Int conf 03

Abstract– Use of scalp EEG for the diagnosis of various
cerebral disorders is progressively increasing. Though the
advanced neuroimaging techniques such as MRI and CT-
SCAN still stay as principal confirmative methods for detecting
and localizing brain tumors, the development of automated
systems for the detection of brain tumors using the scalp EEG
has started attracting the researchers all over the world
notably since 2000. This is because of two important facts: (i)
cheapness and easiness of methods of recording and analyzing
the scalp EEG and (ii) lower risk and possible early detection.
This paper presents a method of detecting the brain tumor
using the first, second and third order statistics of the scalp
EEG with a Modified Wavelet-Independent Component
Analysis (MwICA) technique and a multi-layer feed-forward
neural network.
I. INTRODUCTION
IAGNOSIS and following (early) treatment are either
missed or delayed in 69% of the brain tumor cases due
to the fact that the most of the brain tumor symptoms are
highly misleading according to the survey [12]. The
advanced neuroimaging techniques such as MRI and CT or
biopsy are not immediately suggested due to the following
facts: they are either costly or invasive or do involve risks
like hazardous radiation, especially in case of children,
pregnant women and patients with implant devices [15]. The
delay in diagnosis worsens the outcome [14]. Hence a better
method that does not involve much cost, risks or complexity
is required to detect the presence of a brain tumor (structural
pathology) at an early stage [14].
II. EEG IN BRAIN TUMOR
Generally it is accepted that brain tumors on superficially
accessible portions of cerebral hemispheres involve some
localized loss of electrical activity causing some localized
slow waves on the scalp EEG [1]-[9] [16]. The general
findings on the brain tumor symptoms on EEG are [2] [6]
[9]: Polymorphic delta activity (PDA), Intermittent rhythmic
Manuscript received March 18, 2011.
V. Salai Selvam is with the Sriram Engineering College, Chennai, India.
Phone: +914427689029; fax: +914427689364; e-mail:
vsalaiselvam@yahoo.com.
S. Shenbadevi is with the Anna University, Chennai, Tamil Nadu, India.
E-mail: s_s_devi@annauniv.edu
delta activity (IRDA), Diffuse or localized theta activity,
Localized loss of activity over the area of the tumor,
Asymmetric beta activity, Disturbance of the alpha rhythm
and Spikes, sharp waves, or spike-wave discharges.
Reactivity and persistence of these abnormalities often are
the best indicators of the degree of damage: continuous slow
activity (e.g., persistent PDA) indicates severe structural
pathology such as large, deep hemispheric lesions whereas
intermittent slow activity (e.g., frontal IRDA) generally
indicates small lesions [16].
III. EARLIER WORKS ON BRAIN TUMOR DETECTION USING
EEG
Noteworthy earlier works on the detection of brain tumor
using scalp EEG are [1] [2] [3] [4] and [5]. In [4] it has been
shown how the one- and two- dimensional minimum orders
of non-linear Markov models, which approximate the
structure of the hidden dynamics in the EEG time-series of
the pair of channels F3 and F4, vary with respect to the age
and the structural pathologies (the tumors). In [2] it has been
shown that a multilayer Self-Organizing Map (SOM) trained
with the wavelet and frequency features can be used to
classify the scalp EEG traces of normal, Glioma and
Meningioma patients. In [3] it has been discussed how the
graphs of the scalp EEG patterns of healthy subjects from
those of subjects with brain tumors can be classified using
Multi Layer Feed Forward (MLFF) network. In [1] it has
been studied to separate EEG signals from tumor patients
into spatially independent source signals using a
probabilistic ICA algorithm modified by kernel-based
source density estimation. In [5] the authors have presented
their work in classifying the tumor EEG using Support
Vector Machine (SVM) with FFT-based spectral features.
In this paper, a successful proposal on the use of a
combination of time-domain and frequency-domain features
of the independent components of the scalp EEG obtained
using a Modified Wavelet-ICA (MwICA) in training a Multi
Layer Feed Forward (MLFF) Neural Network, popularly
known as Back Propagation Network (BPN), to classify a
brain tumor EEG segment from a normal one has been
presented. Two first order statistical features, namely the
Mean Square Amplitude (MSA) and the Mean Slope Rate
Brain Tumor Detection using Scalp EEG with
Modified Wavelet-ICA and Multi Layer Feed
Forward Neural Network
V. Salai Selvam and S. Shenbagadevi
D
978-1-4244-4122-8/11/$26.00 ©2011 IEEE 6104
33rd Annual International Conference of the IEEE EMBS
Boston, Massachusetts USA, August 30 - September 3, 2011

(MSR) that track the time-domain (morphological)
variations in the EEG signal, one second order statistical
feature, namely the Mean-to-Maximum Ratio of Power
Spectrum (mmrPS) and one third order statistical feature,
namely the Peak Bispectrum (pBS) that track the frequency-
domain (spectral) variations in the EEG signal have been
chosen. The literatures, [18] & [19] present the efficiency in
the use of these four statistical features in classifying various
characteristic waves (alpha, delta, spindle and K-complex)
of sleep EEG.
IV. MATERIALS AND METHODS
A. Materials
Nineteen Common Average Referenced (CAR) EEG
Channels in the standard 10-20 electrode system, were
obtained in digital format from 3 healthy subjects and 6
subjects with brain tumor (any type of brain structural
pathology was considered) in the age group of 8 to 60 years
for 20-25 minutes in the awake state with their eyes closed
at a sampling rate of 256 Hz. All EEG records were
bandpass-filtered to 1-70 Hz, 50-Hz-notch-filtered and
EMG-filtered using the software accompanied with the EEG
recorder. However some artifacts such as eye blinks, eye
movements, forehead and head movements, transient noises
and muscle noise resulting from facial muscle movements
were still present. Eliminating the epochs containing these
artifacts by visual inspection, only the artifacts-free 1000
seconds (256000 data points) of all the EEG records were
retained for the analysis.
B. Methods
Fig. 1 shows the block diagram representation of the
entire proposed method. The proposed method comprises
the following steps: the preprocessing of the EEG signal, a
new Independent Component Analysis (ICA) approach,
namely the Modified Wavelet-ICA (MwICA) for the
separation of EEG components, the extraction of features
that track the morphological and spectral variations of the
signal and the process of detection by a Multi Layer Feed
Forward (MLFF) neural network popularly known as the
Back Propagation Network (BPN).
C. Preprocessing
All the 19-channel, 1000-second (normal and brain
tumor) EEG records were split into 2-second (512 data
points) EEG epochs for further analysis in order to account
for the quasi-stationarity of the EEG signal. The issue of
stationarity is not a problem for the wavelet transform [25]
and the ICA [26]. However this is required for the features
to be extracted. The quasi-stationarity of the EEG signals
varies from 1 second to several minutes [25] [29]. However
a quasi-stationarity period of 1-2 seconds is typical for most
of the EEG signal analysis [28]. All these 19-channel, 2-
second EEG epochs were then lowpass-filtered to 40 Hz
using a 128-tap FIR filter as the EEG components of interest
were only below this frequency.
D. Modified Wavelet Transform based ICA (MWT-ICA)
The wavelet transform of a time-series is its multiband,
multiresolution decomposition using orthogonal (lowpass
and highpass) filters. The concept of wavelet transform and
its practical implementation version known as the discrete
wavelet transform (DWT) are very well discussed in [40]
[41] [42] and [43]. The Independent Component Analysis
(ICA) is a blind source separation technique that separates
statistically independent (rather uncorrelated) sources or
components from their linear mixtures [30]. The concept and
algorithms of the ICA techniques are discussed in [30] and
[44]. The application of the ICA to biomedical signals,
especially EEG is discussed in [21], [22], [23], [24], [45]
and [46].
The noteworthy earlier works on the efficient
combination of the wavelet transform and ICA are [35]-[38].
The Modified Wavelet-ICA (MwICA) is a modified version
of the Wavelet-ICA (wICA) techniques discussed in [36] &
[38]. However this proposal was a direct consequence of the
article in the literature [27]. According to [26], the number
of data points required to separate n sources is preferably
some multiples (at least equal to) n2
. But the wavelet
decomposition not only decorrelates the data but also
reduces the data size thereby increasing the speed of
convergence by ICA in the blind source separation process.
Fig. 2 (a) to (d) depict the entire process of MwICA. First
each of the 4500 19-channel, 2-second EEG epochs was
decomposed to a depth of level 3 using the Symlets wavelet,
‘sym5’ on a channel-by-channel basis. The choice of
decomposition level and the wavelet type was made based
on trial and error. The wavelets, Daubechies (db1 to db9),
Coiflets (coif1 to coif5) and Symlets (sym2 to sym8) were
tried for 1 to 10 decomposition levels. After the wavelet
decomposition, the ICA of the 3rd
level approximate
coefficients was performed using the SOBI-RO algorithm.
The resulting demixing (separating) matrix was used to
demix the detail coefficients of third, second and first levels.
The demixed wavelet coefficients were then reconstructed
on a channel-by-channel basis to obtain the final set of
independent components (ICs). However the result showed
that the ICs obtained from the 3rd
level approximate
coefficients alone were very much sufficient. This was
evident from the MwICA of the simulated data.
E. Feature extraction
The following features were then extracted from all the
4500 19-component 2-second independent components sets
on a component-by-component basis.
1) First Order Statistics: The Mean Square Amplitude
(MSAci) of an ith component, xci(n) of a 19-component, 2-
second (512 data points) IC set was calculated as the mean
of the squares of the samples of the component [18] [19] i.e.,
6105

MSAci={Σn[xci
2
(n)]/length(xci(n))}. The Mean Slope Rate
(MSRci) for an ith component, xci(n) of a 19-component, 2-
second (512 data points) IC set was calculated as described
in [18] and [19] i.e., MSRci=mean{Σk[xci(k)-xci(k+1)]/[tk-
tk+1]}.
2) Higher Order Statistics: The Power Spectral Density
(PSD) or simply the Power Spectra (PS) of a stationary
time-series is defined by the Wiener-Khintchine theorem as
the Fourier transform of the autocorrelation sequence of the
time-series [32] and in this work, the Welch method was
used to estimate the PSD of the given short-time series. The
Maximum-to-Mean Ratio of Power Spectrum (mmrPSci) of
an ith component, xci(n) of a 19-component, 2-second (512
data points) IC set was computed as the ratio of the
maximum value of the power spectrum, Pci(f) computed to
its mean value [18] [19] i.e., mmrPSci=max{Pci(f)}/mean{
Pci(f)}. Here two values of mmrPS, one being measured
below the frequency 6.5 Hz, named as mmrPSslwci (Max-to-
Mean Ratio of Slow Power Spectrum), and another above it,
named as mmrPSfstci (Max-to-Mean Ratio of Fast Power
Spectrum), were considered.
2-Level Wavelet
Decomposition
using ‘sym5’
2-Level Wavelet
Decomposition
using ‘sym5’
2-Level Wavelet
Decomposition
using ‘sym5’
Channel 1
Channel 2
Channel 19
cA21 cD21 cD11
cA22 cD22 cD12
cA219 cD219 cD119
19-channel,2-second(normalorbraintumor)
EEGepoch
Wavelet Coefficients
Fig. 2 (a) A 2-level wavelet decomposition of a 19-channel, 2-second (normal or brain
tumor) EEG epoch using the Symlets wavelet, ‘sym5’ on a channel-by-channel basis is
shown.
ICA using
SOBI-RO
Demixedsecondlevel
approximatecoefficients
ofallchannels
Secondlevelapproximate
coefficientsfromall
channels
W, demixing
matrix
cA21
cA22
cA219
dcA21
dcA22
dcA219
Fig. 2 (b) The ICA of the 3rd
level approximate coefficients from all
channels using the SOBI-RO algorithm is shown. The resulting
demixing matrix, W was used to demix the detail coefficients from
all channels.
Fig. 2 (d) Reconstruction of the demixed wavelet coefficients of all channels using the Symlets wavelet
‘sym5’ on a channel-by-channel basis is shown. The output of this step was the final set of independent
components.
Demixedwavelet
coefficientsofalllevels
fromallchannels
Finalsetofindependent
components
dcA21 dcD21 dcD11
dcA22 dcD22 dcD12
dcA219 dcD219 dcD119
IC1
IC2
IC19
2-Level Wavelet Reconstruction
using ‘sym5’
using ‘sym5’
using ‘sym5’
Fig. 1 The block diagram representation of
the proposed method is shown.
1000-second 19-channel (normal /
brain tumor) EEG record
Preprocessing (Lowpass-filtering to
40 Hz using a 128-tap FIR filter)
Modified Wavelet-ICA (MwICA)
(Separation of spatially independent
but temporally correlated EEG
components by SOBI-RO ICA
algorithm)
Feature Extraction (Extraction of
time-domain and frequency-domain
features from all EEG components)
Multi Layer Feed Forward (MLFF)
Neural Classifier (Training, testing
& simulation of MLFF or BPN
classifier to detect a given brain
tumor EEG segment)
Detection (Logical output as normal
or brain tumor EEG segment)
Detailcoefficientsofall
levelsfromallchannels
Demixing
by W
Demixed detail
coefficients of all levels
from all channels
cD219 cD119
cD22 cD12
cD21 cD11
dcD219 dcD119
dcD22 dcD12
dcD21 dcD11
Fig. 2 (c) The demixing of the detail coefficients of all levels from all
channels using the demixing matrix, W obtained from the step shown in Fig.
2 (b) is shown.
6106

The bispectrum of a stationary time series, x(n) is defined
as the Fourier transform of its third order cumulant [33] i.e.,
B(f1,f2)=FFT[Rxx(m1,m2)] where Rxx(.) is the third-order
cumulant of x(n) defined as the expected value of the triple
product i.e., Rxx(m1,m2)=E{x(n)x(n+m1)x(n+m2)}.
The bispectrum can be shown [33] to be
B(f1,f2)=X(f1)X(f2)X*
(f1+f2) where X(f) is the discrete Fourier
transform of the sequence, x(n).
The minimum variance estimation of bispectrum requires
a large number of data points. However it has been shown in
[39] that 512 data points (2 seconds) sampled at a rate of
256 Hz are sufficient to make a reasonable estimate of
bispectrum. The Peak Bispectrum (pBS) of an ith
component, xci(n) of a 19-component, 2-second (512 data
points) IC set was computed as the maximum value of the
bispectrum, Bci(f1,f2) of xci(n) computed using the Fast
Fourier Transform (FFT) as explained in [31]. Here again
two values of pBS, one being measured below 6.5 Hz,
named as pBSslw (Slow Peak Bispectrum), and another
above it, named as pBSfst (Fast Peak Bispectrum), were
considered.
At the end of this step there were 4500 feature vectors,
each of length 114 (6 features per component for each 19-
component, 2-second (512 data points) IC set), of which
1500 belonged to normal EEG and 3000 to brain tumor
EEG. Of these 4500 feature vectors, 3000 (1000 from
normal set and 2000 from brain tumor set) feature vectors
were chosen as the training set for the network to be
discussed in section F and the remaining for testing the
trained network.
F. Detection by Multi Layer Feed Forward (MLFF) neural
network
The choice of multi layer network, which is a non-linear
classifier [17], is based mainly on the fact that the scatter
plots of features within and between the classes (normal and
brain tumor cases) exhibit non-linearity. The other reasons
are the generalization of network, the ease of
implementation, the lesser computation overhead and the
availability of large options of network architectures with
simple addition or deletion of layers and/or neurons,
efficient learning and training algorithms etc. The factor for
the success of the training process and the generalization of
the network are discussed in [17] and [47] respectively. The
formulation of this aspect has been presented in [19].
A 3-layer MLFF network such as the one shown in Fig. 3
was chosen for the proposed work. The number of input
layer neurons was made equal to the dimension of the input
vector, i.e., 114. As there were two possible outcomes
whether the feature vector that was input to the network
belonged to normal EEG or brain tumor EEG, the logical
TABLE I
FOUR POSSIBILITIES OF NETWORK OUTCOMES IN DETECTION PROCESS
Actual case
P N
Network
decision
P/
True Positive False Positive
N/
False Negative True Negative
TABLE II
VALUES OF FOUR POSSIBILITIES OF NETWORK OUTCOMES LISTED IN TABLE I
Actual case
P N Total
Network
decision
P/
930 53 983
N/
70 447 517
Total 1000 500 1500
P-Actual brain tumor cases; N-Actual normal cases; P/
-Brain tumor cases as
per network; N/
-Normal cases as per network
TABLE III
VALUES OF PARAMETERS GIVEN BY EQUATIONS (9) TO (12)
Sensitivity or TPR 0.930
FPR 0.106
Accuracy 0.918
Specificity or TNR 0.894
Fig. 3 A 3-layer Multi Layer Feed Forward (MLFF) neural network is
shown.
Fig. 4 Receiver Operating Characteristics: The proposed method has the
point encouragingly at (0.106, 0.930).
6107

outputs that correspond to these two outcomes were chosen
to be the target vectors. The number of the output layer
neurons was chosen to be the size of the target vector. For
this proposed work the number of hidden layer neurons was
randomly chosen to be one-thirtieth of the number of
training vectors available i.e., 100.
V. RESULT AND DISCUSSION
The result of the testing phase has been shown in Table II.
The testing phase included the remaining 1500 cases, of
which 500 belonged to normal case and 1000 to brain tumor
case. The status that the chosen EEG epoch belonged to a
brain tumor case was considered as ‘positive’ and that it
belonged to a normal case as ‘negative’. Then the four
possibilities of the network outcomes were [48]: True
Positive (TP) if the network decided that a chosen EEG
belonged to a brain tumor case when it actually did, True
Negative (TN) if the network decided that a chosen EEG
belonged to a normal case when it actually did, False
Positive (FP) if the network decided that a chosen EEG
belonged to a brain tumor case when it actually belonged to
a normal case and False Negative (FN) if the network
decided that a chosen EEG belonged to a normal case when
it actually belonged to a brain tumor case. This is depicted in
Table I. From Table II the following parameters were
calculated to estimate the performance of the proposed
method [48] [49]: Sensitivity or True Positive Rate (TPR) as
[TP/(TP+FN)], Accuracy (ACC) as [(TP+TN/(P+N)] and
Specificity or True Negative Rate (TNR), which is one minus
False Positive Rate (FPR), as [TN/(FP+TN)] where P
stands for the total number of positive (brain tumor) cases
considered and N for that of negative (normal) cases
considered. The values of these parameters have been listed
in Table III. The ROC (Receiver Operating Characteristics)
was obtained from the values listed in Table III. Fig. 4
shows the ROC. From the ROC it is clear that the
performance of the proposed method in detecting the brain
tumor using the scalp EEG is very much encouraging.
VI. FUTURE DEVELOPMENT
To improve the detection (or classification) rate, not only
the features, such as the ones (except, possibly, the
bispectrum [34]) discussed in this paper, which track the
linear dynamics of the EEG signal but also the features
which track the nonlinear dynamics of the EEG signal can
be considered since the EEG exhibit both the linear and
nonlinear properties [50].
ACKNOWLEDGMENT
We would like to thank Dr. N. Kumaravel, Professor and
Head, Department of Electronics and Communication
Engineering, Anna University, Chennai, India – 600 025 for
his consistent encouragement in all our research activities.
We would like to thank the private clinics who provided us
the required data.
REFERENCES
[1] M. Habl, Ch. Bauer , Ch. Ziegaus, E. W. Lang and F. Schulmeyer,
Can ICA Help Identify Brain Tumor Related EEG Signals?, Proc.
International Workshop on INDEPENDENT COMPONENT
ANALYSIS and BLIND SIGNAL SEPARATION, 19-22 June 2000.
[2] Fadi N. Karameh and Munther A. Dahleh, Automated classification of
EEG signals in brain tumor diagnostics, Proceedings of the American
Control Conference, 2000, Vol.6, Issue 2000, pp. 4169-4173.
[3] S. Chetty (South Africa) and G.K. Venayagamoorthy (USA), A Neural
Network based Detection of Brain Tumours using
Electroencephalography, Proc. Artificial Intelligence and Soft
Computing, 2002.
[4] Rosaria Silipo, Gustavo Deco and Helmut Bartsch, Brain tumor
classification based on EEG hidden dynamics, Intelligent Data
Analysis, ELSEVIER, Volume 3, Number 4, October 1999, pp. 287-
306.
[5] M. Murugesan and Dr.(Mrs.).R. Sukanesh, Towards Detection of
Brain Tumor in Electroencephalogram Signals Using Support Vector
Machines, International Journal of Computer Theory and Engineering,
Vol. 1, No. 5, December, 2009, pp. 622-631.
[6] Edward B Bromfield (July 2009), The eMedicine website. [Online].
Available: http://emedicine.medscape.com/article/1137982-overview.
[7] S. C. O’Connor and P. A. Robinson, Analysis of the
electroencephalographic activity associated with thalamic tumors,
Journal of Theoretical Biology, ELSEVIER, 233, pp.271–286, 2005.
[8] Dr. Sudhir Kumar, Asymmetric Depression of Amplitude in
Electroencephalography leading to a Diagnosis of Ipsilateral
Cerebral Tumor, Annals of Indian Academy of Neurology, Vol.8,
pp.33-35, 2005.
[9] Yasser Metwally (December 2009), Online Braimmapping, Version
15, November 2009. [Online]. Available:
http://yassermetwally.wordpress.com/downloadable-case-
records/issues-in-brainmapping/
[10] Yasser Metwally (December 2007), Online Braimmapping, Beta
Version 1, December 2007. [Online]. Available:
http://yassermetwally.wordpress.com/downloadable-case-
records/issues-in-brainmapping/.
[11] American Brain Tumor Association (October 2010), Facts and
Statistics, October 2010. [Online]. Available:
http://www.abta.org/kids/learning/facts/.
[12] The Musella Foundation for Brain Tumor Research & Information
(April 2010), Brain Tumor Symptoms Survery Results, April 1020.
[Online]. Available:
http://www.virtualtrials.com/braintumorsymptomssurvey/.
[13] Brain Tumor Society (May, 2007), Brain Tumor Facts and Statistics,
May 2007. [Online]. Available:
http://www.revolutionhealth.com/articles/brain-tumor-facts-amp-
statistics/article/.
[14] Bettina C. Reulecke, Christian G. Erker, Barbara J. Fiedler, Thomas-
Ulrich Niederstadt and Gerhard Kurlemann (February 2008), Brain
Tumors in Children: Initial Symptoms and Their Influence on the Time
Span Between Symptom Onset and Diagnosis, Journal of Child
Neurology, Volume 23 Number 2, February 2008 178-183.
[15] National Brain Tumor Society (April 2010), Symptoms & Diagnosis,
April 2010. [Online]. Available: http://www.braintumor.org/patients-
family-friends/about-brain-tumors/symptoms-and-diagnosis/.
[16] Neil Schaul (1998), The fundamental neural mechanisms of
electroencephalography, Electroencephalography and Clinical
Neurophysiology, ELSEVIER, 1998, 101-107.
[17] Martin T. Hagan, Howard B. Demuth and Mark Beale (1996), Neural
Network Design, PWS Publishing Company, ch. 2, 10, 11 & 12.
[18] V. Salai Selvam and S. Shenbagadevi (2010), Variations of First and
Second Order Statistics for Characteristic Waves of Sleep EEG,
International Conference on Intelligent Information Systems and
Management (IISM’2010), June10-12, 2010.
[19] V. Salai Selvam and S. Shenbagadevi (August 2010), Classification of
Characteristic Waves of Sleep EEG Using First and Higher Order
Statistics and Neural Network, CiiT International Journal of Artificial
Intelligent Systems and Machine Learning, Vol 2, No 8, August 2010,
142-150.
6108

[20] Nikias, C. L. and Petropulu, A. P. (1993), Higher Order Spectra
Analysis: A Nonlinear Signal Processing Framework, Prentice Hall,
Englewood Cliffs, NJ, 1993.
[21] TZYY-PING JUNG, SCOTT MAKEIG, COLIN HUMPHRIES, TE-
WON LEE, MARTIN J. MCKEOWN, VICENTE IRAGUI and
TERRENCE J. SEJNOWSKI (2000), Removing
electroencephalographic artifacts by blind source separation,
Psychophysiology, 37 (2000), 163–178. Cambridge University Press.
[22] TZYY-PING JUNG, COLIN HUMPHRIES, TE-WON LEE, SCOTT
MAKEIG, MARTIN J. MCKEOWN, VICENTE IRAGUI and
TERRENCE J. SEJNOWSKI (1998), Extended ICA Removes Artifacts
from Electroencephalographic Recordings, Advances in Neural
Information Processing System, 10:894-900, 1998.
[23] Tzyy-Ping Jung, Scott Makeig, Marissa Westerfield, Jeanne
Townsend, Eric Courchesne and Terrence J. Sejnowski, Removal of
eye activity artifacts from visual event-related potentials in normal
and clinical subjects, Clinical Neurophysiology, ELSEVIER, 111
(2000) 1745-1758.
[24] L. Albera, P. Comon, L.C. Parra, A. Karfoul, A. Kachenoura and L.
Senhadji (2009), ICA and biomedical applications, Handbook of
Blind Source Separation, Independent Component Analysis and
Applications, ch. 18, October 2009.
[25] Rodrigo Quian Quiroga (1998), Quantitative analysis of EEG signals:
Time-frequency methods and Chaotic Theory, Institute of Physiology
and Institute of Signal Processing, Medical University, Luneck, 1998.
[26] Scott Mekeig (2010), Frequently Asked Questions about ICA applied
to EEG and MEG data, February 2010. [Online]. Available:
http://sccn.ucsd.edu/~scott/tutorial/questions/.
[27] Aapo Hyvarinen, Pavan Ramkumar, Lauri Parkkonen and Riitta Hari
(August 2009), Independent component analysis of short-time Fourier
Transforms for spontaneous EEG/MEG analysis, Preprint submitted
to ELSEVIER, August 2009.
[28] Alexander A. Fingelkurts and Andrew A. Fingelkurts (2008), EEG
Oscillatory States: Temporal and Spatial Microstructure, 14th World
Congress of Psychophysiology, St Petersburg, 2008.
[29] Maryann D’Alessandro, Rosana Esteller, George Vachtsevanos,
Arthur Hinson, Javier Echauz, and Brian Litt (May 2003), Epileptic
Seizure Prediction Using Hybrid Feature Selection Over Multiple
Intracranial EEG Electrode Contacts: A Report of Four Patients,
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL.
50, NO. 5, MAY 2003.
[30] Aapo Hyvarinen and Erkki Oja (2000), Independent Component
Analysis: Algorithms and Applications, Neural Networks 13 (4-
5):411-430, 2000.
[31] Nikias C. L and Raghuveer M. R (1993), Higher-Order Spectra
Analysis: A Nonlinear Signal Processing Framework, Prentice Hall
Inc., Englewood Cliffs, NJ, 1993.
[32] John G. Proakis and Dimitris G. Manolakis (2000), Digital Signal
Processing Principles, Algorithms and Applications, Prentice-Hall of
India Pvt Ltd.
[33] Tayfun Akgül, Mingui Sun, Robert J. Sclabassi and A. Enis Çetin
(2000), Characterization of Sleep Spindles Using Higher Order
Statistics and Spectra, IEEE Transactions On Biomedical
Engineering, VOL. 47, NO. 8.
[34] Taikang Ning and Joseph D. Bronzino (1988), Detecting Phase
Coupling of Sleep EEG via Bispectra, Bioengineering Conference,
1988, Proceedings of the 1988 Fourteenth Annual Northeast.
[35] V. Vigneron, A. Paraschiv-Ionescu, A. Azancot, O. Sibony,C. Jutten
(2003), Fetal Electrocardiogram Extraction Based on Non-Stationary
ICA and Wavelet Denoising, ISSPA 2003.
[36] Bruno AZZERBONI, Fabio LA FORESTA, Nadia MAMMONE, and
Francesco Carlo MORABITO (April 2005), A New Approach Based
On Wavelet-ICA Algorithms For Fetal Electrocardiogram Extraction,
ESSAN’2005 Proceedings – European Symposium on Artificial
Neural Networks, Bruges (Belgium), 27-29 April 2005.
[37] Xiaomei Ren, Xiao Hu, Zhizhong Wang and Zhiguo Yan (April
2006), MUAP extraction and classification based on wavelet
transform and ICA for EMG decomposition, Med Biol Eng Comput
(2006) 44: 371–382.
[38] Joachim Taelman, Sabine Van Huffel and Arthur Spaepen (April
2007), Wavelet-Independent Component Analysis to remove
Electrocardiography Contamination in surface Electromyography,
April 2007.
[39] A. Miller, J. W. Sleigh, J. Barnard and D. A. Steyn-Ross (2004), Does
bispectral analysis of the electroencephalogram add anything but
complexity?, British Journal of Anaesthesia 92 (1): 8-13 (2004).
[40] Hans-Georg Stark (2005), Wavelets and Signal Processing An
Application-Based Introduction, Springer 2005.
[41] Martin Vetterli and Jelena Kovacevic (2007), Wavelets and Subband
Coding, Prentice Hall, Englewood Cliﬀs, New Jersey.
[42] C. Valens (1999), A Really Friendly Guide to Wavelets, [Online].
Available: http://polyvalens.pagesperso-
orange.fr/clemens/wavelets/wavelets.html.
[43] Daubechies, I (1992), Ten Lectures on Wavelets, SIAM, 1992.
[44] Aapo Hyvarinen, The Fixed-Point Algorithm and Maximum
Likelihood Estimation for Independent Component Analysis, [Online].
Available: http://www.cis.hut.fi/~aapo/.
[45] Radu Mutihac and Radu Cristian Mutihac (2007), A COMPARATIVE
STUDY OF INDEPENDENT COMPONENT ANALYSIS
ALGORITHMS FOR ELECTROENCEPHALOGRAPHY, Romanian
Reports in Physics, Vol. 59, No. 3, P. 831–860, 2007.
[46] Ricardo Vigário, Jaakko Särelä, Veikko Jousmäki, Matti Hämäläinen,
and Erkki Oja (May 2000),Independent Component Approach to the
Analysis of EEG and MEG Recordings, IEEE TRANSACTIONS ON
BIOMEDICAL ENGINEERING, VOL. 47, NO. 5, MAY 2000.
[47] Donna L. Hudson and Maurice E. Cohen (2001), ‘Neural Networks
and Artificial Intelligence for Biomedical Engineering’, IEEE
Engineering in Medicine and Biology Society, Sponsor, IEEE Press
Series in Biomedical Engineering, Prentice Hall of India, Pvt., Ltd.,
New Delhi, India, pp. 51-52.
[48] Fawcett T. (2006), An introduction to ROC analysis, Pattern
Recognition Letters, 27, 861-874.
[49] P. J. Durka, H. Klekowicz, K. J. Blinowska, W. Szelenberger, and Sz.
Niemcewicz (2003), A Simple System for Detection of EEG Artifacts
in Polysomnographic Recordings, IEEE TRANSACTIONS ON
BIOMEDICAL ENGINEERING, VOL. 50, NO. 4, APRIL 2003.
[50] Das, A. and Das, P. (2005), Nonlinearity in EEG: Investigation by
Surrogate Data Analysis, Complexity International, Volume 10, 2005.
6109

Int conf 03

More Related Content

What's hot (18)

Similar to Int conf 03 (20)

Recently uploaded (20)

Int conf 03