EEG NOISE CANCELLATION BASED ON NEURAL NETWORK
J. Mateo
1
, A. Torres
1
,
C. Soria
2
, Mª. García
2
and C. Sánchez
1
1
Innovation in Bioengineering Research Group, University of Castilla-La Mancha, Cuenca, Spain
2
Clinical Neurophysiology Service, Virgen de la Luz Hospital (SESCAM), Cuenca, Spain
Keywords: Biomedical signals, Muscle noise, Electrocardiogram, Neural network.
Abstract: Electroencephalogram (EEG) recordings often experience interference by different kinds of noise, including
white and muscle, severely limiting its utility. Artificial neural networks (ANNs) are effective and powerful
tools for removing interference from EEGs, but the quality of the separation is highly dependent on the type
and degree of contamination. Several methods have been developed, but ANNs appear to be the most
effective for reducing muscle contamination, especially when the contamination is greater in amplitude than
the brain signal. We propose an ANN as a filter for EEG recordings, developing a novel framework for
investigating and comparing the relative performance of an ANN incorporating real EEG recordings from
the Clinical Neurophysiology Service at the Virgen de la Luz Hospital in Cuenca (Spain). This method was
based on a growing ANN that optimised the number of nodes in the hidden layer and the coefficient
matrices, which were optimised by the simultaneous perturbation method. The ANN improved the results
obtained with the conventional EEG filtering techniques: wavelet, singular value decomposition, principal
component analysis, adaptive filtering and independent components analysis. The system was evaluated
within a wide range of EEG signals in which noise was added. The present study introduces a method of
reducing all EEG interference signals with low EEG distortion and high noise reduction.
1 INTRODUCTION
Noise reduction is a matter of considerable
importance in biomedical signal processing
applications, especially electroencephalogram
(EEG) analysis (Sörnmo, 2005); (Bronzino, 2000);
(Rangayyan, 2002).
Noncortical biological artifacts are the principal
source of contamination in EEG recordings and are
generated primarily by movements, cardiac pulse,
and muscle activity, particularly that of the face
(especially the jaw) and neck. EEG experimental
design is generally constrained by the desire to
minimise the effect of these artifacts.
Several methods have been suggested for muscle
noise reduction. Signal processing techniques used
for noise elimination include bandpass filtering, fast
Fourier transform, autocorrelation, autoregressive
modelling, adaptive filtering, Kalman filtering,
Bayesian filtering, singular value decomposition
(SVD) (Paul 2000); (Shao, 2009); (Zhang, 2006);
(Sameni, 2008) one of the common approaches is
the adaptive filtering (AF) architecture which has
been used for the noise cancellation of ECG (Olmos,
2002) and wavelet (Castellanos, 2006). Recently,
principal component analysis (PCA) (Lagerlund,
2004) and independent component analysis (ICA)
(Crespo-Garcia, 2008) have become popular for
analysing biomedical data. One of the main
advantages of these approaches relates to their
applicability to multisensory observations of mixed
signals. However, PCA is unable to separate some
artifact signals from brain signals when they have
similar amplitudes. In addition, both PCA and SVD
perform well only if the noise level is low enough
and a signal subspace and noise subspace are
orthogonal to each other. For practical applications,
the orthogonality requirement is usually not valid.
On the other hand, ICA cannot guarantee that some
individual independent components (ICs) contain
only noise and not information about useful sources,
especially in biomedical applications. Thus, the
problem of detection and filtering the "useful" part
of each IC is still open, and additional tools are
needed to solve it.
In the present study, we created an artificial
neural network (ANN) that can act as a filter for
EEG recordings. The network was trained using the
330
Mateo J., Torres A., Soria C., García M. and Sánchez C..
EEG NOISE CANCELLATION BASED ON NEURAL NETWORK.
DOI: 10.5220/0003657103300333
In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2011), pages 330-333
ISBN: 978-989-8425-84-3
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
simultaneous perturbation (SP) method. The ANN
was chosen mainly because of its adaptability to the
nonlinear and time-varying features of the noise.
This system was evaluated within a wide range of
EEG signals in which white noise and muscle noise
were added from the Clinical Neurophysiology
Service at Virgen de la Luz Hospital. Thus, this
algorithm could serve as an effective framework for
filtering noise in EEG recordings. We expect that the
distortion of this signal will be reduced compared to
conventional methodologies. The results
demonstrate that this method can maintain the
original shape of the EEG signal in very low SNR
conditions in which the brain signal is mixed with
the noise.
2 MATERIALS
The signals considered in the present study
originated from patients at the referred hospital. All
signals were recorded using Viasys Healthcare –
NicoletOne equipment, which had been
implemented within a concrete period of time. Sleep
studies were also included.
All signals obtained from the hospital were
randomly classified into three groups, and each of
them used a different phase in the filtering process
with ANN. Forty signals were chosen to integrate
the first group, which was used for network training.
The second group was used to validate and compare
proper ANN function, and the third group was used
to compare ANN with the other systems.
The first and second groups were comprised of
80 signals that lasted 25-50 minutes. Also, 10 sleep
testing of approximately eight hours, have been
included in these groups. These signals were filtered
through ICA (Crespo-Garcia, 2008) to remove any
current noise.
Once this process was completed, white and
muscle noise were sequentially added to the EEG
signal as defined by equation. 1, where d(k) is the
EEG signal after ICA filtering, A is the amplitude of
the added noise, and n(k) is the noise signal.
Amplitude A was modified in order to obtain an
SNR margin between -10dB and 30dB. The main
aim was to estimate the clean signal d(k) from the
noisy signal p(k). These recordings are synthetic
signals (EEG records to which different noises were
added).
p(k) = d(k) +An (k) (1)
The third group of signals was made up of 50
signals; 40 signals that lasted 30-60 minutes and 10
sleep testing of 8 hours, and neither noise nor
variation was added or modified in them (real
signals). These signals were also used to compare
the above mentioned methods to ANN.
3 METHOD
The Adaline network uses supervised learning and
involves a sum of products using the input and
weight vectors. An adaptive operation means that
there is a mechanism by which w
i
, v
i
can be adjusted,
usually iteratively, to achieve correct values.
Regarding Adaline, the properties of the perturbation
vectors are assumed to be as described by Maeda
and De Figueiredo (Maeda, 1997).
Figure 1: Proposed Neuronal Network with one neuron in
the hidden layer. The black coefficients are constants.
For every input, the network output differs from
the expected target value d
n
by (d
n
– p
n
), where p
n
is
the current output and n is the number of signals.
This network structure was initially made up of
three layers: an input layer, one hidden layer made
up of 40 neurons, and an output layer. Once this
network was trained, its work was re-evaluated and,
if necessary, more neurons added to the hidden layer
(Figure 1). This procedure was repeated until
expected results are obtained. At all these stages, the
ANN was adapted using the SP method in order to
obtain the best results. The process of training and
initialization was modified and implemented as
described by Maeda (Maeda, 1995, 1997). The
detailed strict convergence conditions of SP were
described by Spall (Spall ,1992).
3.1 Learning Algorithm using
Simultaneous Perturbation
Simultaneous perturbation technique for training
neural networks has been introduced by Spall (Spall,
1992). Other authors (Maeda, 1995) have also
reported results of similar methods. To adapt the
1
Fs
+
Y1
X1
Fs
+
Y2
Xn
Fs
+
Ym
Input
W
V
Output
Fs
+
Fs
+
Fs
+
Fs
+
Hidden layer
EEG NOISE CANCELLATION BASED ON NEURAL NETWORK
331
weights of the system is necessary to consider the
gradient of the error function, this is:
w
w
J
)(
(2)
Defining the error function like:
()
2
2
1
)( dywJ =
(3)
where;
()
dy =
ε
(4)
Using the equation (4) it is possible to measure the
error between the present exit and the wished exit.
On the other hand, the approach of differences is a
procedure known to obtain the derived from a
function, so this approach to reduce the complexity
can be used (Haykin, 1994). c is a perturbation
added to the i component (Maeda, 1997).
The neural network exit, Y, is a function of the
vector of weights:
c
wYfwYf
w
wJ
i
i
))(())(()(
(5)
Nevertheless the above idea which is very simple,
needs more operations. It is due to evaluate
()
i
Jw
for
all the network weights to obtain the amount
modified for all the weights.
i
t
tttttt
i
t
c
VWYDVCWY
w
),(),( ++
=Δ
(6)
The weights of the neural network are adapted using
the following rule:


=
−∆
.
The best results have been obtained when 15
neurons were added to the hidden layer. When more
than 15 neurons are added, there is no improvement
of both the computational load and the noise
reduction.
4 RESULTS
Noise reduction is important for obtaining a clear
and useful signal. Some signals, such as EEG, are
non-stationary, and the noise statistical property is
complicated because of the complexity of the signal.
Different techniques have been proposed to reject
muscle noise in EEG signals; these conventional
filtering techniques can contain ripples that do not
correspond to the original EEG. ANN improves all
results obtained by wavelet, SVD, PCA, AF and
ICA, significantly reducing the interference, Figure
2. The methods are referred in introduction section.
Figure 2: Comparison of the muscle noise removal by
ANN and traditional techniques for F7-T3 derivation.
a) Original recording without processing. b) Input signal
of 8 dB muscle noise used to compare the different
methods. c) Filtering results for muscle noise with the
wavelet method. d) SVD method. e) PCA method. f) AF
method. g) ICA method and h) ANN
=
2
2
{
{
log20
xxE
xxE
SIR
out
in
(7)
Table 1: Obtained results of the cross correlation and SIR
for muscle noise, average values.
Methods
Synthetic
(CC)
Real(CC) SIR
Wavelet
SVD
PCA
AF
ICA
ANN
03.085.0 ±
02.088.0 ±
04.082.0 ±
03.089.0 ±
02.091.0 ±
02.096.0 ±
03.082.0 ±
03.086.0 ±
03.080.0 ±
04.086.0 ±
03.088.0 ±
02.095.0 ±
2.19.12 ±
13.25.13 ±
3.15.12 ±
27.20.14 ±
5.12.15 ±
1.11.19 ±
Table 1 shows the cross-correlation average
values and standard deviation in synthetic signals
and real signals. Nevertheless table 1 shows the SIR
average values calculated for synthetic recordings.
Equation 7, shows SIR expression where x
in
shows
the input to the system, x
out
the exit and x the
original registry without noise. As can be seen from
table 1, with ANN the values are higher than the
values obtained with the other systems, both
synthetic signals and real signals.
Figure 3 shows the time-frequency analysis of
EEG signal with muscle noise. As can be seen the
ANN system reduces fluctuations due to muscle
noise and gets a more uniform signal.
NCTA 2011 - International Conference on Neural Computation Theory and Applications
332
Figure 3: The time-frequency analysis for a noisy signal is
shown in the upper figure and for ANN output is shown in
the lower figure.
5 CONCLUSIONS
The present study demonstrates how ANN can be
used to reduce muscle noise in EEG data.
Throughout all stages, the ANN method was adapted
using the SP method, which was improved to
achieve our target. Our ANN method was shown to
be an effective enhancement tool. The techniques
proposed here can be applied in multichannel EEG.
In all of the practical cases studied, different kinds
of noise components appear in the recordings. For
this reason, the removal of noise facilitates the
clinical analysis for medical professional use.
As a way of conclusion, suffice is to say that the
ANN - based approach obtains both more signal
reduction and low distortion of the signal results in
comparison with traditional filtering techniques. The
results of this study show the maintenance of clinical
information. The technique which has been
proposing through this paper, finds its application by
means of denoising biological signals (EEG, ECG,
etc).
ACKNOWLEDGEMENTS
This work was sponsored by University of Castilla-
La Mancha (Spain) and Virgen de la Luz Hospital of
Cuenca (Spain).
REFERENCES
Bronzino, J., The Biomedical Engineering Handbook, 2nd
Ed, CRC Press, Springer, 2000.
Castellanos, N. P., Makarov,V. A., Recovering EEG brain
signals: Artifact suppression with wavelet enhanced
independent component analysis, J. Neuroscience.
Methods 158 (2006). 300–312.
Crespo-Garcia, M., Atienza M., Cantero, J., Muscle
artifact removal from human sleep eeg by using
independent component analysis, Annals Biomed Eng
36 (3) (2008) 467–475.
Haykin S., (1994) Neural Networks: A Comprehensive
Approach, IEEE Computer Society Press 1994,
Piscataway, USA.
Lagerlund, T, Sharbrough, F., Busacker N. Spatial
filtering of multichannel electroencephalographic
recordings through principal component analysis by
singular value decomposition, Brain Topography 17
(2) (2004) 73–84.
Maeda, Y., Hirano, H., Kanata, Y. A learning rule of
neural networks via simultaneous perturbation and its
hardware implementation, Neural Networks 8 (2)
(1995) 251–259.
Maeda, Y; Figueiredo, R. J. P. D., Learning rules for
neuro-controller via simultaneous perturbation, IEEE
Trans. on Neural Networks 8 (6) (1997) 1119–1130.
Olmos, S., Sörnmo, L. Laguna, P., Block adaptive filters
with deterministic reference inputs for event-related
signals: Blms and brls, IEEE Trans. Sign. Proces. 50
(5) (2002) 1102–1112.
Paul, J. S., Reddy, M. R., Kumar, V. J, A transform
domain SVD filter for suppression of muscle noise
artefacts in exercise ECG’s, IEEE Trans. Biomed.
Eng. 47 (5) (2000) 654–663
Rangayyan R., Biomedical Signal Analysis: A Case-Study
Approach, IEEE Press Series in Biomedical
Engineering, 2002.
Sameni, R; Shamsollahi, M.; Jutten, C. Model-based
bayesian filtering of cardiac contaminants from
biomedical recordings, Physiological Measurement 29
(5) (2008) 595–613.
Shao, S.-Y; Shen, K.-Q.; Ong, C. J.; Wilder-Smith, E. P.
V; Li, X.-P, Automatic eeg artifact removal: A
weighted support vector machine approach with error
correction, IEEE Trans. on Biomed. Eng. 56 (2)
(2009) 336–344.
Sörnmo, L. and Laguna, P., Bioelectrical Signal
Processing in Cardiac an Neurological Applications,
Elsevier Academic Press, 2005.
Spall, J. C. Multivariate stochastic approximation using a
simultaneous perturbation gradient approximation,
IEEE Trans. Automatic Control 37 (3) (1992) 332–
341.
Zhang, T; Okada, Y; Recursive artifact windowed-single
tone extraction method (raw-stem) as periodic noise
filter for electrophysiological signals with interfering
transients, J. Neuroscience Methods 155 (2) (2006)
308–318.
EEG NOISE CANCELLATION BASED ON NEURAL NETWORK
333