Académique Documents
Professionnel Documents
Culture Documents
com
Procedia
Engineering
www.elsevier.com/locate/procedia
Abstract
In this paper it is proposed to separate the ECG signal taken from skin electrodes located on a pregnant womans
body into Maternal Electro Cardiogram signal and Fetal Electro Cardiogram signal. Blind Source Separation is the
technique used for separating these source signals. ICA is applied on the mixed signals and the separated signals are
reconstructed using wavelet reconstruction. Comparison results show that Lifting Wavelet Transformation and
FASTICA algorithm produces the best SNR value of 11.39 for maternal and 10.10 for fetal Electro Cardio Gram
signals.
2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of ICCTSD 2011
Keywords: Blind Source Separation; Independent Component Analysis; Lifting wavelet transformation; FASTICA; SNR.
1.
Introduction
ECG signal is the record of electrical activity of the heart that results when the heart muscle cells
in atria and ventricles contract. ECG is a non stationary signal and composed of Atrial depolarization (P
wave), ventricular depolarization (QRS complex) and repolarization of the ventricles (T wave).The FECG
signal is comparatively weak in amplitude (less than 20% of the mothers ECG) [1] and often mixed with
noise. The Fetal Heart Rate (FHR) lies in the range from 1.3 to 3.5 Hz [1] and sometimes it is possible for
the maternal and some of the fetal ECG signals to be closely overlapping. So far, research and extensive
works have been made in this area, developing better algorithms, upgrading existing methodologies, and
improving detection techniques to reduce noise and acquire accurate FECG signals.
_____________________
Wavelet transform is well fitting to non stationary signals like ECG. The combination of wavelet analysis
and BSS methods. also shows potential attitude for the separation of the maternal and fetal signals from
ECG [1]. Blind adaptive-filtering tactic overcomes the theoretical limitations in applying conventional
BSS methods based on ICA for FECG signal extraction problem.
In this regard different ECG databases were collected from San Joe hospital, Perumbavoor, Kerala, CSI
Mission Hospital, Neyyoor , James hospital, Nagercoil and some were also collected from MIT and
physionet websites [15] which were in digital format.
hese signals were scanned and converted to jpeg format as images, which were then read by MATLAB
and converted into a one dimensional format to be used in the algorithm.
2.
Methodology
This graphical image is read and then converted into binary format and finally it is scanned and
mapped to two-dimensional format. This signal is then preprocessed using Wavelet transform to remove
noise. It is then decomposed to time-frequency domain using Wavelet Lifting decomposition. The noisy
FECG signal is decomposed to five levels of wavelet transform by using the daubechies wavelet (db4)
[2]. The fetal and maternal signals are then separated by applying FAST ICA algorithm to the
decomposed signal. The separated signals are finally reconstructed using the Inverse Wavelet Lifting
Transform.
357
358
INFORMATION SOURCE
INFORMATION SOURCE
INFORMATION TO SIGNAL MAPPING
ALGORITHM
WAVELET RECONSTRUCTION 1
MATERNAL ECG
WAVELET RECONSTRUCTION 2
FETAL ECG
2.1 Preprocessing
First step in separation algorithm is to remove the noise (due to graph to signal conversion process) from
the input ECG signal [2],[6]. In this case wavelet denoising technique is used.
Fig.4. Denoising
359
J. Jasper Jasper
Raj Immanuel
et al. / Procedia
Engineering
30 (2012)
363
Raj Immanuel,et.,al
/ Procedia
Engineering
00 356
(2011)
000000
In case of noisy FECG signal, it is decomposed to five levels of wavelet transform by using daubechies
wavelet. The transform consists of a set of high pass and low pass filters in the input stage and the same
number in output stage. One factor that affects the number of levels that can be reached to achieve the
satisfactory noise removal result is signal-to-noise ratio (SNR) in the original signal.
The threshold values are calculated at each decomposition level using the formula,
(1)
3.
Wavelet Decomposition
Even samples
Input signal
Dual Lifting
Primal Lifting
Odd samples
From the Fig. 5 it can be inferred that the input is divided into even and odd stream, followed by lifting
one of the streams by applying Laurent polynomial to other and adding it to the first.The lifting theorem
now states that any other finite filter gnew complementary to h is of the form,
gnew(z) = g(z) + h(z)s(z2)
h
4.
new
(2)
(3)
ICA is a signal processing technique that has demonstrated the ability of separating independent
sources from mixed recorded signals [3]. The input signal is acquired from Magnetocardiography (MCG),
a technique that allows the non-invasive detection of the extremely low magnetic field variations
associated to the spontaneous electrical activity of the human heart. The Fetal Magnetocardiography
(FMCG) contains electrical activity of both the fetal and maternal hearts, since both are included in the
same space volume, i.e. in the maternal abdomen. The ICA problem is rather complex in the case of
FECG because
Both signals are quasi periodic and have a similar waveform.
The fetal and maternal cardiac signals are characterized by different values of frequency(fetal
heart rate is greater than maternal) and intensity(the strength of maternal heart is greater than the
fetus).
360
J. Jasper
Raj Immanuel et
al. / Procedia
Engineering
30 (2012)
356 363
Jasper
Raj Immanuel,et.,al
/ Procedia
Engineering
00 (2011)
000000
In FMCG some weak fetal bats may be hidden behind the more intense maternal beats and heart
rate is not perfectly constant.
To use non - gaussianity in ICA estimation, we must have a quantitative measure of non gaussianity of a
random variable such as Mean, Kurtosis [12], Standard deviation, Negentropy [12].
ICA is a statistical method that searches from multivariate statistical data for underlying factors or
components that are independent [4]. The recorded data is a linear superposition of n different sources of
the form,
X = A*S,
(4)
A-Mixing Matrix, S-Source Matrix.
The goal of ICA is to find a Separating Matrix W, which is as close to A-1 as possible, based upon
proper statistical criteria in order to optimally recover the original source signal,
y(t) =W*x(t)=W*A*S(t)=S(t)
(5)
4.1. FASTICA
The basic concept of Fast ICA is to take a neural network leaning rule and convert it into a fixedpoint iteration. This yields an algorithm that is very simple, does not depend on any user-defined
parameters and is fast to converge to the most accurate solution allowed by the data. The FASTICA
require 10 % of the floating point operations used by the neural algorithm [14],[5].
4.2. Fixed point FASTICA
In order to obtain something that could be implemented in hardware or hardware-software
systems, it is necessary to have any algorithm that can work with purely fixed point numbers [14]. The
algorithm to satisfy this requirement is developed in MATLAB. The data was converted to double for
input to both whitening and PCA, and the output whitened data was converted to the same 16 bit numbers
as above before input to the ICA algorithm [10]. This involved several multiplication and additions of
arrays of numbers. Typically normalization requires taking square root and division which done by
existing methods for fast square root algorithm of integers and division could also implemented by an
embedded CPU.
4.3. ERICA
ERICA known as equivariant robust independent component analysis is one of the FASTICA
technique which is being used here to implement separation. The ERICA algorithm can be applied to both
real valued and complex-valued data. The term equivariance implies that the variance of both the original
and the reconstructed signal has to be similar.
5. Reconstruction
In the inverse transform, upsampling is to be done after which filtering is performed [3]. The
drawback of this method is that the filter will perform a lot of multiplications by zero. Inorder to
overcome this, upsampler is added behind the filters. In the inverse lifting transform it is possible to undo
this lifting step by again applying the same Laurent polynomial to the other stream and then subtract it
from the first. In this case low pass sub band is lifted first with the help of high pass sub band, called as
primal lifting. The high pass sub band is then lifted with the help of low pass sub band, called as dual
lifting. Reconstructed signal is taken out from switch as shown in fig.6,
361
High frequency
component
Primal Lifting
Dual Lifting
Low frequency
component
Fig. 6. Inverse Lifting Wavelet Transform.
6. Result
6.1. Gaussianity Test
Table 1. Parameters for gaussianity.
PARAMETER
SIGNAL 1
SIGNAL 2
MEAN
43.7941
-137.17
69.414
11.4052
1.371
1.6315
NON GAUSSIAN
S.D.
116.91
KURTOSIS
3.3246
SKEWNESS
0.5509
NEGENTROPY
2.4324
RESULT
NON GAUSSIAN
ALGORITHM
SIGNAL 1
SIGNAL 2
ICA
1.8
11.32
8.68
9.32
ERICA
7.78
11.98
FAST ICA
10.10
18.3
Switch
362
J. Jasper
Raj Immanuel et
al. / Procedia
Engineering
30 (2012)
356 363
Jasper
Raj Immanuel,et.,al
/ Procedia
Engineering
00 (2011)
000000
ALGORITHM
SIGNAL 1
SIGNAL 2
ICA
10.00
9.39
ERICA
10.06
11.09
FAST ICA
11.39
11.76
J. Jasper
et al. // Procedia
363
Jasper Raj
Raj Immanuel
Immanuel,et.,al
ProcediaEngineering
Engineering30
00(2012)
(2011)356
000000
7. References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
M.A Hasan, M. B. I. Reaz, et al., Detection and Processing Techniques of FECG Signal for Fetal Monitoring,
Biological Procedures Online, vol. 11(1), pp. 263-295. March 2009. ISSN: 1480-9222.
Mikhled Alfaouri and Khaled Daqrouq, ECG Signal Denoising By Wavelet Transform Thresholding, American Journal of
Applied Sciences 5 (3):276-281, 2008, ISSN 1546-9239.
Silvia Comania, Dante Mantinic, et al., Independent Component Analysis: Fetal Signal Reconstruction from
magnetocardiographic recordings, Elsevier Computer Methods and Programs in Biomedicine, (2004) 75, pp.163-177.
Burghoff, M., Van Leeuwen, P., Separation of fetal and maternal magnetocardiographic signals in twin pregnancy using
independent component analysis (ICA), Neurology, Neurophysiology and Neuroscience, 2004.
V.K.Ananthashayana, and Jyothirmayi M., Blind Source Separation Using Modified Gaussian FastICA, World Academy of
Science, Engineering and Technology 56, 2009.
Sylvain Sardy, Paul Tseng, et al., Robust Wavelet Denoising, IEEE Transactions on Signal Processing, Volume 49, No. 6,
pp. 1146-1152, June 2001.
D.L.Donoho, Denoising by Soft Thresholding, IEEE Transactions on Information Theory, Vol. 41, pp. 613-627, May 1995.
M.Kania, M.Fereniec, et al., Wavelet Denoising for Multi-Lead high resolution ECG signals, Measurement Science Review,
Vol.7, Sec. 2, No.4, 2007.
W. Chen, T. Nemoto, et al., Fetal ECG Extraction from Maternal Body Surface Measurement using Independent Component
Analysis, Engineering in Medicine and Biology Society, 2001. Proceedings of the 23rd Annual International Conference of
the IEEE, vol. 2, 2001,pp 1990-1993,.
Aapo Hyvrinen, Fast and Robust Fixed-Point algorithms for Independent Component Analysis, IEEE Transactions on
Neural Networks, 10(3):626-634, 1999.
Book: Aapo Hyvarinen, Juha Karhunen, and Erkki Oja, Independent Component Analysis, Copyright @ 2001 by John
Wiley & Sons, Inc.
Book: Hyvarinen.A and Oja.E (2000). Independent Component Analysis: Algorithms and Applications Neural Networks,
13(4-5):411-430,2000
Wei Yu Liu Zhenxing Li Changhai WuHan Univ.of Sci. & technol., Wuhan Improved Particle Swarm to Nonlinear Blind
Source Separation , IEEE International Symposium on Microwave, Antenna & EMC technologies for wireless
communications.
Hyvarinen.A and Bingham.E (2000). A Fast Fixed- point algorithm for independent component analysis of complex valued
signals. Helsinki University Of technology, http://www.cis.hut.fi/projects/ica .
MITBIH Database Distribution, http://www.ecg.mit.edu, and Physio Bank Archive Index
http://www.physionet.org/pn3/nifecgdb/
363