is much higher than scalp
impedance (skull conductivity is around 0.018 S.m−1 and scalp conductiv-
ity is around 0.44 S.m−1 ) [30]. We modeled the physical structure of the
electrical impedance of the electrode-skin interface and also the impedance
of the head between the points that the EEG and the EVS electrodes were
placed, using a resistive-capacitive circuit [33] as shown in Fig. 6.2.
49
Figure 6.2: Electrical equivalent circuit for the electrode-skin interface and
the underlying skin [33]
In this electrical equivalent circuit, Ehe is the half cell potential of the elec-
trode/gel interface, the parallel combination of resistive Rd and capacitive
Cd components represent the impedance associated with the electrode-gel
interface. Rs is the series impedance associated with the resistance of the
electrode gel. Ese is the potential difference across the epidermis, whose
impedance is represented by the resistance Re and capacitance Ce .
In general, the dermis and the subcutaneous layer under it behave as an

equivalent pure resistance Ru . The deeper layers of the skin, containing
vascular, hair follicles, nervous components, contribute very less to the elec-
50
trical skin impedance. For example the range of the resistivity of the inner
tissues (Ru ) is typically 100 − 200Ωcm2 and for the outer layers (Re ), it
is between 104 − 106 Ωcm2 along with a capacitances (Ce ) in the range of
1 − 50nF/cm2 [24]. However, the sweat glands and ducts can add an equiv-
alent parallel RC network (represented by broken lines in Fig. 6.2) and a
potential difference between sweat glands, ducts, dermis and subcutaneous
layers [33].
Since the pure resistance of the deeper layers of skin and the resistance of
the electrode gel are negligible, we can simplify the impedance structure as
following (6.1):
Rd Re Rp sB1 + B0
Z(s) ≈ ( + )≈ 2 (6.1)
sRd Cd + 1 sRe Ce + 1 sRp Cp + 1 s A2 + sA1 + 1
where, s is the complex frequency variable, A2 ,A1 ,B2 and B1 represent

specific combinations of Rd ,Re ,Rp ,Cd , Ce and Cp for each electrode. This
model-based identification approach suggests the following relation between
the injected EVS current and the collected EEG voltage at a given elec-
trode(6.2):
sB1 + B0
Em = Xin + E + Wnoise (6.2)
s2 A2+ sA1 + 1
where, Em is the measured EEG, Xin is the injected EVS current, E is the
original neural signals or artifact-free EEG and Wnoise is the measurement
noise.
In the simulation study, the impedance structure is modeled and used to
51
simulate the EVS contribution (artifact) at each EEG channel (6.3):
∗ sB1 + B0
Em = Xin (6.3)
s2 A2+ sA1 + 1
∗ represents the EVS artifacts in the measured EEG signals. The

where, Em
modeled impedance structure between EVS electrodes and EEG electrodes
was used to calculate the output voltage due to the EVS current (the EVS
artifact) at each EEG electrode. We calculated the fit percentage of the
simulation output and the measured EEG at each EEG electrodes in the
20-channel EEG/EVS studies and the result is shown in Fig. 6.3.
100
90
80
70
Fit percentage
60
50
40
30
20
10
0
0 5 10 15 20
EEG channels
Figure 6.3: Fit percentage between the simulation output and the measured
EEG at each channel
The fit percentage is a measure of the relative energy fraction in the simu-
52
lated EVS artifact, calculated as (6.6):
∗ (t))2
(Em (t) − Em
f it = 100(1 − ) (6.4)
( (Em (t) − mean(Em (t))2
As shown in this figure, the fitness of simulated EVS artifact is higher at

the EEG electrodes which are closer to the EVS electrodes and it is lower
at the further channels, such as mid-line Parietal channel (Channel 15: Pz),
mid-line Central channel (channel 10: Cz) and Frontal channels (Channel
5: Fz,channel 1: FP1, and channel 2: FP2). Please refer to Fig. 4.2 for the
location of these channels.
According to (6.1), the skin impedance can be modeled by a low-order,

continuous-time transfer function with one zero and two poles. The transfer
function of the skin impedance between each EEG electrode and the EVS
electrode was calculated through a system identification process based on
an iterative nonlinear least-squares algorithm, where the EVS current was
taken as the input (I) and the measured EEG signals were taken as the
output of the system (V) and then the parameters of the impedance model
at each EEG electrode were identified. For instance, the impedance at the
channel 18 (O1, occipital) was modeled by (6.5):
1 + sTz
Z ∗ (s) = Kp (6.5)
s2 Tw2 + 2sζ.Tw + 1
with Kp = −40921 , Tw = 0.10848, ζ = 4.7863 and Tz = −2.3726.
53
The transfer function of the modeled impedance is shown in Fig. 6.4.
Figure 6.4: The transfer function of the modeled impedance at the occipital
channel
The fit percentage of this model is about 87%. The fit percentage is calcu-

lated as given by (6.6):

(V (t) − Z ∗ .I(t))2
f it = 100(1 − ) (6.6)
( (V (t) − mean(V (t))2
54
We separately computed the impedance models based on the collected EEG

data during the EVS application in each trial for each subject. We obtained
different impedance model for each trial with different EVS. The impedance
models are different in different trials, because the impedance depends on
the skin condition and may vary based on the variation of the physiological
condition due to biochemical additives, perspiration and degree of hydra-
tion, valence and concentration of electrolytes, temperature, dermatological
condition (e.g. diseases), thyroid activity, and different emotional states.
Even the properties of the EVS current stimulation such as waveform, am-
plitude and duration may also affect the electrical properties of the body
[24].
To address the concern about the time-variant properties of the scalp impedance
even during each trial, we computed the impedance models for shorter time
intervals during EVS. We observed that the length of time intervals does
not significantly affect the fitness of the models, which means the impedance
model does not change during each trial with duration up to few minutes.
For example, in one of our EEG/EVS studies where the EVS duration was
70 seconds, we calculated the impedance models for 1s, 2s, 5s, 7s, 10s, and
14s time intervals and analyzed the fitness of the impedance model at each
channel. Fig.6.5 and Fig.6.6) show how the fitness of the impedance model
at channel 18 varies in different time intervals. The results show that for
most of the intervals, the fitness of the model is very close to the fitness
obtained by using the entire trial, which is around 87%.
55
100
90
80
70
Fit percentage
60
50
40
30
20
10
0
0 10 20 30 40 50 60 70
Number of time intervals
(a)
100
90
80
70
Fit percentage
60
50
40
30
20
10
0
0 5 10 15 20 25 30 35
(b)
Figure 6.5: The fit percentage for the impedance model at channel 18 for
time intervals (a) 1sec and (b) 2sec
56
100
90
80
Fit percentage 70
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
(c)
100
90
80
70
Fit percentage
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10
(d)
100 100
90 90
80 80
70 70
Fit percentage
Fit percentage
60 60
50 50
40 40
30 30
20 20
10 10
0 0
1 2 3 4 5 6 7 1 2 3 4 5
Number of time intervals Number of time intervals
(e) (f)
Figure 6.6: The fit percentage for the impedance model at channel 18 for
time intervals (a) 5sec, (b) 7 sec, (c) 10sec, and (d) 14sec
57
6.3. Optimization of The Proposed Method Using Simulated Data
This indicates that the impedance of the scalp does not change during each
EVS trial and the impedance model can be represented by one transfer
function for the entire trial with a duration of few minutes in the EEG/EVS
studies.
To generate the simulated EVS artifacts, we applied the EVS current sig-
nals to the transfer functions of impedance models. In order to generate a
data set with known EEG and EVS artifact components, we combined the
simulated EVS artifacts and the clean EEG data which were acquired under
rest condition without application of any stimulation. This facilitates opti-
mization of the proposed method as well as comparison of the performance
of different methods in removing the EVS artifacts.
6.3 Optimization of The Proposed Method Using
Simulated Data
We used the simulated data to optimize the performance of the proposed

method. Having the simulated artifact, we could compare the artifact-
removed EEG signals with the original artifact-free EEG signals. We mea-
sured the performance of the proposed method using the correlation analysis
and the normalized residual sum of squares, which is also called the normal-
ized quadratic error.
For the correlation analysis, we calculated a Pearson’s correlation coefficient

between the artifact-removed EEG signals and the original artifact-free EEG
58
signals. The correlation coefficient is defined by:
Cov(Et , Êt )
Corr(Et , Êt ) = (6.7)
σEt · σÊt
where, Et is the original artifact-free EEG signal and Êt is the artifact-
removed EEG signal.
Another measurement of the performance is the normalized residual sum of

squares (RSS), defined by :

(Et (t) − Êt (t))2
RSSN = (6.8)
(Et (t) − mean(Et (t)))2
where, Et (t) represents the original artifact-free signal and Êt (t) is the
artifact-removed signal.
The first step in optimizing the proposed method is to choose the wavelet
algorithm and the mother wavelet, so that the performance of the method
is maximized, based on correlation and RSS analysis. To compare different
wavelet algorithms and mother wavelets, we employed a number of mother
wavelets from two different wavelet family which have been commonly used
in EEG signal processing, such as Daubechies (db3, db4, and db5) and Symlets
(sym3, sym4, and sym5). Both SWT and DWT algorithms were used with
these mother wavelets in the proposed artifact removal method and applied
to the simulated data and decomposed the data into 12 frequency bands.
Then the EVS artifacts were estimated using separate OE model structures
59
with order 2, within each frequency band.
In Table 6.1, we tabulated the normalized residual sum of squares and the
correlation between the original artifact-free signals and the artifact-removed
signals, reconstructed by the components in the frequency range lower than
32 Hz.
Table 6.1: Correlation and normalized residual sum of squares between the
artifact-removed signals and the original artifact-free EEG signals for simu-
lated data using different wavelets
DWT db3 DWT db4 DWT db5 DWT db6

Corr 0.8781 0.9023 0.9155 0.9242
RSSN 0.5517 0.4870 0.4503 0.4255
DWT sym3 DWT sym4 DWT sym5 DWT sym6
Corr 0.8781 0.9023 0.9156 0.9242
RSSN 0.5517 0.4870 0.4503 0.4255
SWT db3 SWT db4 SWT db5 SWT db6
Corr 0.9932 0.9933 0.9933 0.9932
RSSN 0.1710 0.1700 0.1705 0.1714
SWT sym3 SWT sym4 SWT sym5 SWT sym6
Corr 0.9932 0.9933 0.9933 0.9932
RSSN 0.1710 0.1700 0.1705 0.1714
As shown, the SWT algorithm has a superior performance compared to

DWT algorithm and between different mother wavelets both Daubechies
and Symlet wavelets with order 4, performed better than the others with the
highest correlation and the least RSS. The results also illustrate the relative
robustness of the method (small variations in the performance, when varying
the order).
60
Another step to improve the performance of the method, is to find an op-

timum regression method to estimate EVS artifacts as accurate as pos-
sible. We used three different classes of model structures: Output-Error
(OE) as a simple special case of the general polynomial model, non-linear
Hammerstein-Wiener with the piecewise-linear function and Space-State
models. We previously introduced all of these models in the section “Regression-
based artifact removal methods”.
We applied the proposed method to the simulated data, using each of these
models with different orders. Data was decomposed in 12 frequency bands
using SWT with Daubechies 4. In Table 6.2, we tabulated the normalized
residual sum of squares and the correlation between the original artifact-free
signals and the artifact-removed signals reconstructed by the components in
the frequency range lower than 32 Hz.
Table 6.2: Correlation and normalized residual sum of squares between the
artifact-removed signals and the original artifact-free EEG signals for simu-
lated data using different regression models
OE2 OE3 OE4

Corr 0.9933 0.9933 0.9933
RSSN 0.1700 0.1701 0.1704
NLHW2 NLHW3 NLHW4
Corr 0.9934 0.9926 0.9851
RSSN 0.1711 0.1230 0.1725
SS2 SS3 SS4
Corr 0.9933 0.8105 0.7466
RSSN 0.1704 0.7628 0.9174
As already discussed in the section “Regression-based artifact removal meth-
61
ods”, we used the piecewise-linear function for nonlinear Hammerstein-

Wiener models, as nonlinear algebraic functions and broke down the EEG
signals into a different number of intervals, from 2 to 20 intervals. We ob-
served that, when we use Hammerstein-Wiener model with up to 4 intervals
in the piecewise-linear function, we could get the highest correlation and
the least residual sum of squares. This indicates that electrical properties
of the scalp during the EVS application (for few minutes) can be modeled
almost linearly, and they present only mild nonlinearity. The outcome of
the simulation study indicates the same point, as previously shown in Fig.
6.6.
The results of the regression analysis, using different regression model struc-
tures, show between all those models the Output-Error order 2 outperforms
the other models, followed by the nonlinear Hammerstein-Wiener order 2
and 3 and State-Space order 2. We employed the Output-Error order 2
model in the proposed method to achieve the best performance, when it is
applied to the real data.
62
Chapter 7
Study Results and

Assessment
To study the effectiveness of different methods in removing the EVS ar-

tifacts, we applied the artifact removal methods on real EEG data, con-
taminated with artifacts, induced by different EVS stimulations, such as
stochastic EVS (SVS) and galvanic EVS (GVS). In these EEG/EVS stud-
ies, we applied EVS electric current to a number of subjects, at different
amplitude levels, frequencies and durations and collected a large amount of
EEG/EVS data (over 30 GigaByte of data). Then, we applied the artifact
removal methods on randomly chosen EEG/EVS data from different studies
(and different subjects), to assure the accuracy and reliability of the results.
We observed consistent results in analysis of the collected experimental data.
However, here we just reported the best results that we could achieve.
To compare these artifact removal methods, we applied these methods on

the EEG signals acquired from occipital channel O1 (please refer to Fig.
4.2), while the same EVS stimulation were delivered to different subjects.
63
7.1. Assessment of The Artifact Removal Methods in Stochastic EVS Studies
We chose the occipital channel because the data acquired from this channel,
which contains important physiological information, gets heavily contami-
nated by the EVS artifacts, since the EEG electrodes for occipital channels
are located very close to the EVS electrodes.
7.1 Assessment of The Artifact Removal
Methods in Stochastic EVS Studies
To assess the performance of the artifact removal methods in the SVS stud-
ies, we applied the methods to the EEG/EVS data, collected from four
different subjects, where the stimulus was a recorded zero-mean pink noise
current, with a 1/f type power spectrum within frequency range of 0.1Hz to
10Hz and duration of 72 seconds, delivered at 5 different RMS levels between
60uA to 450uA.
First we employed the ICA-based methods, the adaptive filters and the
regression-based methods. Then we selected the best methods with higher
performance from each type of these methods and compared them with
our proposed method, where we used the wavelet analysis to improve the
performance of the regression-based artifact removal methods.
To quantify the performance of these methods and compare them, we used

the correlation analysis. We calculated the correlation coefficients between
the EVS signals and the estimated EVS artifacts for each method.
64
Between different ICA algorithms, we observed that the extended Infomax

performed better. Although in some cases, the Fast-ICA algorithm provided
as good results as the extended Infomax, it is not as robust as the extended
Infomax ICA. In many cases, the Fast-ICA does not converge, especially
in the studies with the higher EVS amplitude levels. We therefore only
reported the result of the extended Infomax ICA (Table 7.1).
Between adaptive filters, as previously discussed in the section “Adaptive

filter-based artifact removal methods”, we employed both RLS-based and
LMS-based filters. We used these filters with different properties, such as
length, forgetting factors and adaptation gains.
In Fig. 7.1 and Fig. 7.2, we depicted the result of the correlation analysis for
the RLS-based filters with different length and different forgetting factors.
We observe that the filter with the forgetting factor of 0.99997 and the length
of 2 had the best performance.
In Fig. 7.2 and Fig. 7.3, we depicted the result of the correlation analysis
for the LMS-based filters. It is shown that between these LMS-based filters
with different length and adaptation gains, the filter with the adaptation
gain of 0.5 and the length of 3, had the best performance.
65
Figure 7.1: Correlation between the EVS signal and the estimated EVS
artifact using the RLS filters
66
artifact using the RLS filters (top) and the LMS filters (bottom)
67
artifact using the LMS filters
68
Between the regression-based methods, as we already examined in the opti-

mization process, the second order Output-Error model with the frequency
weighted maximum likelihood estimation method, provided the best results.
In Table 7.1, we tabulated the results of the correlation analysis for these
methods and compared the correlation coefficients between the EVS signals
and the EVS artifacts estimated by each of these methods in the real data
sets.
Table 7.1: Correlation between the EVS signals and the estimated EVS
artifact extracted from EEG signals for real data using different methods
Method Correlation
ICA-Infomax method (Removing the artifact component) 0.6859
ICA-Infomax method (Applying threshold to the artifact component) 0.6858
Regression method with OE2 0.7673
RLS Adaptive filter (Forgetting factor: 0.99997, length:2) 0.7615
LMS Adaptive filter (Adaptation gain: 0.5, length:3) 0.7010
The results show that between all the methods, that mentioned earlier, the
second order Output-Error regression model can estimate the EVS artifacts
with a higher correlation with the original EVS signals. This is also in con-
formance with our skin impedance analysis, where we analyzed the resistive-
capacitive equivalent circuit of the electrode and skin physical structures as
shown in Fig. 6.2.
We conclude that, as we use a model based on the actual physical structure,

our method can deal with the artifactual contamination better than other
methods, including statistical methods such as ICA. We had the same ob-
servation in the simulation study as well, where the second order regression
69
models matched well with the skin and electrodes impedance model.
Although the regression-based methods alone performed better than the

other methods, they cannot completely clean up the EEG signals and the
remaining artifacts are still large that the original neural signals are still
buried in the residual. We employed the wavelet decomposition along with
the regression-based method in our proposed method to achieve the best
performance in removing EVS artifact.
The wavelet decomposition method improves the estimation of the EVS

artifacts in both correlation performance and robustness. This is due to the
separate transfer function estimations for each frequency band, which makes
the regression model less prone to nonlinear skin behavior and other noise
sources.
Using the wavelet decomposition method, we focus on the frequency bands

of interest, and estimate the artifact in each frequency separately, thus the
regression models can deal better with the frequency nonlinearity attributed
to the nonlinear behaviors of the skin. To investigate the performance of
the proposed method in each frequency band separately, we applied the
proposed method to the simulated EVS contaminated data. Fig. 7.1 and
Fig. 7.5 show the fit percentage and the correlation between the estimated
EVS artifacts and the simulated EVS contaminated signals in frequency
bands introduced in Table 5.1.
70
100
90
80
70
Fit percentage
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10 11 12
Frequency bands
Figure 7.4: Fit percentage of the estimated EVS artifacts with simulated
EVS contaminated signals using the OE order 2 in 12 frequency bands
0.9
0.8
Correlation coefficient
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9 10 11 12
Frequency bands
Figure 7.5: Correlation between the estimated EVS signals and simulated
EVS contaminated signals using the OE order 2 in 12 frequency bands
71
The results show that in the frequency range of the EVS signals, we can
achieve better results in estimating the artifacts in terms of fit percentage
and correlation coefficients. For example, when the EEG signals sampled
at 1kHz, are split into 12 frequency components, in the frequency compo-
nents L6 to L10, which approximately correspond to 8-16Hz, 4-8Hz, 2-4Hz,
1-2Hz and 0.5-1Hz bands, we can achieve 96% fit percentage with the cor-
relation coefficient of 0.99. It explains why the regression-based estimation
achieve higher performance in rejecting the EVS artifacts, when the artifact
is estimated in each frequency band separately.
Another benefit of using wavelet decomposition is that we can filter out the
frequency components that are out of interest. Removing those frequency
components can, in effect, be equivalent to reducing noise in the estimation
process and improve the results of the regression analysis.
We removed the frequency components, which were out of our interest, (out-
side the band between 0.98Hz and 31.25Hz), and reconstructed the artifact-
removed EEG data in the frequency range of interest. Through a correlation
analysis, we can demonstrate that the wavelet based time-frequency analy-
sis approach significantly enhances the performance of the artifact removal
method.
We calculated the correlation coefficients between the EVS signals and the
estimated EVS artifacts reconstructed from different frequency bands, and
we illustrated how the wavelet analysis improves the regression estimates.
The result of the correlation analysis is tabulated in Table 7.1.
72
Table 7.2: Correlation between the EVS signal and the estimated EVS ar-
tifact reconstructed from different frequency components for real data
Frequency band Correlation

Estimated EVS artifact without wavelet decomposition 0.7673
Estimated EVS artifact from 0.12Hz to 250Hz 0.8463
Estimated EVS artifact from 0.24Hz to 125Hz 0.9168
Estimated EVS artifact from 0.49Hz to 62.5Hz 0.9725
The results show that the correlation between the EVS signal and the esti-
mated EVS artifact significantly increases by using wavelet decomposition
method. After removing the frequency components lower than 0.98Hz and
higher than 31.25Hz, which are not of the main interest, the correlation
coefficient between the EVS signal and estimated EVS artifact range was
increased up to 0.9899, while without using wavelet decomposition, the cor-
relation coefficient between the EVS signal and estimated EVS artifact was
0.7673. As shown in the Table 7.1, removing different frequency components
affects the results of the regression estimates.
We could significantly suppress the EVS artifact and achieve high Signal to
Artifact Ratio (SAR), defined as:
RM SEEG
SAR = 20Log10 ( ) (7.1)
RM SArtif act
73
After estimating the EVS artifacts in separate frequency bands, we could

reconstruct the artifact-removed EEG signals from the frequency bands of
interest, which resulted in the removal of a major portion of the artifact.
For instance, in the EEG/EVS studies, where the stimulus was a pink noise
with a RMS value of 100uA, we could achieve a SAR of −0.7057dB in the
frequency range of 1Hz-16Hz (Fig. 7.6), which means a large amount of the
artifact is removed, considering that the correlation coefficient between the
removed artifactual components and the EVS signals is 0.9899.
In the other frequency ranges, the artifact residual is larger and we have
lower SAR. For example, in the frequency range of 1Hz-32Hz (Fig. 7.1), we
could obtain a SAR of −4.5592dB and in the range of 1Hz-64Hz, the SAR
was −6.0206dB (Fig. 7.8).
It is worth mentioning that in the original contaminated EEG signals, as

a result of the SVS application with a RMS value of 100uA, the SAR was
−15.171dB before removing the EVS artifact (Fig. 7.9).
74
150
100
EEG voltage (uV)

50
−50
−100
−150
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
5
Time (msec) x 10
Figure 7.6: The Occipital EEG channel data after applying the proposed
artifact removal method using the frequency components between 1-16Hz
250
200
150
100
EEG voltage (uV)
50
−50
−100
−150
−200
−250
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
5
Time (msec) x 10
artifact removal method using the frequency components between 1-32Hz
75
300
200
100
EEG voltage (uV)
−100
−200
−300
−400
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
5
Time (msec) x 10
artifact removal method using the frequency components lower than 64Hz
2500
2000
1500
EEG voltage (uV)
1000
500
−500
−1000
−1500
−2000
−2500
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

5
Time (msec) x 10
Figure 7.9: The Occipital EEG channel data before removing EVS artifacts
76
After removing the EVS artifact in the frequency band of 1Hz to 16Hz, we
analyzed the power spectrum of the EEG data before, during and after EVS
application. Fig. 7.10 shows that during the EVS application, there is an
increase in the power in frequency bands of 1-12Hz and 20-28Hz.
Power density (dB)

50
45 20
40 10
35 0
Frequency (Hz)
30 −10
25 −20 dB
20
−30
15
−40
10
−50
5
−60
20 40 60 80 100 120 140 160 180
Time(s)
Power density (dB)

50
50
45
40
40
35 30
Frequency (Hz)
30 20
25 10 dB
20
0
15
−10
10
−20
5
−30
20 40 60 80 100 120 140 160 180
Time (s)
Figure 7.10: Power spectrum of the EEG data before, during and after EVS
application (top) and Power spectrum of the EVS (bottom)
77
As discussed so far, compared to the traditional methods, the proposed

artifact removal method has superior performance when it is applied to the
low-amplitude stochastic EVS (up to 1mA).
To study the performance of the proposed method in dealing with larger

artifacts due to high-amplitude EVS stimulations, we conducted some high-
amplitude EVS studies in collaboration with the Sensorimotor Physiology
Laboratory. We collected some EEG/SVS data, where the stimulation signal
was a pink-noise delivered at amplitude levels from 100uA to 3100uA (each
300uA).
We collected these data to investigate the performance of artifact removal

methods, as the EVS amplitude is increased. We applied different artifact
methods to remove the high-amplitude EVS artifacts and compared their
performances. The regression-based methods provide more robust results
than the ICA-based methods.
We used a correlation analysis to compare the proposed method and the

ICA method, and the results show that while the performance of the ICA
method deteriorates as the stimulation amplitude is increased, the proposed
method provides a superior robust performance (Fig. 7.11).
78
7.2. Assessment of The Proposed Artifact Removal Method in Pulsed GVS Studies
Figure 7.11: Correlation between the EVS signals and the estimated EVS
artifacts using the proposed method (red) and the ICA method (blue) for
different stimulation amplitudes
7.2 Assessment of The Proposed Artifact
Removal Method in Pulsed GVS Studies
To assess the performance of the proposed artifact removal method in the

square pulse EVS (or GVS) studies, we conducted some EEG/GVS studies
in collaboration with the Sensorimotor Physiology Laboratory.
79
We collected the EEG/GVS data; the applied stimulation signal was a

square wave GVS pulses with amplitude levels between 0.5mA to 3mA and
duration between 500ms and 10 seconds. These EEG/GVS experimental
data were collected for the postural/balance studies, which is the area of
interest of our collaborator in the Sensorimotor Physiology Laboratory.
This type of stimulation has been commonly used to study the GVS evoked
postural responses and characterization of muscles and whole-body responses
[9, 28]. Even, in some of the studies of this type, square wave GVS pulses
with higher amplitude (e.g. 4mA) and shorter duration (sometimes as short
as 20ms) are applied [28].
In the vestibular-evoked postural/balance response studies, since the muscles

onset latency is in the order of tens of millisecond (e.g. around 50ms [28]),
we need to clean the EEG data in a very short period of time (as short as
few milliseconds) after stimulus onset, which introduces a new challenging
problem to solve, and it is out of the scope of this project.
The proposed method is not suitable when we need to acquire the EEG data
immediately after the GVS onset, because we cannot accurately reconstruct
large sharp jumps by means of wavelet transform. When we applied the
proposed method to remove the high-amplitude GVS artifact, we observed
additional large ripples near the sharp edges of the square pulses, due to
the Longo phenomenon. The Longo oscillations near the edge of the high-
amplitude GVS pulses reflect the difficulty that we faced in approximating
a huge sharp edge by a finite series of wavelets. As a result, the original
80
EEG signal immediately after the GVS onset was not completely recovered.
Except for the short period after the GVS onset, we are still able to remove
a major portion of the GVS artifacts from the EEG signals. Following we
discuss about how we employed the proposed method to remove the GVS
artifacts.
Fig. 7.12 shows a sample of the measured EEG signal during the GVS with
the amplitude of 0.5 mA. We applied the proposed artifact removal method
and using the wavelet decomposition, split the signals into a number of fre-
quency components. Considering that if the EEG signals are reconstructed
from all the frequency components, the artifact is still dominant, as shown in
Fig. 7.13. In this figure (Fig. 7.13), we can also observe the large oscillations
(Longo phenomenon), near the edge at both ends of the GVS signal.
Hence, we reconstructed the cleaned up EEG signals from the frequency

components of interest, to improve the results. The remaining GVS artifact
was removed from the EEG signals, when we reconstructed the EEG signals
with the components in a frequency band between 4Hz to 32Hz. However,
about 300 milliseconds of the EEG signals were lost at both ends of the
GVS, due to the Longo phenomenon, as shown in Fig. 7.14.
81
Figure 7.12: EEG signal during the GVS pulses with the amplitude of 0.5mA
Figure 7.13: The cleaned up EEG signal during the third GVS pulse using
the proposed method, reconstructed from all the frequency components
82
Figure 7.14: The cleaned up EEG signal reconstructed from the frequency
components between 4-32Hz (top) and GVS artifact component (bottom)
83
We observed that the proposed method is very efficient in removing the high-
amplitude GVS artifacts, but the immediate data after the GVS onset were
not fully recovered and we lost some data around few hundreds of millisec-
onds. This aspect is important in the vestibular-evoked postural/balance
response studies, where researchers are looking for data from 5 milliseconds
after the GVS onset. There is therefore, a need to develop another method,
which can deal with the sharp edges of the GVS artifacts and recover the
EEG data right after the GVS application.
84
Chapter 8
Conclusion and Future Work
8.1 Summary and Conclusion
In this project, we introduced a novel method to resolve one the most

challenging problems in regards with EEG/EVS studies. To optimize the
method, assess its performance and assure the reliability of the results, we
performed a simulation study. In the simulation study, we analyzed the skin
and the electrode-skin interface impedance and we demonstrated that by
modeling the impedance and estimating the transfer function of the skin-
electrode structure, we can simulate the EVS artifacts, which showed a good
fit with the original EVS signals. By combining the simulated artifact and
clean EEG data, we simulated EVS-contaminated EEG signals. Thus, we
were able to assess the performance of the proposed method using the sim-
ulated signals, where both neural signals and artifacts were known. We
showed that we could reconstruct up to 96% of the EVS artifact compo-
nents in some frequency bands, especially in the frequency range of the EVS
signals (Fig. 7.1).
85
8.1. Summary and Conclusion
We concluded that employing the wavelet decomposition along with the

time-domain regression methods can greatly improve the estimation of the
EVS artifacts. By combining the regression method and wavelet analysis
in the proposed artifact removal method, we were able to focus on different
frequency bands and significantly enhanced the SAR of the contaminated
EEG data in specific frequency bands.
Most of the current researches in the field of EEG analysis, have been focused
on the component-based methods, such as ICA, to remove different types
of artifacts from EEG signals. The proposed method behaves differently in
rejecting the EVS artifact and it was shown that our method can outperform
the methods commonly used in the field.
At present, using the proposed method, we achieved a high signal to arti-

fact ratio and a good correlation between the estimated EVS artifacts and
the original EVS signals, while using the ICA-based methods we could not
obtain good results in terms of both signal to artifact ratio and correla-
tion coefficients. This indicates that the ICA-based methods may introduce
some undesirable variation in the EEG signals, while the proposed artifact
removal method introduces the least distortion in the EEG signals, com-
pared to other methods.
The specific contribution of this thesis consists in developing a novel artifact

removal method, which outperforms the other methods in removing EVS
artifacts. Artifact removal methods are the most essential tool in analyzing
EEG studies, especially for different stimulations studies. In this project, we
86
8.1. Summary and Conclusion
also analyzed the impedance properties of the head (specifically scalp) and
skin-electrode interface. A simulation study was performed and its results
showed that the impedance models for the skin and skin-electrodes interfaces
remain fairly constant over the entire trial (up to few minutes) and the small
variation in impedance during the EVS stimulation is not considerable. Also,
when we applied the proposed method on the real data, using the second
order nonlinear Hammerstein-Wiener model with few breakpoints in the
piecewise-linear block, it provided the same results as the Output-Error
model of the second order. This implies that the scalp impedance does
not vary much during the EVS application, and relationships between the
injected EVS current and the EVS artifacts at the EEG electrodes remain
constant over the entire trial.
Furthermore, using the proposed method, we don’t need to collect and pro-
cess multiple EEG channels as oppose to the ICA-based analysis, therefore
this method is much faster than the ICA-based methods. This also allows
us to have a simple experimental set-up for collecting EEG signals with few
EEG channels for the EVS studies, which makes the data acquisition easy
and less time consuming in preparation. When the acquisition session takes
a long time, the subject gets tired and as a result more myogenic and oc-
ular artifacts are introduced in the EEG data. On the other hand, as the
in proposed method less data are processed, it is easier to implement in a
real-time system.
However, as illustrated earlier, we could not completely remove the EVS

contamination in all frequency ranges and in the GVS studies, we could not
87
8.2. Areas of Future Studies
recover the EEG signals in a short period of time (in the order of millisec-
onds) right after the GVS onset.
8.2 Areas of Future Studies
Since there is still some room to reduce the EVS-induced artifacts, which re-
main in the EEG data after applying the proposed method in the frequency
band out side the 1Hz to 16Hz, the immediate future work would be the
improvement of the proposed artifact removal method. We suggest that a
combination of the proposed method with other methods may improve the
results. For example, employing the ICA methods after using the proposed
method, may improve the results. Applying the proposed method can re-
move a major part of the EVS artifacts and reduce the artifact to signal ratio
significantly, while ICA-based methods may help to remove the remaining
artifacts in broader frequency bands such as 1Hz to 32Hz, by considering
broader statistical correlations, beyond the linear ones.
Moreover, the proposed artifact removal method may need some modifica-
tions when EVS with different waveforms, power spectra and duration are
employed in EVS studies. Nevertheless, the proposed method may not be
suitable for some special studies. For example, in the case of vestibular-
evoked studies where high-amplitude GVS stimulation is used, we lose some
EEG data right after the GVS onset, while those data are important in some
physiological studies. In such cases, a refined method must be developed.
88
After cleaning up the EEG signals from the artifacts, the next phase of
this research would be analyzing the EEG signals during the EVS appli-
cation, identifying the effects of the EVS on the EEG signals and finding
the correlation between EVS and the induced EEG patterns. This leads to
characterization of the central nervous system responses to EVS stimula-
tion, e.g. vestibular-evoked postural/balance responses. Researchers need
this information to understand the vestibular system functions and its ef-
fects on different regions in brain, which may lead to find a way to alter the
brain dynamics using the EVS signals. Manipulating the dynamics of the
brain and its oscillatory activities may introduce new alternative non-drug
treatments for the vestibular and some of the neurophysiological disorders.
The stochastic EVS has been already applied in neurophysiological stud-

ies to investigate the effects of noisy stimulation on the patients with neu-
ral degenerative disorders, specifically multiple system atrophy (MSA) and
Parkinson’s disease (PD). Low amplitude pink noise EVS produces fluctua-
tions in biological systems which may lead the system to resonate at a par-
ticular noise level and it can improve the sensorimotor functions in patients
with neurodegenerative disorders [22, 23, 36]. In these types of procedures,
the EVS should be applied in accordance with the patients’ needs, thus the
stimulation can effectively suppress or control the symptoms of the disease.
To be able to find the best operational EVS parameters such as amplitude,

frequency and waveforms in any of these diagnostic or therapeutic applica-
tions, researchers need to have an optimized artifact removal method that
can be run in a fully automated manner. The proposed method has been
89
developed in a way that it is less computationally demanding and can be

implemented in embedded systems. This method is very promising, as it can
eventually be used in real time closed-loop applications, where the system
can automatically adjust the EVS, based on the EEG signals that it receives
(Fig 8.1).
Figure 8.1: Schematic of a closed-loop EVS/EEG system
90
Bibliography
[1] Tatyana I. Aksenova, Dimitri V. Nowicki, and Alim-Louis Benabid. Fil-

tering out deep brain stimulation artifacts using a nonlinear oscillatory
model. Neural Computation, 21(9):2648–2666, 2009.
[2] Anthony J. Bell and Terrence J. Sejnowski. An information-

maximization approach to blind separation and blind deconvolution.
NEURAL COMPUTATION, 7:1129–1159, 1995.
[3] A Garcs Correa, E Laciar, H D Patio, and M E Valentinuzzi. Artifact

removal from eeg signals using adaptive filters in cascade. Journal of
Physics: Conference Series, 90(1):012081, 2007.
[4] Maite Crespo-Garcia, Mercedes Atienza, and Jose Cantero. Muscle

artifact removal from human sleep eeg by using independent component
analysis. Annals of Biomedical Engineering, 36:467–475, 2008.
[5] Christopher J. Dakin, Billy L. Luu, Kees van den Doel, John Timo-
thy Inglis, and Jean-Sbastien Blouin. Frequency-specific modulation of
vestibular-evoked sway responses in humans. Journal of Neurophysiol-
ogy, 103(2):1048–1056, 2010.
91
Bibliography
[6] Arnaud Delorme and Scott Makeig. Eeglab: an open source toolbox for
analysis of single-trial eeg dynamics including independent component
analysis. Journal of Neuroscience Methods, 134(1):9–21, 2004.
[7] Valentina Dilda, Hamish G. MacDougall, and Steven T. Moore. Tol-

erance to extended galvanic vestibular stimulation: Optimal exposure
for astronaut training. Aviation, Space, and Environmental Medicine,
82(8):770–774, 2011.
[8] Yaara Erez, Hadass Tischler, Anan Moran, and Izhar Bar-Gad. Gener-
alized framework for stimulus artifact removal. Journal of Neuroscience
Methods, 191(1):45–59, 2010.
[9] Richard C. Fitzpatrick and Brian L. Day. Probing the human vestibu-
lar system with galvanic stimulation. Journal of Applied Physiology,
96(6):2301–2316, 2004.
[10] Jun Gao, Yong Yang, Pan Lin, Pei Wang, and Chong Zheng. Automatic
removal of eye-movement and blink artifacts from eeg signals. Brain
Topography, 23:105–114, 2010.
[11] Frederic Grouiller, Laurent Vercueil, Alexandre Krainik, Christoph

Segebarth, Philippe Kahane, and Olivier David. A comparative study
of different artefact removal algorithms for eeg signals acquired during
functional MRI. NeuroImage, 38(1):124–137, 2007.
[12] Takao Hashimoto, Christopher M. Elder, and Jerrold L. Vitek. A

template subtraction method for stimulus artifact removal in high-
92
Bibliography
frequency deep brain stimulation. Journal of Neuroscience Methods,

113(2):181–186, 2002.
[13] Ping He, Glenn Wilson, and Christopher Russell. Removal of ocular
artifacts from the eeg: a comparison between time-domain regression
method and adaptive filtering method using simulated data. Medical
and Biological Engineering and Computing, 45:495–503, 2007.
[14] T Jung, S Makeig, C Humphries, T Lee, M Mckeown, I Vicente,

and T Sejnowski. Removing electroencephalographic artifacts by blind
source separation. Psychophysiology, (37):163–178, 2000.
[15] Te Won Lee, Mark Girolami, and Terrence J. Sejnowski. Independent

component analysis using an extended infomax algorithm for mixed
subgaussian and supergaussian sources. Neural Comput., 11(2):417–
441, 1999.
[16] Lennart Ljung. System Identification: Theory for the User (2nd Edi-
tion). Prentice Hall PTR, December 1998.
[17] Elie Lobel, Justus F. Kleine, Denis Le Bihan, Anne Leroy-Willig, and
Alain Berthoz. Functional mri of galvanic vestibular stimulation. Jour-
nal of Neurophysiology, 80(5):2699–2709, 1998.
[18] Ratko Magjarevic, Manousos A. Klados, C. Papadelis, C. D. Lithari,

and P. D. Bamidis. The removal of ocular artifacts from eeg signals: A
comparison of performances for different methods. volume 22 of IFMBE
Proceedings, pages 1259–1263. Springer Berlin Heidelberg, 2009.
93
Bibliography
[19] B. W. McMenamin, A. J. Shackman, J. S. Maxwell, L. L. Greischar,

and R. J. Davidson. Validation of regression-based myogenic correction
techniques for scalp and source-localized eeg. Psychophysiology, 46:578–
592, 2009.
[20] Brenton W. McMenamin, Alexander J. Shackman, Jeffrey S. Maxwell,

David R.W. Bachhuber, Adam M. Koppenhaver, Lawrence L. Greis-
char, and Richard J. Davidson. Validation of ica-based myogenic
artifact correction for scalp and source-localized eeg. NeuroImage,
49(3):2416–2432, 2010.
[21] F. Morbidi, A. Garulli, D. Prattichizzo, C. Rizzo, and S. Rossi. Appli-

cation of kalman filter to remove tms-induced artifacts from eeg record-
ings. Control Systems Technology, IEEE Transactions on, 16(6):1360–
1366, nov. 2008.
[22] S. Pal, S. M. Rosengren, and James G. Colebatch. Stochastic galvanic

vestibular stimulation produces a small reduction in sway in parkinson’s
disease. Journal of Vestib Res, 19(3-4):137–142, 2009.
[23] W. Pan, R. Soma, S. Kwak, and Y. Yamamoto. Improvement of motor

functions by noisy vestibular stimulation in central neurodegenerative
disorders. Journal of Neurology, 255:1657–1661, 2008.
[24] Mark R. Prausnitz. The effects of electric current applied to skin: A

review for transdermal drug delivery. Advanced Drug Delivery Reviews,
18(3):395 – 425, 1996.
[25] Alois Schloegl, Andreas Ziehe, and Klaus-Robert Mller. Automated
94
Bibliography
ocular artifact removal: comparing regression and component-based

methods. Nature Precedings, 2009.
[26] A. Schlogl, C. Keinrath, D. Zimmermann, R. Scherer, R. Leeb, and

G. Pfurtscheller. A fully automated correction method of eog artifacts
in eeg recordings. Clinical Neurophysiology, 118(1):98–104, 2007.
[27] Alexander Shackman, Brenton Mcmenamin, Heleen Slagter, Jeffrey

Maxwell, Lawrence Greischar, and Richard Davidson. Electromyogenic
artifacts and electroencephalographic inferences. Brain Topography,
22(1):7–12, June 2009.
[28] Gregory Martin Lee Son, Jean-Sbastien Blouin, and John Timothy In-
glis. Short-duration galvanic vestibular stimulation evokes prolonged
balance responses. Journal of Applied Physiology, 105(4):1210–1217,
2008.
[29] A T Tidswell, A Gibson, R H Bayford, and D S Holder. Electrical

impedance tomography of human brain activity with a two-dimensional
ring of scalp electrodes. Physiological Measurement, 22(1):167175, 2001.
[30] A T Tidswell, A Gibson, R H Bayford, and D S Holder. Electrical

impedance tomography of human brain activity with a two-dimensional
ring of scalp electrodes. Physiological Measurement, 22(1):167–175,
2001.
[31] Kathrin S. Utz, Violeta Dimova, Karin Oppenlander, and Georg

Kerkhoff. Electrified minds: Transcranial direct current stimula-
tion (tdcs) and galvanic vestibular stimulation (gvs) as methods of
95
non-invasive brain stimulation in neuropsychology. Neuropsychologia,
48(10):2789–2810, 2010.
[32] Garrick L. Wallstrom, Robert E. Kass, Anita Miller, Jeffrey F. Cohn,

and Nathan A. Fox. Automatic correction of ocular artifacts in the
eeg: a comparison of regression-based and component-based methods.
International Journal of Psychophysiology, 53(2):105–119, 2004.
[33] John G. Webster. Medical Instrumentation-Application and Design, 4th

Edition. New York: Wiley, 2009.
[34] David Wilkinson, Sophie Nicholls, Charlotte Pattenden, Patrick Kil-

duff, and William Milberg. Galvanic vestibular stimulation speeds vi-
sual memory recall. Experimental Brain Research, 189:243–248, 2008.
[35] Y. Yamamoto, R. Soma, Z. R. Struzik, and Shin Kwak. Can electrical

vestibular noise be used for the treatment of brain diseases? AIP
Conference Proceedings, 800(1):279–286, 2005.
[36] Y. Yamamoto, Z. R. Struzik, R. Soma, K. Ohashi, and S. Kwak. Noisy

vestibular stimulation improves autonomic and motor responsiveness in
central neurodegenerative disorders. Annals of Neurology, 58(2):175–
181, 2005.
96
Appendix A
DS5 Isolated Bipolar

Constant Current Stimulator
The DS5 has been developed for the new and increasingly important field of
Clinical Nerve Excitability Studies 2 . These studies, including electrotonus
and threshold electrotonus, are giving us added understanding of human
nerve excitability and the pathophysiology of diseases such as diabetic neu-
ropathy, carpal tunnel syndrome (CTS), amyotrophic lateral sclerosis (ALS),
multifocal motor neuropathy (MMN), motor neurone disease (MND) taxol-
cisplatin neuropathy and neuromyotonia.
Multiple Applications
Although the DS5 has been primarily designed for clinical studies of periph-
eral nerves using threshold tracking techniques, because the unit provides
a current output proportional to the voltage input and can deliver up to
50mA (from a compliance voltage of greater than 100V), it should appeal
2
All the material in this appendix is from the Digitimer company’s website:
www.digitimer.com
97
Appendix A. DS5 Isolated Bipolar Constant Current Stimulator
to any researchers wishing to safely apply a computer controlled constant

current stimulus to a patient or research volunteer.
Unlimited Stimulus Profile Possibilities
Unlike other clinical stimulators, the DS5 is not a traditional square wave
pulse stimulator. Instead, the DS5 will produce an isolated constant current
stimulus proportional to a voltage applied at its input, with the shape of
the input waveform describing the shape of the stimulus. As a result, when
driven by a computer running appropriate software, the DS5 is capable
of generating a stimulus consisting of multiple components including sine
waves, ramps, square waves or totally arbitrary waveforms. Patient safety is
assured at all times by the implementation of safety features which, amongst
other things, also limit the amount of energy passed to the subject.
Stimulus Control via Commercial or Specialist Hardware/Software
The DS5 has been developed in collaboration with Professor Hugh Bostock
at the Institute of Neurology in London, who uses the stimulator to facili-
tate the threshold tracking measurements alluded to above. Briefly, muscle
action potentials evoked by a test stimulus from the DS5 are amplified and
compared with a desired target response. The resultant error signal is used
to modulate the amplitude of the test stimulus. Once a baseline level of ex-
citability is established, conditioning pulses can be applied to examine their
effect on the excitability of the nerve being studied. By carrying out this
procedure using his own specialist software, diagnostic tests can be partially
automated and completed more rapidly.
98
Appendix A. DS5 Isolated Bipolar Constant Current Stimulator
The DS5 stimulator has four input voltage ranges: 1V, 2.5V, 5V and 10V
making it widely compatible with other hardware. There are three output
ranges: 10mA, 25mA and 50mA, with selection either by front panel con-
trols or Windows PC Control Software. On the rear of the device are two
BNC sockets, one accepting the voltage input waveform, the other provid-
ing a Monitor Output, allowing the stimulus waveform to be fed into a data
acquisition system so it can be recorded alongside your EMG data.
For more information about this device, please refer to the following link.
http://www.digitimer.com/clinical/ds5b ipolarc onstantc urrents timulator.htm
99
Appendix B
Model 2200 Analog Stimulus

Isolator
The Model 2200 Analog Stimulus Isolator is designed to electrically isolate

any waveform shape generated by an external signal source 3 . The signal
is DC-coupled, and can take on any waveshape within its wide bandwidth,
including pulses. Six output scaling options are available, covering a wide
range of constant current and constant voltage levels. For a given change in
the input signal, the Model 2200 output would scale according to the front
panel settings. If for any reason the instrument cannot follow the input
signal, an error indicator lights.
For this unit, the isolated output section is battery-powered and optically-
coupled to the input section, for the ultimate in clean isolation. Two
rechargeable batteries are utilized: a 9V battery that powers the control
circuitry, and a 12V battery that provides the signal power. Both recharge-
able batteries provides full function for at least 8 hours of continuous use.
3
All the material in this appendix is from the A-M Systems company’s website: www.a-
msystems.com
100
Appendix B. Model 2200 Analog Stimulus Isolator
During non-use periods, it is recommended that the unit remain on a trickle

charge by connection to the provided external battery charger.
For more information about this device, please refer to the following link.
http://www.a-msystems.com/s-139-analog-stimulus-isolator-model-2200.aspx
101
Appendix C
Matlab CNT Converter
% loadcnt() - Load a Neuroscan signal file.

% Usage:
% >> cnt = loadcnt(file, varargin)
%
% Inputs:
% filename - name of the file with extension
% Optional inputs:
% ’t1’ - start at time t1, default 0
% ’sample1’ - start at sample1, default 0, overrides t1
% ’lddur’ - duration of segment to load,
% ’ldnsamples’ - number of samples to load,
% ’scale’ - [’on’|’off’] scale data to microvolt
% ’dataformat’ - [’int16’|’int32’]
% ’blockread’ - [integer] by default it is automatically
% determined from the file header, though
% sometimes it finds an incorect value
%
102
Appendix C. Matlab CNT Converter
% Outputs:
% cnt - structure with the continuous data
% cnt.header
% cnt.electloc
% cnt.data
% cnt.tag
%
% Authors: Sean Fitzgibbon, Arnaud Delorme, 2000-
%
% Note: function original name was load_scan41.m
% Copyright (C) 2000 Sean Fitzgibbon

% Copyright (C) 2003 Arnaud Delorme, Salk Institute
%
% This program is free software;
& you can redistribute it and/or modify
% it under the terms of the GNU General Public License
% as published by the Free Software Foundation;
%
% This program is distributed in the hope that
% it will be useful, but WITHOUT ANY WARRANTY;
%
% You should have received a copy of
% the GNU General Public License
% along with this program;
% if not, write to the Free Software
% Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
103
% $Log: loadcnt.m,v $
% Revision 1.21 2005/08/16 22:46:55 arno
% allowing to read event type 3
%
% Revision 1.20 2005/05/12 15:50:37 arno
% keypad modified by Andreas
%
% Revision 1.19 2004/11/23 17:08:57 hilit
% fixed a typo
%
% Revision 1.18 2004/09/14 23:31:57 arno
% dataformat
%
% Revision 1.17 2004/09/14 23:27:44 arno
% opening file as little endian
%
% Revision 1.16 2004/03/19 18:52:42 arno
% blockread msg
%
% Revision 1.15 2004/03/19 18:51:26 arno
% allowing blockread option
%
% Revision 1.14 2003/11/05 16:38:08 arno
% reading events for 32-bit data
%
% Revision 1.13 2003/10/30 19:41:01 arno
104
% updating error message

%
% Revision 1.12 2003/10/30 19:23:54 arno
% adding revision line
%
function [f,lab,ev2p] = loadcnt(filename,varargin)
if ~isempty(varargin)
r=struct(varargin{:});
else r = [];
end;
try, r.t1; catch, r.t1=0; end

try, r.sample1; catch, r.sample1=[]; end
try, r.lddur; catch, r.lddur=[]; end
try, r.ldnsamples; catch, r.ldnsamples=[]; end
try, r.scale; catch, r.scale=’on’; end
try, r.blockread; catch, r.blockread = []; end
try, r.dataformat; catch, r.dataformat = ’int16’; end
sizeEvent1 = 8 ; %%% 8 bytes for Event1

type=’cnt’;
105
if nargin ==1
scan=0;
end
fid = fopen(filename,’r’, ’l’);

disp([’Loading file ’ filename ’ ...’])
h.rev = fread(fid,12,’char’);
h.nextfile = fread(fid,1,’long’);
h.prevfile = fread(fid,1,’ulong’);
h.type = fread(fid,1,’char’);
h.id = fread(fid,20,’char’);
h.oper = fread(fid,20,’char’);
h.doctor = fread(fid,20,’char’);
h.referral = fread(fid,20,’char’);
h.hospital = fread(fid,20,’char’);
h.patient = fread(fid,20,’char’);
h.age = fread(fid,1,’short’);
h.sex = fread(fid,1,’char’);
h.hand = fread(fid,1,’char’);
h.med = fread(fid,20, ’char’);
h.category = fread(fid,20, ’char’);
h.state = fread(fid,20, ’char’);
h.label = fread(fid,20, ’char’);
h.date = fread(fid,10, ’char’);
h.time = fread(fid,12, ’char’);
h.mean_age = fread(fid,1,’float’);
106
h.stdev = fread(fid,1,’float’);
h.n = fread(fid,1,’short’);
h.compfile = fread(fid,38,’char’);
h.spectwincomp = fread(fid,1,’float’);
h.meanaccuracy = fread(fid,1,’float’);
h.meanlatency = fread(fid,1,’float’);
h.sortfile = fread(fid,46,’char’);
h.numevents = fread(fid,1,’int’);
h.compoper = fread(fid,1,’char’);
h.avgmode = fread(fid,1,’char’);
h.review = fread(fid,1,’char’);
h.nsweeps = fread(fid,1,’ushort’);
h.compsweeps = fread(fid,1,’ushort’);
h.acceptcnt = fread(fid,1,’ushort’);
h.rejectcnt = fread(fid,1,’ushort’);
h.pnts = fread(fid,1,’ushort’);
h.nchannels = fread(fid,1,’ushort’);
h.avgupdate = fread(fid,1,’ushort’);
h.domain = fread(fid,1,’char’);
h.variance = fread(fid,1,’char’);
h.rate = fread(fid,1,’ushort’);
h.scale = fread(fid,1,’double’);
h.veogcorrect = fread(fid,1,’char’);
h.heogcorrect = fread(fid,1,’char’);
h.aux1correct = fread(fid,1,’char’);
h.aux2correct = fread(fid,1,’char’);
h.veogtrig = fread(fid,1,’float’);
107
h.heogtrig = fread(fid,1,’float’);
h.aux1trig = fread(fid,1,’float’);
h.aux2trig = fread(fid,1,’float’);
h.heogchnl = fread(fid,1,’short’);
h.veogchnl = fread(fid,1,’short’);
h.aux1chnl = fread(fid,1,’short’);
h.aux2chnl = fread(fid,1,’short’);
h.veogdir = fread(fid,1,’char’);
h.heogdir = fread(fid,1,’char’);
h.aux1dir = fread(fid,1,’char’);
h.aux2dir = fread(fid,1,’char’);
h.veog_n = fread(fid,1,’short’);
h.heog_n = fread(fid,1,’short’);
h.aux1_n = fread(fid,1,’short’);
h.aux2_n = fread(fid,1,’short’);
h.veogmaxcnt = fread(fid,1,’short’);
h.heogmaxcnt = fread(fid,1,’short’);
h.aux1maxcnt = fread(fid,1,’short’);
h.aux2maxcnt = fread(fid,1,’short’);
h.veogmethod = fread(fid,1,’char’);
h.heogmethod = fread(fid,1,’char’);
h.aux1method = fread(fid,1,’char’);
h.aux2method = fread(fid,1,’char’);
h.ampsensitivity = fread(fid,1,’float’);
h.lowpass = fread(fid,1,’char’);
h.highpass = fread(fid,1,’char’);
h.notch = fread(fid,1,’char’);
108
h.autoclipadd = fread(fid,1,’char’);
h.baseline = fread(fid,1,’char’);
h.offstart = fread(fid,1,’float’);
h.offstop = fread(fid,1,’float’);
h.reject = fread(fid,1,’char’);
h.rejstart = fread(fid,1,’float’);
h.rejstop = fread(fid,1,’float’);
h.rejmin = fread(fid,1,’float’);
h.rejmax = fread(fid,1,’float’);
h.trigtype = fread(fid,1,’char’);
h.trigval = fread(fid,1,’float’);
h.trigchnl = fread(fid,1,’char’);
h.trigmask = fread(fid,1,’short’);
h.trigisi = fread(fid,1,’float’);
h.trigmin = fread(fid,1,’float’);
h.trigmax = fread(fid,1,’float’);
h.trigdir = fread(fid,1,’char’);
h.autoscale = fread(fid,1,’char’);
h.n2 = fread(fid,1,’short’);
h.dir = fread(fid,1,’char’);
h.dispmin = fread(fid,1,’float’);
h.dispmax = fread(fid,1,’float’);
h.xmin = fread(fid,1,’float’);
h.xmax = fread(fid,1,’float’);
h.automin = fread(fid,1,’float’);
h.automax = fread(fid,1,’float’);
h.zmin = fread(fid,1,’float’);
109
h.zmax = fread(fid,1,’float’);
h.lowcut = fread(fid,1,’float’);
h.highcut = fread(fid,1,’float’);
h.common = fread(fid,1,’char’);
h.savemode = fread(fid,1,’char’);
h.manmode = fread(fid,1,’char’);
h.ref = fread(fid,10,’char’);
h.rectify = fread(fid,1,’char’);
h.displayxmin = fread(fid,1,’float’);
h.displayxmax = fread(fid,1,’float’);
h.phase = fread(fid,1,’char’);
h.screen = fread(fid,16,’char’);
h.calmode = fread(fid,1,’short’);
h.calmethod = fread(fid,1,’short’);
h.calupdate = fread(fid,1,’short’);
h.calbaseline = fread(fid,1,’short’);
h.calsweeps = fread(fid,1,’short’);
h.calattenuator = fread(fid,1,’float’);
h.calpulsevolt = fread(fid,1,’float’);
h.calpulsestart = fread(fid,1,’float’);
h.calpulsestop = fread(fid,1,’float’);
h.calfreq = fread(fid,1,’float’);
h.taskfile = fread(fid,34,’char’);
h.seqfile = fread(fid,34,’char’);
h.spectmethod = fread(fid,1,’char’);
h.spectscaling = fread(fid,1,’char’);
h.spectwindow = fread(fid,1,’char’);
110
h.spectwinlength = fread(fid,1,’float’);
h.spectorder = fread(fid,1,’char’);
h.notchfilter = fread(fid,1,’char’);
h.headgain = fread(fid,1,’short’);
h.additionalfiles = fread(fid,1,’int’);
h.unused = fread(fid,5,’char’);
h.fspstopmethod = fread(fid,1,’short’);
h.fspstopmode = fread(fid,1,’short’);
h.fspfvalue = fread(fid,1,’float’);
h.fsppoint = fread(fid,1,’short’);
h.fspblocksize = fread(fid,1,’short’);
h.fspp1 = fread(fid,1,’ushort’);
h.fspp2 = fread(fid,1,’ushort’);
h.fspalpha = fread(fid,1,’float’);
h.fspnoise = fread(fid,1,’float’);
h.fspv1 = fread(fid,1,’short’);
h.montage = fread(fid,40,’char’);
h.eventfile = fread(fid,40,’char’);
h.fratio = fread(fid,1,’float’);
h.minor_rev = fread(fid,1,’char’);
h.eegupdate = fread(fid,1,’short’);
h.compressed = fread(fid,1,’char’);
h.xscale = fread(fid,1,’float’);
h.yscale = fread(fid,1,’float’);
h.xsize = fread(fid,1,’float’);
h.ysize = fread(fid,1,’float’);
h.acmode = fread(fid,1,’char’);
111
h.commonchnl = fread(fid,1,’uchar’);
h.xtics = fread(fid,1,’char’);
h.xrange = fread(fid,1,’char’);
h.ytics = fread(fid,1,’char’);
h.yrange = fread(fid,1,’char’);
h.xscalevalue = fread(fid,1,’float’);
h.xscaleinterval = fread(fid,1,’float’);
h.yscalevalue = fread(fid,1,’float’);
h.yscaleinterval = fread(fid,1,’float’);
h.scaletoolx1 = fread(fid,1,’float’);
h.scaletooly1 = fread(fid,1,’float’);
h.scaletoolx2 = fread(fid,1,’float’);
h.scaletooly2 = fread(fid,1,’float’);
h.port = fread(fid,1,’short’);
h.numsamples = fread(fid,1,’ulong’);
h.filterflag = fread(fid,1,’char’);
h.lowcutoff = fread(fid,1,’float’);
h.lowpoles = fread(fid,1,’short’);
h.highcutoff = fread(fid,1,’float’);
h.highpoles = fread(fid,1,’short’);
h.filtertype = fread(fid,1,’char’);
h.filterdomain = fread(fid,1,’char’);
h.snrflag = fread(fid,1,’char’);
h.coherenceflag = fread(fid,1,’char’);
h.continuoustype = fread(fid,1,’char’);
h.eventtablepos = fread(fid,1,’ulong’);
h.continuousseconds = fread(fid,1,’float’);
112
h.channeloffset = fread(fid,1,’long’);
h.autocorrectflag = fread(fid,1,’char’);
h.dcthreshold = fread(fid,1,’uchar’);
for n = 1:h.nchannels
e(n).lab = fread(fid,10,’char’);
e(n).reference = fread(fid,1,’char’);
e(n).skip = fread(fid,1,’char’);
e(n).reject = fread(fid,1,’char’);
e(n).display = fread(fid,1,’char’);
e(n).bad = fread(fid,1,’char’);
e(n).n = fread(fid,1,’ushort’);
e(n).avg_reference = fread(fid,1,’char’);
e(n).clipadd = fread(fid,1,’char’);
e(n).x_coord = fread(fid,1,’float’);
e(n).y_coord = fread(fid,1,’float’);
e(n).veog_wt = fread(fid,1,’float’);
e(n).veog_std = fread(fid,1,’float’);
e(n).snr = fread(fid,1,’float’);
e(n).heog_wt = fread(fid,1,’float’);
e(n).heog_std = fread(fid,1,’float’);
e(n).baseline = fread(fid,1,’short’);
e(n).filtered = fread(fid,1,’char’);
e(n).fsp = fread(fid,1,’char’);
e(n).aux1_wt = fread(fid,1,’float’);
e(n).aux1_std = fread(fid,1,’float’);
e(n).senstivity = fread(fid,1,’float’);
113
e(n).gain = fread(fid,1,’char’);
e(n).hipass = fread(fid,1,’char’);
e(n).lopass = fread(fid,1,’char’);
e(n).page = fread(fid,1,’uchar’);
e(n).size = fread(fid,1,’uchar’);
e(n).impedance = fread(fid,1,’uchar’);
e(n).physicalchnl = fread(fid,1,’uchar’);
e(n).rectify = fread(fid,1,’char’);
e(n).calib = fread(fid,1,’float’);
end
% finding if 32-bits of 16-bits file

% ----------------------------------
begdata = ftell(fid);
enddata = h.eventtablepos; % after data
if strcmpi(r.dataformat, ’int16’)
nums = (enddata-begdata)/h.nchannels/2;
else nums = (enddata-begdata)/h.nchannels/4;
end;
% number of sample to read

% ------------------------
if ~isempty(r.sample1)
r.t1 = r.sample1/h.rate;
else
r.sample1 = r.t1*h.rate;
end;
114
startpos = r.t1*h.rate*2*h.nchannels;
else startpos = r.t1*h.rate*4*h.nchannels;
end;
if isempty(r.ldnsamples)
if ~isempty(r.lddur)
r.ldnsamples = round(r.lddur*h.rate);
else r.ldnsamples = nums;
end;
end;
% channel offset
% --------------
if ~isempty(r.blockread)
h.channeloffset = r.blockread;
end;
if h.channeloffset > 1
fprintf(’WARNING: reading data in blocks of %d,
if this fails, try using option "’’blockread’’, 1"\n’);
end;
disp(’Reading data .....’)

if type == ’cnt’
% while (ftell(fid) +1 < h.eventtablepos)

%d(:,i)=fread(fid,h.nchannels,’int16’);
%end
115
fseek(fid, startpos, 0);

if h.channeloffset <= 1
dat=fread(fid, [h.nchannels r.ldnsamples], r.dataformat);
else
h.channeloffset = h.channeloffset/2;
% reading data in blocks
dat = zeros( h.nchannels, r.ldnsamples);
dat(:, 1:h.channeloffset) =
fread(fid, [h.channeloffset h.nchannels])’;
counter = 1;
while counter*h.channeloffset < r.ldnsamples
dat(:, counter*h.channeloffset+1:
counter*h.channeloffset) = ...
fread(fid, [h.channeloffset h.nchannels]
, r.dataformat)’;
counter = counter + 1;
end;
end;
%ftell(fid)
if strcmpi(r.scale, ’on’)
disp(’Scaling data .....’)
%%% scaling to microvolts
for i=1:h.nchannels
bas=e(i).baseline;
sen=e(i).senstivity;cal=e(i).calib;
116
mf=sen*(cal/204.8);
dat(i,:)=(dat(i,:)-bas).*mf;
end
end
ET_offset = (double(h.prevfile) * (2^32))

+ double(h.eventtablepos);
fseek(fid, ET_offset, ’bof’);
disp(’Reading Event Table...’)

eT.teeg = fread(fid,1,’uchar’);
eT.size = fread(fid,1,’ulong’);
eT.offset = fread(fid,1,’ulong’);
if eT.teeg==2
nevents=eT.size/sizeEvent2;
if nevents > 0
ev2(nevents).stimtype = [];
for i=1:nevents
ev2(i).stimtype = fread(fid,1,’ushort’);
ev2(i).keyboard = fread(fid,1,’char’);
temp = fread(fid,1,’uint8’);
ev2(i).keypad_accept = bitand(15,temp);
ev2(i).accept_ev1 = bitshift(temp,-4);
ev2(i).offset = fread(fid,1,’long’);
ev2(i).type = fread(fid,1,’short’);
ev2(i).code = fread(fid,1,’short’);
117
ev2(i).latency = fread(fid,1,’float’);
ev2(i).epochevent = fread(fid,1,’char’);
ev2(i).accept = fread(fid,1,’char’);
ev2(i).accuracy = fread(fid,1,’char’);
end
else
ev2 = [];
end;
elseif eT.teeg==3
if nevents > 0
if r.dataformat == ’int32’
bytes_per_samp = 4;
else
bytes_per_samp = 2;
end
for i=1:nevents
os = fread(fid,1,’ulong’);
ev2(i).offset = os * bytes_per_samp * h.nchannels;
ev2(i).type = fread(fid,1,’short’);
ev2(i).code = fread(fid,1,’short’);
118
ev2(i).latency = fread(fid,1,’float’);
ev2(i).epochevent = fread(fid,1,’char’);
ev2(i).accept = fread(fid,1,’char’);
ev2(i).accuracy = fread(fid,1,’char’);
end
else
ev2 = [];
end;
elseif eT.teeg==1
if nevents > 0
for i=1:nevents
% modified by Andreas Widmann 2005/05/12 14:15:00

%ev2(i).keypad_accept = fread(fid,1,’char’);
% end modification
ev2(i).offset = fread(fid,1,’long’);
end;
else
ev2 = [];
119
end;
else
disp(’Skipping event table (tag != 1,2,3 ;
theoritically impossible)’);
ev2 = [];
end
end
fseek(fid, -1, ’eof’);

t = fread(fid,’char’);
f.header = h;
f.electloc = e;
f.data = dat;
f.Teeg = eT;
f.event = ev2;
f.tag=t;
%%%% channels labels

for i=1:h.nchannels
plab=sprintf(’%c’,f.electloc(i).lab);
if i>1
lab=str2mat(lab,plab);
else
lab=plab;
end
end
120
%%%% to change offest in bytes to points

if ~isempty(ev2)
ev2p=ev2;
ioff=900+(h.nchannels*75);
ev2p(i).offset=(ev2p(i).offset-ioff)/
(2*h.nchannels) - r.sample1;
end
else % 32 bits
for i=1:nevents
ev2p(i).offset=(ev2p(i).offset-ioff)/
(4*h.nchannels) - r.sample1;
end
end;
f.event = ev2p;
end;
frewind(fid);
fclose(fid);
121
Appendix D
Extended-Infomax ICA
Algorithm
This toolbox performs Independent Component Analysis (ICA) decomposi-
tion of input data using the logistic infomax ICA algorithm of Bell Sejnowski
(1995) with the natural gradient feature of Amari, Cichocki and Yang, or
optionally the extended-ICA algorithm of Lee, Girolami and Sejnowski, with
optional PCA dimension reduction. Annealing based on weight changes is
used to automate the separation process.
Authors:
Scott Makeig with contributions from Tony Bell, Te-Won Lee, Tzyy-Ping
Jung, Sigurd Enghoff, Michael Zibulevsky, Delorme Arnaud, CNL/The Salk
Institute, La Jolla, 1996
Reference:
Makeig, S., Bell, A.J., Jung, T-P and Sejnowski, T.J., ”’Independent com-
ponent analysis of electroencephalographic data”’, Information Processing
122
Appendix D. Extended-Infomax ICA Algorithm
Systems 8:145-151, MIT Press, Cambridge, MA (1996)
Toolbox Citation:
Makeig, Scott et al. ”EEGLAB: ICA Toolbox for Psychophysiological Re-
search”, Swartz Center for Computational Neuroscience, Institute of Neural
Computation, University of San Diego California, 2000.
For more information:
http://www.sccn.ucsd.edu/eeglab
http://www.sccn.ucsd.edu/eeglab/icafaq.html
http://www.cnl.salk.edu/ tony/ica.html
Copyright (C) 1996 Scott Makeig et al, SCCN/INC/UCSD
This program is free software; and can be redistributed and/or modified
under the terms of the GNU General Public License as published by the
Free Software Foundation.
123

Ubc 2012 Fall Adib Mani

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ubc 2012 Fall Adib Mani

Uploaded by

Copyright:

Available Formats

A New Method for Removing

Electrical Vestibular Stimulation

B.Sc., Sharif University of Technology

M.Sc., Sharif University of Technology

A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF

THE REQUIREMENTS FOR THE DEGREE OF

MASTER OF APPLIED SCIENCE

The Faculty of Graduate Studies

(Electrical and Computer Engineering)

THE UNIVERSITY OF BRITISH COLUMBIA

Electro-Encephalography (EEG) signals recorded during Electrical Vestibu-

functions can be altered during the EVS application.

tool in physiology and neurology. Various applications of EVS have been

implemented in the health-care industry and also in other industries such as

cations, it still remains a challenging problem to understand how the EVS

distribution of the EVS current throughout the scalp generates an artifact

during EVS, we have to eliminate this artifact.

We developed a method to remove this artifact by estimating the contri-

posed method is a hybrid method, which combines time series regression

and wavelet decomposition methods to estimate the artifact and remove

estimate the EVS current distribution in each frequency band separately.

We optimized the proposed method using simulated data. Then we assessed

others, in terms of achieving higher signal to artifact ratio and introducing

less distortion to the original EEG signals.

• In this project, all the experimental EVS/EEG data was collect by

our collaborators in the Paciﬁc Parkinson’s Research Centre and the

Sensorimotor Physiology Laboratory.

• The novel method proposed in this study, was submitted as a journal

• All experimental procedures in this work were reviewed and approved

by the University of British Columbia Clinical Research Ethics Board

under Certiﬁcate Number H09-02016.

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

1.2 Research Motivation and Contribution . . . . . . . . . . . . 3

1.3 Future Research Goals and Potential Innovations . . . . . . . 5

2 EVS Brain Interaction . . . . . . . . . . . . . . . . . . . . . . . 7

3 EVS Studies Problem . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . 17

4 Experimental Set-up and Data Acquisition . . . . . . . . . 23

4.1 EVS Instrumentation and Set-up . . . . . . . . . . . . . . . . 23

4.2 EEG Acquisition Set-up . . . . . . . . . . . . . . . . . . . . . 26

5 A Review of Artifact Removal Methods and Their Mathe-

5.2 ICA-based Artifact Removal Methods . . . . . . . . . . . . . 33

5.3 Regression-based Artifact Removal Methods . . . . . . . . . 36

5.4 Adaptive Filter-based Artifact Removal Methods . . . . . . . 39

5.5 Wavelet-based Artifact Removal Methods . . . . . . . . . . . 42

6 A Novel Hybrid Artifact Removal Method for EVS Arti-

6.1 Design a Hybrid Wavelet-based Artifact Removal Method . . 46

6.2 Impedance Analysis and Simulation Study for Method Opti-

6.3 Optimization of The Proposed Method Using Simulated Data 58

7 Study Results and Assessment . . . . . . . . . . . . . . . . . 63

7.1 Assessment of The Artifact Removal Methods in Stochastic

7.2 Assessment of The Proposed Artifact Removal Method in

Pulsed GVS Studies . . . . . . . . . . . . . . . . . . . . . . . 79

8 Conclusion and Future Work . . . . . . . . . . . . . . . . . . 85

8.1 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . 85

8.2 Areas of Future Studies . . . . . . . . . . . . . . . . . . . . . 88

A DS5 Isolated Bipolar Constant Current Stimulator . . . . 97

B Model 2200 Analog Stimulus Isolator . . . . . . . . . . . . . 100

C Matlab CNT Converter . . . . . . . . . . . . . . . . . . . . . . 102

D Extended-Infomax ICA Algorithm . . . . . . . . . . . . . . . 122

5.1 Frequency bands for approximation and details components

of the measured EEG when sampling rate is 1KHz . . . . . . 45

6.1 Correlation and normalized residual sum of squares between