When trying to record foetal ECG (FECG), using non

advertisement
SEPARATION OF TWINS FETAL ECG BY MEANS OF BLIND SOURCE
SEPARATION (BSS)
A. Kam and A. Cohen
Signal Processing Laboratory, Dept. of Electrical and Computer Engineering,
Ben-Gurion University, Beer-Sheva, Israel.
E-mail: amitk(arnon)@eesrv.bgu.ac.il
Abstract
The Maternal ECG (MECG) is the main source of interference in Fetal ECG (FECG)
monitoring. The MECG is detected at all electrodes placed on the mother’s skin (thoracic and
abdominal). In the case of multi-fetal pregnancies the traditional adaptive filtering technique
provides a “maternal clean “ signal consisting of the two fetal ECG signals. It cannot however
provide the desired individual FECG. The work reported here suggests the use of Blind
Source Separation (BSS) algorithm to deal with the problem. Simulation studies showed that
the BSS algorithm could perform separation of twins FECG. Real signals were recorded from
pregnant women at their 28-30th week of pregnancy, carrying 2-3 fetuses. The noise was
found to be too strong for the algorithm (and the naked eye) to notice any fetal heart signal.
The monitoring of FECG has clinical importance. If the physician could obtain a reliable
reading of the FECG, he could detect problems in the fetal heart activity even before he is
born. The procedure for obtaining the FECG should be non-invasive. The fetal heart is a small
heart so that the electrical current it generates is very low. In order to record the FECG,
electrodes are placed on the maternal abdomen as close as possible to the fetal heart. The
FECG may be acquired by placing a number of electrodes around the general area of the fetus
and hoping that at least one of the electrodes will have the FECG with high enough SNR.
Beside the problem of electrode placement, noise from electromyographic activity effects the
signal due to the fetus low voltage signal. Another interfering signal is the maternal ECG
(MECG) which can be 5-1000 times higher in its intensity and ability to induce surface
potentials [1]. The MECG effects all the electrodes, those that are placed on the chest
(thoracic electrodes), and those placed on the abdomen (abdominal electrodes) of the mother.
Because the FECG is a very weak signal, an electrode placed on the thorax of the pregnant
woman will hardly record any of it, if at all.
This fact implies that an adaptive
Abdominal electrode
cancellation algorithm may be employed.
An illustration of this conventional
Thoracic electrode
-
approach is given in figure 1.
that n sources are suffice to represent the
Σ
+ Estimated FECG
Adaptation
algorithm
Another approach to eliminate maternal
interference was suggested [2]. Assume
H(z)
Figure 1:Adaptive scheme for MECG
cancellation.
activity of all the bioelectric signals including the MECG and FECG. Define the source vector
as all the source signals si(t):
s(t ) T  [ s1 (t ), s2 (t ),, sn (t )]
(1)
Suppose that m measurement signals from surface electrode pairs are measured and arranged
in a vector x(t), called the measurement signal vector:
x(t ) T  [ x1 (t ), x2 (t ),, xm (t )]
(2)
Due to the resistive behavior of the body as a conductive medium in the frequency range of
0.5-100Hz there exist a linear memoryless relationship between x(t) and s(t). The problem can
then be formulated as :
x(t )  As(t )
(3)
Where A is an n×m constant real matrix. It can be assumed that all bioelectric sources are
mutually independent. By applying the algorithm for the BSS problem (JADE algorithm [3])
to the recorded data x(t) recovery of the source signals can be achieved yielding the source
signal of the FECG and the source signal of the MECG.
When comparing the results of the adaptive filtering technique to that of the BSS in the
single fetus case no apparent advantage was detected in favor of the BSS separation approach
[4]. This is however not the case with a multiple fetus pregnancy. Let us assume we are
dealing with the case of two fetuses i.e. twins. The thoracic electrodes still record only the
MECG but the abdominal electrodes record all three signals, the maternal ECG (MECG), first
fetus ECG (FECG1), and the second one (FECG2). The adaptive filtering approach can
separate the MECG from the fetus signals (FECG1 & FECG2), but can not separate FECG1
from FECG2 as both appear in all the abdominal electrodes. Using the Blind Source
Separation approach, theoretically, the problem can be solved.
In order to study the problem, simulation studies were performed. The coupling system
used (matrix A) was that of linear instantaneous combination form. Three ECG signals taken
from three persons (men or non-pregnant women) were chosen to simulate each source
(MECG, FECG1, FECG2). The MECG was about
1
ten times larger in amplitude from the FECG1,2.
0
Figure 2 shows the three source signals. One of
-1
(a)
0
0.1
the observations was simulated as a thoracic
0
electrode and the other two as abdominal ones. A
-0.1
0.5
1
1.5
2
2.5
3
3.5
4
0.5
1
1.5
2
2.5
3
3.5
4
0.5
1
1.5
2
2.5
3
3.5
4
(b)
0
0.1
(c)
mixing matrix, A, was chosen to simulate the
0
generation of the three observation vectors
(electrodes recording). Figure 3 shows the three
simulated electrode signals. Infinite SNR was
assumed. Applying the JADE algorithm to this
-0.1
0
Figure 2: The three source signals
used in the simulation, (a) MECG,
(b) FECG1, (c) FECG2.
1
problem yielded the three source signals, up to
sign and amplitude. The simulation results were
very encouraging suggesting the method could be
(a)
0
-1
0
0.5
0.5
1
1.5
2
2.5
3
3.5
4
0.5
1
1.5
2
2.5
3
3.5
4
0.5
1
1.5
2
2.5
3
3.5
4
(b)
0
applied to real signals.
-0.5
With cooperation from the Soroka medical
0.5
center, several pregnant women carrying two to
0
-0.5
three fetuses were sampled. The pregnant women
were in their 26th-28th week of pregnancy. In none
of the recorded signals, abdominal or thoracic
0
(c)
0
Figure 3:The three simulated
electrode signals, (a) Thoracic
signal, (b) first abdominal signal,
(c) second abdominal signal.
electrodes, the FECG was visually noticed. Even
when the location of the heart was determined by ultra-sound and the electrodes were placed
in its vicinity. We assume that the FECG of the fetus in the 28th week and less, is very low
and the signal to noise ratio does pose a severe problem. In order to verify the assumption that
no FECG was detected in real recordings due to a very low SNR, more simulation studies
were carried out. Three of the sources used were the three ECG signals used previously. Two
more sources were defined, one was white noise with uniform distribution and the second was
white noise with exponential distribution. The two noise sources were scaled to provide an
SNR of 25dB between the noise and the MECG signal. This ratio was constant throughout all
the simulations. The FECG signals were then scaled to various SNR’s. The gains used
(elements of the matrix A) were between 0.2 and 0.9 to simulate an electrode that is close to
the source (gain of around 0.85) and further away from the source (gain of around 0.25).
When the number of simulated electrodes was equal or more than the number of sources full
separation of the sources was achieved even when the SNR between the FECG and the noise
signals was as low as –25dB. In the case where the number of simulated electrodes (three)
was less then the number of sources (five) three situations have been noticed, depending on
the SNR between the FECG’s and the noise signals. From the definition of the JADE
algorithm the maximum number of sources that can be estimated from three observations
(electrodes) is three. So at least two of the estimated sources are expected to be a combination
of two or more sources. The first situation was for SNR of more then 15dB. In this case each
one of the estimated sources contained only one of the ECG signals and maybe some noise,
figure 4 shows the three outputs for this case. In the second case where the SNR was less
then 15dB the two FECG were not separated (The MECG was separated). In the third
situation where the SNR was less then –20dB and lower both FECG where not detected at all,
i.e. the algorithm managed to separate the
MECG from the noise but the FECG was non
existent in any of the estimations.
In conclusion the BSS algorithm may perform
with real recorded signals from a pregnant
woman with multiple fetuses in the condition
that the FECG’s signals are strong enough (SNR
10
(a)
0
-10
0
5
0.5
1
1.5
2
2.5
3
3.5
4
0.5
1
1.5
2
2.5
3
3.5
4
0.5
1
1.5
2
2.5
3
3.5
4
(b)
0
-5
0
5 (c)
0
between the noise and the FECG is high). The
simulation studies assessed the assumption made
that the SNR between the FECG’s and the noise
was low (lower than 15dB) in the real recordings
-5
-10
0
Figure 4: Output for SNR of 15dB.
(a) MECG. (b) FECG1 + noise.
(c) FECG2 + noise.
made from pregnant women in their 28th-30th week of pregnancy. Future work planned will
include pregnant women in more advanced stages of their pregnancy.
References
Adam, D., and Shavit, D. “Complete foetal ECG morphology recording by synchronized
[1]
adaptive filtration”, Medical and biological engineering and computing, 28, 287-292. 1990.
[2]
Cardoso, J.-F., “ Multidimensional independent component analysis”, Proc. of ICASSP
98. 4, 1941-1944. 1998.
[3]
Cardoso, J.-F., and Souloumiac, A., “Blind beamforming for non gausian signals”, IEE
Proc. –F, 140 (6), 362-370. 1993.
[4]
Kam, A. and Cohen, A., “Maternal ECG elimination and Foetal ECG Detection –
Comparison of Several Algorithms”, Proc. Of the 20th Ann. Int. Conf. IEEE EMBS, HongKong, 1998.
Download