2746 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO.
11, NOVEMBER 2010
Quantification of Motion Artifact Rejection Due
to Active Electrodes and Driven-Right-Leg
Circuit in Spike Detection Algorithms
Antoine Nonclercq*, Member, IEEE, and Pierre Mathys
Abstract—Identification of spikes in the EEG plays an impor-
tant role during the diagnosis of neurological disorders, such as
epilepsy. Automatic spike detection (ASD) is attractive because it
reduces the diagnostic time and improves objectivity of the scor-
ing. Unfortunately, automatic detection is sometimes confounded
by artifacts, particularly motion artifacts, which can be frequent
in ambulatory recording, in the ICU, when recording from restless
patients or children, etc. EEG systems have recently been improved
by using active electrodes and driven-right-leg circuits (DRL) to
reduce motion artifacts. However, the performances of ASD algo-
rithms, both with unimproved and improved EEG systems, are
difficult to quantify in patients because of poor reproducibility of
the results. In this paper, a test setup was used to evaluate the
performance of active electrodes and DRL, and assess if they can
be complemented or substituted by a spike detection algorithm in
avoiding motion artifact. Results show that motion artifacts can
largely degrade spike detection when a traditional EEG system
is used, whereas an EEG fitted with active electrodes and a DRL
allows high-quality detection. When using a traditional EEG, the
choice of a spike detection algorithm has a large influence on de-
tection quality.
Index Terms—Amplifying electrodes, bioelectric recordings,
driven-right-leg (DRL) circuit, motion artifacts, spike detection.
I. INTRODUCTION
EG systems are a widely used clinical tool for diagnos-
E ing and monitoring neurological disorders. Its major ap-
plication is epilepsy, a disorder with an incidence of 47.4 per
100 000 [1], which is characterized by sudden recurrent and
transient disturbances of mental functions and/or movements of
the body that result from the excessive discharge of groups of
brain cells [2]. The presence of spikes on an interictal scalp EEG
is accepted as confirmation for the diagnosis of epilepsy [2], [3],
and therefore, interictal spike detection plays a crucial role in
diagnosis.
Spike detection is traditionally performed by an experienced
electroencephalographer and is based on visual inspection of the
Manuscript received August 9, 2009; revised December 19, 2009 and April
21, 2010; accepted May 28, 2010. Date of publication July 8, 2010; date of
current version October 15, 2010. Asterisk indicates corresponding author.
∗ A. Nonclercq is with the Bio-, Electro- and Mechanical Systems De-
partment, Université Libre de Bruxelles, Bruxelles B1050, Belgium (e-mail:
anoncler@ulb.ac.be).
P. Mathys is with the Bio-, Electro- and Mechanical Systems Depart-
ment, Université Libre de Bruxelles, Bruxelles B1050, Belgium (e-mail:
pmathys@ulb.ac.be).
Digital Object Identifier 10.1109/TBME.2010.2055867
0018-9294/$26.00 © 2010 IEEE
recording. Unfortunately, this method tends to be very time con-
suming, error prone, and too subjective [4]. Therefore, there is a
need for an efficient automated spike detection (ASD) method to
reduce the diagnostic time and improve the objectivity of scor-
ing. Furthermore, such an algorithm would allow studies related
to a quantitative description of spike density, topology, and mor-
phology, which would help to determine patient syndrome and
surgical outcome. The comprehensive spike marking required
for these types of studies is too time consuming for visual iden-
tification by electroencephalographers [5]. Hundreds of ASD
methods have been published (see [5]–[7] for review). While
reasonably accurate spike detection can be obtained with an
EEG free from artifacts [8], artifacts cause false positives, and
therefore, overestimate the number of spikes. This has long been
the plague of automatic systems [6], [9]. ASD algorithms are
still less accurate than expert electroencephalographers, which
limit their use and make artifact rejection the fundamental dif-
ficulty still facing today’s algorithms [5], [6].
Motion artifacts are especially problematic because they may
have a spike-like shape [10], [11]. They are not confined to a
small spectral band, so they cannot be removed by frequency
filtering. Motion artifacts can be frequent, especially in the case
of ambulatory recording, in the ICU—where patients are fre-
quently moving or being moved [12]—when recording of rest-
less patients or children, etc., and have, therefore, long been a
problem in biopotential measurements [13].
Motion artifacts are introduced into the EEG recording stream
by movement of the patient, of other people around the patient,
or of the electrode leads. The resulting geometric change al-
ters the magnetic and capacitive coupling of the patient and
the leads, which, in turn, alters the parasitic current flowing
into the leads [14], [15]. This current produces a parasitic volt-
age drop when flowing into the electrode/gel/skin interface,
which interferes with the EEG signal [16]. Patient movements
also cause impedance variation [17]–[20] and potential varia-
tion [20]–[23] at the electrode/gel/skin interface. However, the
use of Ag/AgCl electrodes and skin preparation minimizes both
electrode/gel/skin potential variation [13], [18], [21], [23] and
impedance variation [19], making the magnetic and capacitive
coupling effect dominant.
The use of active electrodes has been reported to greatly
reduce motion artifacts [17], [21], [24]. By having a built-in
amplifier, active electrodes have low-output impedance, pre-
venting the parasitic current flowing into the leads to produce
large voltage drops, thereby reducing the impact of artifacts on
the recording. Driven-right-leg (DRL) circuit [16], [25], [26] is
NONCLERCQ AND MATHYS: QUANTIFICATION OF MOTION ARTIFACT REJECTION DUE TO ACTIVE ELECTRODES 2747

often used jointly with the active electrodes and helps minimiz-
ing the common-mode voltage.
Publications about motion artifacts in the EEG [13] and their
possible reduction by active electrodes [17], [21], [24] describe
measurements on patients. Such an approach is limited because
of normal variability in the EEG signal among different record-
ings, in frequency, shape and strength of artifact, in patient
stoutness, as well as in geometrical arrangement of patient,
electrodes, leads, and EEG system used. Because the recording Fig. 1. Top view of setup.
conditions are not reproducible, it is not possible to compare dif-
ferent types of system and of spikes detection algorithms with
respect to motion artifacts. Since the original EEG signal of
each patient is unknown, it is impossible to quantify the impact
of the artifact on the recording and, a fortiori, on the detection.
In this way, a quantitative study of the diminution of the effect
of motion artifact is not feasible in studies on patients.
It was to address this lack that an EEG setup was designed
and built to characterize the performance of EEG systems and
spike detection algorithms. It comprises a generator that pro-
duces cerebral-like waves, a head with adjustable motion, and
hence, adjustable artifacts strength, a electrode/gel/skin inter-
face model, electrodes, and leads. These are used with the EEG Fig. 2. Spike and waves comparison. (Above) output of the electroencephalo-
system and the spike detection algorithm to be tested. Com- graphic signals generator. (Below) epileptic patient.
pared to tests on patients, this has the major advantages of being
able to generate reproducible EEG-like signals and motion arti-
facts on demand, while excluding other kinds of artifacts. The
5) the EEG system, which amplifies, filters, and digitizes the
EEG signal at the output of the amplifier, when contaminated
signal;
by motion artifact, can be compared to the unattenuated input
6) the spike detection algorithm.
signal from the generator. While simulations could be used to
The setup was placed on a bench with a mains electrical cable
virtually realize, such a test bench, an evaluation using data col-
running in parallel to it, at a distance of 1 m from each module.
lected directly from a physical system is advantageous because
This section describes these modules, as well as the method
the results more closely reflect performance, since they include
used for recording on the physical model and on the human
the factors previously mentioned that are difficult to accurately
subjects.
model.
Overall spike detection accuracy depends on the ability of
both the EEG system and the spike detection algorithm to pre-
vent or reject artifacts. Using the proposed setup, the acquisition A. EEG Signal Generator
chain—from a generated EEG-like signal on the dummy head The generator is able to produce an arbitrary waveform, built
to detected spikes on the recording—is considered as a whole. up of 1000 analog voltages samples per second, with a resolu-
In this paper, a traditional EEG system type and a system fitted tion of 15 nV and an overall noise level lower than 0.2 µVrms.
with active electrode and a DRL type were tested together with Performances and application examples of this EEG signal gen-
four spike detection algorithms. In this way, this paper helps erator have been previously published [27].
to evaluate the usefulness of active electrodes and a DRL, and In order to compare the different spike detection algorithms,
assess if they can be complemented or substituted by a spike the signal generator was set to produce spikes at a constant rate
detection algorithm in terms of motion artifact rejection. For of one spike per second on eight channels simultaneously. An
comparison, measurements were also performed on two human epileptic spike has a pointed peak and a duration of 20–70 ms [3].
subjects. Although it may occur alone, a spike is usually followed by a
slow wave, which lasts 150–350 ms, forming what is known as
a “spike and slow-wave complex” [8]. In this paper, spikes were
stylized by a 200 µV 50 ms triangle followed by a 200 µV 250 ms
II. METHODS half-sine wave, thus fitting Chatrian’s definition [3]. The period
The setup comprises (see Fig. 1) the following: of the spikes is 1000 ms. The similarity between the stylized
1) the EEG signal generator that produces cerebral-like spikes, and wave and real ones cannot be quantified, since real
waves; spikes and wave vary much between patients. However, Fig. 2
2) the moving head that generates motion artifacts; gives an illustration of the EEG signal generator output and
3) the electrode/gel/skin interface model; compares it with an example of spike and wave from an epileptic
4) the leads; patient.
2748 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 11, NOVEMBER 2010
Fig. 3. Head movement as a function of time.
B. Moving Head
The moving head generates reproducible motion artifacts by
rotating alternately clockwise and counterclockwise around a
vertical axis, simulating a patient shaking his head. A TRI-
NAMIC QSH4218 stepping motor with a 1.8◦ step angle and
a 49 N·cm holding torque was driven by a NANOTEC SMC42
driver at a 1/8 step, thereby achieving a 0.225◦ angular resolu-
tion. A microcontroller was programmed to produce sinusoidal
motion (see Fig. 3).
The period of rotation was about T = 980 ms, so the time of the
movement artifacts relative to the spikes continually changes.
Depending on the artifact strength wanted, αm ax can be set
to any angle from 0◦ (i.e., the head does not move) to 90◦
(i.e., the head moves from −45◦ to +45◦ ). The mean angular
velocity is equal to 2αm ax /T, up to ≈180◦ /s. In the following,
the artifact strength of a given recording will be quantified in
terms of mean angular velocity. No electromagnetic interference
or artifact other than motion artifacts was observed.
C. Electrode/Gel/Skin Interface
To model the electrode/gel/skin interface, a 20 kΩ resistor
(corresponding to a typical electrode/gel/skin impedance value
[16]) was placed in series with two Ag/AgCl electrodes back-
to-back.
Eight electrodes were placed on the head (FP1, C3, O1, FP2,
C4, O2, T3, and T4 applied according to the standard 10/20
system [28] of electrode placement). Two additional electrodes
were placed: the mastoid electrode (posterior to the right ear)
and the reference (or ground) electrode (on the forehead).
D. EEG and Leads
One meter leads were used to connect the dummy head to the
EEG systems. Two EEG systems were tested: one traditional
type, and the other fitted with active electrodes and a DRL
type. They have typical characteristics: an overall gain of 1000
(for the system with active electrodes, an amplification of 100
is achieved in the electrode itself), a 0.16–70 Hz passband, a
sampling frequency of 200 Hz, and a noise level lower than
0.8 µVrms.
E. Spike Detection Algorithms
Out of the large number of algorithms previously proposed
in the literature, four were chosen for this study. The stylized
spike series described earlier was chosen to train each algorithm
and to adjust each parameter. In this way, the study was focused
on the ability of the algorithm to reject motion artifacts and
not on spike shape variation. For each algorithm, a threshold
is used to determine the presence of a spike: the higher the
threshold, the more closely a pattern must correspond to the
stylized spike in order to be classified as a spike. The threshold
must be carefully chosen when artifacts are present. It must
be low enough to detect spikes altered by artifacts and high
enough to reject spike-like artifacts. The threshold showing the
best performances as well as the impact of a ±10% threshold
variation are discussed in the Section III.
The first algorithm chosen is based on artificial neural network
(ANN) and was previously proposed by Özdamar and Kalayci
[29]. It uses raw electroencephalographic data as an input. The
neural network structure and training of this study were the
same as proposed in the paper. It is widely known and was
previously reviewed by Pang et al. [7]. Out of the different
ANN architectures proposed, the one chosen was the “simplest
network with the best outcome”: 30 inputs, six hidden layer
neurons, and two outputs. However, since the second-output
neuron was used to detect electromyogram activity, it was out of
focus for this paper and, therefore, was left unused. To train the
ANN, portion of the stylized spike series were presented to the
ANN. A spike is detected if the output stage of the ANN exceeds
a given threshold. In this paper, it will be referred to as the “ANN
algorithm.”
The second algorithm chosen was proposed by Kalayci and
Özdamar [8] and was also previously reviewed by Pang et al. [7].
The neural network structure and training of this study were the
same as the proposed paper. It is also an ANN-based algorithm,
but uses wavelets as the input. Out of the set of algorithms,
the one chosen was the 16–8-1 feedforward, backpropagation
(BP) neural network using Daubechies-4 mother wavelet with
the central eight coefficients from scales 3 and 4 as the feature
vector, which corresponds to the algorithm reviewed by Pang
et al. [7]. A spike is detected if the output stage of the ANN
exceeds a given threshold. In this paper, it will be referred to as
the “ANN/wavelet algorithm.”
Wavelets have also been used independently of ANN. The
third algorithm is based on the wavelet transform proposed by
Kalayci and Özdamar [8], but without adding the ANN stage.
A spike is detected if the coefficients obtained by the algorithm,
at scales 3 and 4, are greater than preset thresholds. These are
expressed as a fraction of the coefficients obtained, for the styl-
ized spike, at the corresponding scales. In this way, the impact
of this additional ANN stage can be determined. In this paper,
it will be referred as the “wavelet algorithm.”
Finally, the fourth algorithm uses template matching. Tem-
plate matching is an old method that can serve as a “reference
algorithm” as it gives a direct quantification of the distortion due
to artifacts. A 30-sample window was chosen for the calcula-
tions, corresponding to the window length used by Özdamar and
NONCLERCQ AND MATHYS: QUANTIFICATION OF MOTION ARTIFACT REJECTION DUE TO ACTIVE ELECTRODES
Kalayci [29]. A spike is detected if the cross correlation between
the signal and the stylized spike is greater than a given thresh-
old and if the amplitude of the signal is greater than 50 µV.
In this paper, it will be referred as the “template matching
algorithm.”
F. Method Used for Recording on Human Subjects
Simulated electroencephalographic signals from the brain-
wave generator with head motion movement were verified by
comparing to signals obtained from two human subjects. The
same electrodes, gel and EEG systems were used both for the
physical model and the subjects. Prior to placing the electrodes,
the patient’s skin was cleaned and a conductive gel was applied
to the skin. Measurements were performed simultaneously with
both systems to minimize the variability of the measurements
described in Section I. To do so, the electrodes of both EEG
systems were placed close to each other (1 cm apart) so as to
measure the same brain activity. Separate reference electrodes
were used and also placed 1 cm apart. However, simultaneous
recording like this, does introduce several complications that
may reduce the accuracy.
1) The stray currents induced by each system will cause in-
terfering potentials for the other system.
2) The symmetrical arrangement described in Figs. 1 and
3 is not quite possible (the electrodes were placed close
together and arranged symmetrically for each system in
order to minimize this effect).
3) The leads of both systems clash during the motion (they
go in different directions from the head because the EEG
systems are placed in different locations).
4) Because the electrodes are not at the exact same place, the
movement artifacts must be slightly different.
The subjects were asked to move their head from side to side.
Since the head movement provided by the subjects could not be
as accurate as the physical model (0.225◦ ) only three angular
velocities were used (0, 90, and 180◦ /s) instead of the 13 used
for the physical model (0 to 180◦ /s with a 15◦ /s step). Only
1-min recordings were taken to avoid changes due to fatigue.
Ten sets of recordings were performed in order to compare the
error between recordings.
G. Method Used for Recording on Physical Model
The EEG systems were placed on the bench one after the
other, in order to avoid the effects 1) to 4) described earlier.
Position error for the electrode is thought to be less than 0.5 cm
(as reported by [30]) and the positioning of the head relative to
the reference axis, when α = 0, less than 5◦ .
In the first study, to compare the physical model with the
human subjects, four set of 1-min recordings were taken. Fur-
thermore, since the patient was not able to produce any EEG
pattern on demand, the output of the EEG signal generator was
set to zero. In the second study, the comparison of spike de-
tection algorithms, 3-min recordings were taken. In this way,
there were 8 channels × 180 s × 1 spike, and waves/second =
1440 spikes and waves per recording. The algorithms treat each
channel independently.
2749
Fig. 4. RMS output signal measured (µVrms) as a function of the mean
angular velocity (◦ /s) for the physical model (dotted line) and the patient
(plain line). The mean, maximum, and minimum values are displayed for each
curve.
III. RESULTS
A. Comparison Between the Physical Model and Human
Subjects
Fig. 4 compares the artifact recorded for the subjects and the
physical model. Degradation was slower for the EEG system
fitted with active electrodes and a DRL, and the level of artifact
was about three times lower than the traditional system with
a 180◦ /s mean angle velocity of the head. Fig. 4 also shows
that the level of artifact is higher for the human subjects, be-
cause of the background EEG, as well as physiological artifacts
(e.g., myoelectric), are also present. The background EEG is
the signal of about 10 µVrms present when mean velocity is
zero.
The standard deviation measured on the subjects equals
24.7 µVrms for a traditional EEG system and 9.8 µVrms for
the system fitted with active electrodes and DRL (averaged on
every level of artifact). It was lower for the physical model:
3.4 and 1.2 µVrms, respectively. This difference is due to the
variability related to measurements on subjects, as described in
Section I. The curves of Fig. 4 were approximated with a best-
fit line (least-squares error). Using the equation for this line,
the mean standard deviation of the mean angular velocity was
calculated from the mean standard deviation of the rms output
signal. For the physical model, it equals 7.2◦ /s for a traditional
EEG and 7.7◦ /s for an EEG fitted with active electrodes and
a DRL. The standard deviation accounts both for variation be-
tween sets of recordings (due to the geometrical arrangement)
and for the variation of movement artifact in time.√This sec-
ond component should be reduced by a factor of 3 in the
following study, since the recording durations are extended to
3 min.
B. Comparison Between Algorithms
For each algorithm, calculations were performed on each
sample of the recording. The threshold of each algorithm was
varied between 40% and 90% with a 1% step. Sensitivity, selec-
tivity, and correlation were calculated for each recording. For
the first study, the sensitivity, the selectivity, and the correlation
2750 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 11, NOVEMBER 2010
TABLE I
BEST MEAN CORRELATION
Fig. 5. Sensitivity, selectivity, and correlation of the spike detection algo-
rithms (shown with circles for the template matching algorithm, diamonds for
the wavelet algorithm, squares for the ANN algorithm, and crosses for the
ANN/wavelet algorithm) obtained as a function of mean angular velocity (◦ /s).
The results are shown for a (left) traditional system and (right) for the system
with active electrodes and DRL.
were calculated as follows:
TP
sensitivity = (1)
TP + FN
TP
selectivity = (2)
TP + FP

and 13.7% for the latter.
correlation = sensitivity × selectivity (3)
where TP is the number of true positives, FP is the number of
false positives, and FN is the number of false negatives.
The threshold showing the best mean correlation (averaged
over every level of artifact strength) is given in Table I. Sensi-
tivity, selectivity, and correlation corresponding to these thresh-
olds are given on Fig. 5. Almost-perfect spike detection (99%
or above on average, above 94% in the worst case) was obtained
with all spike detection algorithms for the system with active
electrodes and DRL. For the traditional system, the correlations
obtained were best with the template matching algorithm (94%
on average and 69% worst case), followed by the wavelet and the
ANN algorithms (both 88% on average and 54% worst case),
and finally, by the ANN/wavelet algorithm (81% on average
and 38% worst case). Fig. 5 shows sensitively, selectivity, and
correlation curves that are similar, illustrating that the best cor-
Fig. 6. Correlation of the four spike detection algorithms (shown with circles
for the system with active electrodes and DRL and diamonds for the traditional
system) obtained as a function of mean angular velocity (◦ /s). The maximum,
minimum, and mean values are displayed for each mean angular velocity (◦ /s).
relation is obtained near equal error rate (i.e., when sensitivity
equals selectivity).
To study the sensitivity of the algorithms to the threshold, a
±10% threshold variation was tested. However, for the last two
algorithms, since the two ANNs are tailored to give a ±80%
maximum output, sensitivity fell to zero when the threshold
exceeded a level close to 80% (79% for the ANN algorithm and
78% for the ANN/wavelet algorithm). Therefore, for these two
algorithms, the 20% threshold variation was shifted in order to
exclude these values (59%–79% for the ANN algorithm and
58%–78% for the ANN/wavelet algorithm). The effect of such
a threshold variation on the correlation of the four algorithms
is shown in Fig. 6. This impact was small for all algorithms
except the ANN/wavelet algorithm, and slightly higher for a
traditional system (mean correlation interval was 0.5% for the
system with active electrodes and DRL and rose to 2.3% for a
traditional system). It was higher in the case of the ANN/wavelet
algorithm (mean correlation interval was 4.4% for the system
with active electrodes and DRL and rose to 6.7% for a traditional
system) and reached a maximum value of 15.6% for the former
IV. DISCUSSION
This study demonstrates that motion artifacts can largely de-
grade spike detection quality when a traditional EEG system is
used (a correlation as low as 38% can be observed), whereas the
newer system with active electrodes and DRL allows good spike
detection, regardless of the spike detection algorithm used, even
for strong motion artifacts (correlation is always above 94%).
When the traditional system is used, the algorithms greatly dif-
fer in rejection of motion artifact, the template matching algo-
rithm being the most immune and the ANN/wavelet algorithm
the least. Threshold variation has a small effect, except for the
ANN/wavelet algorithm, where a 10% threshold variation can
reduce correlation by 15.6%.
When selecting a spike detection algorithm, the results pro-
vided in this study should be complemented by additional
NONCLERCQ AND MATHYS: QUANTIFICATION OF MOTION ARTIFACT REJECTION DUE TO ACTIVE ELECTRODES
reports that address other parameters, such as robustness to
other types of artifacts and the ability of the algorithm to in-
tegrate spike-shape variability. For example, a drawback of the
template matching algorithm is the difficulty in defining a set
of templates suitable for all cases [31]. Such a limitation does
not appear in this study, since the spike shape used is clearly de-
fined. Therefore, it seems beneficial to use more elaborate tem-
plate matching-based algorithms that can overcome this weak-
ness (e.g., [32]). Another interesting finding concerns the ANN
algorithm and the ANN/wavelet algorithm, which, in a previ-
ous comparison [7], yielded similar results. In this study, with
imposed motion artifacts, the ANN algorithm was the better
one.
This study also shows that more complex algorithms do not
necessarily yield better motion artifact rejection. However, a test
performed on a larger number of algorithms would be needed
to investigate the ways in which motion artifact vulnerability of
complex algorithms is related to elements that compose them.
Further work could also include the comparison of algorithms
belonging to other categories, as well as other types of hard-
ware setup, such as EEG systems fitted with dry electrodes.
It would also be interesting to study the ability of electroen-
cephalographers to distinguish the spikes from the artifacts and
to compare the results with the ones obtained for automated
algorithms. Finally, the proposed test bench could also be used
to study the strength of movement artifacts as a function of time
and electrode position. Such a study could place movement ar-
tifacts in relation to the underlying mechanisms that produce
them.
V. CONCLUSION
A physical model was used to quantify the performance of
spike detection algorithms with imposed motion artifacts, as
well as the reduction in motion artifacts by converting a tra-
ditional EEG system to one fitted with active electrodes and
a DRL. The physical model was compared with two human
subjects in an attempt to validate the hypotheses that artifact
degradation followed the same trends and that the reproducibil-
ity of the physical model was superior. The measurements tend
to confirm these hypotheses. However, an additional study with
a larger number of subjects is necessary to confirm the suggested
findings. The main conclusion is that high-detection quality (a
correlation higher than 94%) can be reached with the improved
EEG system, even when strong motion artifacts are present.
Such high correlation was never possible using a traditional
EEG system, regardless of the spike detection algorithm be-
ing used. If a traditional EEG system is used, the choice of a
spike detection algorithm is crucial: the four algorithms that
we tested showed large variations in their sensitivity to motion
artifacts.
This study also shows that more complex algorithms, such as
the ANN/wavelet algorithm, do not necessarily improve detec-
tion quality. Indeed, the ANN/wavelet algorithm is more vul-
nerable to motion artifacts and is more sensitive to threshold
variation than either of the two methods that it comprises.
2751
ACKNOWLEDGMENT
The authors would like to thank Prof. N. Donaldson from the
University College London for his comments and suggestions
on this paper.
REFERENCES
[1] I. A. W. Kotsopoulos, T. van Merode, F. G. H. Kessels, M. C. T. F.
M. de Krom, and J. A. Knottnerus, “Systematic review and meta-analysis
of incidence studies of epilepsy and unprovoked seizures,” Epilepsia,
vol. 43, pp. 1402–1409, May 2002.
[2] E. Niedermeyer and F. L. Da Silva, Electroencephalography Basic Prin-
ciples, Clinical Applications, and Related Fields. Baltimore, MD:
Williams and Wilkins, 1999.
[3] G. Chatrian, L. Bergamini, M. Dondey, D. Klass, M. Lennox-Buchthal,
and I. Petersen, “A glossary of terms most commonly used by clini-
cal electhroencephalographers,” Electroencephalogr. Clin. Neurophysiol.,
vol. 37, pp. 538–548, Nov. 1974.
[4] P. Y. Ktonas, Methods of Analysis of Brain Electrical and Magnetic Sig-
nals. Amsterdam, The Netherlands: Elsevier, 1987.
[5] S. B. Wilson and R. Emerson, “Spike detection: A review and comparison
of algorithms,” Clin. Neurophysiol., vol. 113, pp. 1873–1881, Dec. 2002.
[6] J. D. Frost, “Automatic recognition and characterization of epileptiform
discharges in the human EEG,” J. Clin. Neurophysiol., vol. 2, pp. 231–249,
Jul. 1985.
[7] C. C. Pang, A. R. M. Upton, G. Shine, and M. V. Kamath, “A comparison
of algorithms for detection of spikes in the electroencephalogram,” IEEE
Trans. Biomed. Eng., vol. 50, no. 4, pp. 521–526, Apr. 2003.
[8] T. Kalayci and Ö. Özdamar, “Wavelet preprocessing for the automatic
detection of EEG spikes with neural networks,” IEEE Eng. Med. Biol.
Mag., vol. 14, no. 2, pp. 160–166, Mar./Apr. 1995.
[9] J. R. Glover, N. Raghavan, P. Y. Ktonas, and J. D. Frost, “Context-based
automated detection of epileptogenic sharp transients in the EEG: Elim-
ination of false positives,” IEEE Trans. Biomed. Eng., vol. 36, no. 5,
pp. 519–527, May 1989.
[10] C. W. Bazil, B. A. Malow, and M. R. Sammaritano, Sleep and Epilepsy:
The Clinical Spectrum. Amsterdam, The Netherlands: Elsevier, 2002.
[11] J. Engel, T. A. Pedley, J. Aicardi, and M. A. Dichter, Epilepsy: A Com-
prehensive Textbook. Philadelphia, PA: Williams and Wilkins, 2008.
[12] G. B. Young, J. R. Ives, M. G. Chapman, and S. M. Mirsattari, “A compar-
ison of subdermal wire electrodes with collodion-applied disk electrodes
in long-term EEG recordings in ICU,” Clin. Neurophysiol., vol. 117,
pp. 1376–1379, Jun. 2006.
[13] H. W. Tam and J. G. Webster, “Minimizing electrode motion artifact by
skin abrasion,” IEEE Trans. Biomed. Eng., vol. 24, no. 2, pp. 134–139,
Mar. 1977.
[14] K. E. Misulis and T. C. Head, Essentials of Clinical Neurophysiology.
U.K.: Butterworth-Heinemann, 2003.
[15] T. L. Lee-Chiong, Sleep. Hoboken, NJ: Wiley, 2006.
[16] A. C. Metting van Rijn, A. Peper, and C. A. Grimbergen, “High-quality
recording of bioelectric events. Part 1. Interference reduction, theory and
practise,” Med. Biol. Eng. Comput., vol. 28, pp. 389–397, Sep. 1990.
[17] T. Thompson, T. Steffert, T. Ros, J. Leach, and J. Gruzelier, “EEG appli-
cations for sport and performance,” Methods, vol. 54, pp. 279–288, Aug.
2008.
[18] D. P. Burbank and J. G. Webster, “Reducing skin potential motion artefact
by skin abrasion,” Med. Biol. Eng. Comput., vol. 16, pp. 31–38, Jan. 1978.
[19] H. de Talhouet and J. G. Webster, “The origin of skin-stretch-caused
motion artifacts under electrodes,” Physiol. Meas., vol. 17, pp. 81–93,
May 1996.
[20] S. Odman, “Potential and impedance variations following skin deforma-
tion,” Med. Biol. Eng. Comput., vol. 19, pp. 271–278, May 1981.
[21] J. G. Webster, “Reducing motion artifacts and interference in biopotential
recording,” IEEE Trans. Biomed. Eng., vol. 31, no. 12, pp. 823–826, Dec.
1984.
[22] A. Searle and L. Kirkup, “A direct comparison of wet, dry and insulating
bioelectric recording electrodes,” Physiol. Meas., vol. 21, pp. 271–283,
May 2000.
[23] J. Webster, Medical Instrumentation. New York: Wiley, 1998.
[24] B. Hagemann, G. Luhede, and H. Luczak, “Improved “active” electrodes
for recording bioelectric signals in work physiology,” Eur. J. Appl. Phys-
iol., vol. 54, pp. 95–98, Dec. 1985.
[25] J. G. Webster and B. B. Winter, “Driven-right-leg circuit design,” IEEE
Trans. Biomed. Eng., vol. 30, no. 1, pp. 62–66, Jan. 1983.
2752 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 57, NO. 11, NOVEMBER 2010

[26] E. M. Spinelli, N. H. Martinez, and M. A. Mayosky, “A transconductance Pierre Mathys was born in Nivelles, Belgium, in
driven-right-leg circuit,” IEEE Trans. Biomed. Eng., vol. 46, no. 12, 1954. He received the M.Sc. degree in electrome-
pp. 1466–1470, Dec. 1999. chanical engineering and the Ph.D. degree in ap-
[27] A. Nonclercq and P. Mathys, “High resolution wave generator permits test plied sciences from the Université Libre de Bruxelles
and calibration of medical instrumentation devices,” presented at the 3rd (ULB), Bruxelles, Belgium, in 1977 and 1984,
Eur. Med. Biol. Conf.—IFMBE Eur. Conf. Biomed. Eng., Prague, 2005. respectively.
[28] H. H. Jasper, “The ten-twenty electrode system of the international fed- He is currently a Head of the Bio-, Electro- and
eration,” Electroencephalogr. Clin., Neurophysiol., vol. 10, pp. 370–375, Mechanical Systems Research Group, Applied Sci-
1958. ence Faculty, ULB, where he is engaged in research
[29] Ö. Özdamar and T. Kalayci, “Detection of spikes with artificial neural in digital electronics, real-time embedded systems,
networks using raw EEG,” Comput. Biomed. Res., vol. 31, pp. 122–142, power electronics, and industrial and biomedical
Apr. 1998. instrumentation.
[30] American Board of Registered Electroencephalographic Technologists,
Guidelines for the Oral Board in EEG, ABRET, Nashville, TN, 1984.
[31] H. S. Liu, T. Zhang, and F. S. Yang, “A multistage, multimethod approach
for automatic detection and classification of epileptiform EEG,” IEEE
Trans. Biomed. Eng., vol. 49, no. 12, pp. 1557–1566, Dec. 2002.
[32] A. Nonclercq, M. Foulon, D. Verheulpen, C. De Cock, M. Buzatu,
P. Mathys, and P. Van Bogaert, “Spike detection algorithm automatically
adapted to individual patients applied to spike and waves percentage quan-
tification,” Neurophysiol. Clin., vol. 39, pp. 123–131, Apr. 2009.

Antoine Nonclercq (M’07) was born in Brussels,

Belgium, in 1978. He received the M.Sc. degree in
electrical engineering from the Université Libre de
Bruxelles (ULB), Bruxelles, Belgium, the M.Sc. de-
gree in control and electrical engineering from the
Universidad Politécnica de Madrid, Madrid, Spain,
in 2002, and the Ph.D. degree in applied sciences
from the ULB, in 2007.
He is currently a Development Engineer at Med-
ical Data Technology Company, Brussels, Belgium,
a Researcher at the Implanted Devices Group, Uni-
versity College London, London, U.K., and an Industrial Partner at the ULB.
His research interests include the design of medical monitoring equipment and
biomedical signal processing.