J. Software Engineering & Applications, 2009, 2: 330-334
doi:10.4236/jsea.2009.25043 Published Online December 2009 (http://www.SciRP.org/journal/jsea)
Fetal ECG Extraction from Maternal Abdominal
ECG Using Neural Network
M. A. HASAN1, M. I. IBRAHIMY1, M. B. I. REAZ2
1Department of Electrical and Computer Engineering, International Islamic University Malaysia, Kuala Lumpur, Malaysia;
2Department of Electrical, Electronic and Systems Engineering, University Kebangsaan Malaysia, Selangor, Malaysia.
Received May 25th, 2009; revised June 25th, 2009; accepted July 2nd, 2009.
ABSTRACT
FECG (Fetal ECG) signal contains potentially precise information that could assist clinicians in making more appro-
priate and timely decisions during pregnancy and labor. The extraction and detection of the FECG signal from com-
posite maternal abdominal signals with powerful and advance methodologies is becoming a very important requirement
in fetal monitoring. The purpose of this paper is to illustrate the developed algorithms on FECG signal extraction from
the abdominal ECG signal using Neural Network approach to provide efficient and effective ways of separating and
understanding the FECG signal and its nature. The FECG signal was isolated from the abdominal signal by neural
network approach with different learning constant value and momentum as well so that acceptable signal can be con-
sidered. According to the output it can be said that the algorithm is working satisfactory on high learning rate and low
momentum value. The method appears to be exceedingly robust, correctly isolate the FECG signal from abdominal
ECG.
Keywords: Neural Network, FECG, Abdominal ECG, Heart Rate
1. Introduction
Fetal Heart Rate (FHR) analysis has become a widely
accepted means of monitoring fetal status. Currently,
Doppler ultrasound and FECG have proven to be reliable
techniques for monitoring FHR. The disadvantages of
Doppler ultrasound systems require intermittent reposi-
tioning of the transducer and they are only suitable for
use with highly trained midwifes. The use of Doppler
ultrasound (non invasive manner) is not suitable for long
periods of fetal heart rate monitoring [1]. In contrast,
methods utilizing the abdominal electrocardiogram
(AECG) have a greater prospect for long-term monitor-
ing of FHR and fetal well being using signal processing
techniques [2]. The fetal ECG is an electrical signal that
can be obtained non-invasively by applying a pair of
electrodes to the abdomen of a pregnant woman [3].
Sometimes the FECG is the only information source in
early stage diagnostic of fetal health and status. The
characteristics of the FECG, such as presence of signal,
rate, waveform and dynamic behavior are useful in de-
termining the fetal life, fetal maturity and existence of
fetal distress or congenital heart disease. Therefore, the
extraction of FECG signals from the abdominal ECG
signal with powerful and advance methodologies is be-
coming a very important requirement in biomedical en-
gineering. The ultimate reason for the interest in FECG
signal analysis is in clinical diagnosis and biomedical
applications. The fetal ECG contains potentially valuable
information that could assist clinicians in making more
appropriate and timely decisions during labour, but the
FECG signal is vulnerable to noise and difficulty of proc-
essing it accurately without significant distortion has
impeded its use [4]. A number of difficulties and com-
plication are associated with recording the abdominal
ECG. Electrical activity recorded from the maternal ab-
domen suppresses the FECG (the magnitude of the
FECG signal at the maternal abdomen is of the order of
several microvolts), which is a fraction of the MECG
amplitude recorded at the maternal abdomen. The ab-
dominal ECG contains a weak fetal ECG signal, a rela-
tively sound maternal ECG, maternal muscle noise (elec-
tromyographic activity in the muscles of the abdomen
and uterus) and respiration, mains coupling, and thermal
noise from the electronic equipment (electrodes, amplifi-
ers, etc.), power line interference (A/C) and Baseline
Wandering (BW). The signal processing algorithm needs
to remove the maternal ECG complexes, reduce the ef-
fects of motion artifact, muscle noise and power line in-
terface and then enhance the fetal QRS complexes before

Fetal ECG Extraction from Maternal Abdominal ECG Using Neural Network 331
they can be consistently detected. Therefore, to get pro-
per information of the fetal status and condition, it is
necessary to improve the SNR of the abdominal signal.
The MECG signal is the most predominant interfering
signal with FECG in the abdominal signal. The fre-
quency spectrum of each noise source partially overlaps
that of the FECG and therefore filtering alone is not suf-
ficient to achieve adequate noise reduction. Techniques
to get better FECG signal acquisition remain the subject
of on going research.
Although there are still limitations for extracting the
FECG signal from the abdominal ECG to monitor the fetal
status perfectly, currently, there is a significant amount of
effort being done to improve SNR of fetal ECG signal.
Conventional system reconstruction algorithms have vari-
ous limitations and considerable computational complexity
and many show high variance. Up to date advances in
technologies of signal processing and mathematical mod-
els have made it matter-of-fact to develop advanced FECG
extraction and analysis techniques. Ranges of mathemati-
cal techniques and Artificial Intelligence (AI) have ac-
knowledged comprehensive attraction. Mathematical
models incorporate wavelet transform, time-frequency
approaches, Fourier transform, statistical signal analysis
and higher order statistics. AI approaches towards signal
recognition include Artificial Neural Networks (ANN) [5],
Self-Organizing Map (SOM) neural network [6], Finite
Impulse Response (FIR) neural network [7] and fuzzy
logic system [8] a new technique combining the adaptive
noise canceller and adaptive signal enhancer in a single
recurrent neural network has been anticipated for the
processing of abdominal ECG signal [9].
In the field of fetal ECG extraction, various research
efforts have been carried out, including subtraction of an
averaged pattern, matched filtering, adaptive filtering,
orthogonal basis functions, fractals, temporal structure,
frequency tracking, polynomial networks, wavelets, and
real-time signal processing. Methods in fetal ECG for
extracting abdominal fetal ECGs have been recently in-
troduced for the monitor of fetal heart rate. So far, re-
search and extensive works have been made in the area,
developing better algorithms, upgrading existing meth-
odologies, improving detection techniques to reduce
noise and acquire accurate FECG signals to obtain reli-
able information about the fetus state thus assuring fetus
well-being during pregnancy period. A. K. Barros, et al.
(2001) discovered a semi-blind source separation algo-
rithm to solve the fetal ECG extraction problem [10].
This algorithm requires a priori information about the
autocorrelation function of the primary sources, to ex-
tract the desired signal (FECG). They did not assume that
the sources to be statistically independent but they as-
sumed that the sources have a temporal structure and
have different autocorrelation functions. The main prob-
lem with this method is that if there is fetal heart rate
variability, as is the case when the fetus is not healthy,
the a priori estimate of the autocorrelation function of the
fetal ECG may not be appropriate for the monitoring of
the fetal heart rate.
To enhance the extraction of FECG signal from the
abdominal ECG signal, in this paper, the NN-based
FECG extraction has been proposed. As the neural net-
work is adaptive to the nonlinear and time-varying fea-
tures of ECG signal therefore, the neural network has
been used to extract the FECG signal. Here, the adaptive
linear neural network has been considered with single
neuron. The input signal is considered as maternal ECG
and the target signal is abdominal ECG. Using this neural
network approach, the maternal ECG has been sup-
pressed from the abdominal ECG (maternal and fetal
ECG) by correlation detraction, so that the output can be
considered as only fetal ECG. Therefore, this paper is
paying attention for the accurate extraction of FECG
signal so that the correct decision can be made by the
clinician for well being of fetal during the pregnancy.
2. Factors Affecting the Abdominal ECG
Signal
The main source of interference is the maternal electrical
activity, the amplitude of which is much higher than the
amplitude of the fetus electrical activity, which is often
completely masked by the former. Beside this, the FECG
signals are often obscured by electrical noise from other
sources. Common ECG noise sources, such as power line
interference, muscle contractions, respiration, skin resis-
tance interference, instrumental noise, in addition to
electromyogram and electrohysterogram due to uterine
contractions, can corrupt FECG signals significantly [11].
The shape and structure of the FECG signal also depends
on the placement of the electrodes although there is no
standard electrode positioning for optimal FECG acquisi-
tion [12]. All of the aforementioned constraints make the
FECG extraction a difficult process. Therefore, it is im-
portant to understand the characteristics of the electrical
noise. Electrical noise, which will affect FECG signals,
can be categorized into the following types:
MECG Signal: Maternal ECG is the most predominant
interfering signal with FECG in the abdominal signal.
The frequency spectrum of this noise source partially
overlaps that of the ECG and therefore filtering alone is
not sufficient to achieve adequate noise reduction.
Electrode Contact Noise: Electrode contact noise is
transient interference caused by loss of contact between
the electrode and skin, which effectively disconnects the
measurement system from the subject. Electrode contact
noise can be modeled as a randomly occurring rapid
baseline transition, which decay exponentially to the
baseline value and has a superimposed 60Hz component.
The transition may occur only once or may rapidly occur
several times in succession.

Fetal ECG Extraction from Maternal Abdominal ECG Using Neural Network 332
RR Interval
RR
P
T
U
Q
S
T Peak
Amplitude QRS Amplitude
T Complex
ST Segment
ST Waveform
P Complex
PR Segment
PR Interval
QRS Complex
Figure 1. QRS complex in FECG signal
Motion Artifact: When motion artifact is introduced to
the system, the information is skewed. Motion artifact
causes irregularities in the data. There are two main
sources for motion artifact, Electrode interface and Elec-
trode cable. Motion artifact can be reduced by proper
design of the electronics circuitry and set-up.
Inherent Noise in Electronics Equipment: All elec-
tronic equipments generate noise. This noise cannot be
eliminated; using high quality electronic components can
only reduce it.
Ambient Noise: Electromagnetic radiation is the source
of this kind of noise. The surfaces of the human bodies
are constantly inundated with electric-magnetic radiation
and it is virtually impossible to avoid exposure to ambi-
ent noise on the surface of earth.
3. Clinical Importance of FECG Morphology
Biomedical signal means a collective electrical signal
acquired from any organ that represents a physical vari-
able of interest where the signal is considered in general
Figure 2. Tapped delay line
a function of time and is describable in terms of its am-
plitude, frequency and phase. FECG is a biomedical sig-
nal that gives electrical representation of fetal heart rate
to obtain the vital information about the condition of the
fetus during pregnancy and labor from the recordings on
the mother's body surface. The FECG signal is a com-
paratively weak signal (less than 20 percent of the
mother ECG) and often embedded in noise. The fetal
heart rate lies in the range from 1.3 Hz to 3.5 Hz and
sometimes it is possible for the mother and some of the
fetal ECG signals to be closely overlapping. The FECG
is very much related to the adult ECG shown in Figure 1,
containing the same basic waveforms including the P-
wave, the QRS complex, and the T-wave.
The PQRST complex as shown in Figure 1 it is com-
posed of three parts: The P-wave reflects the contraction
of the atriales. Secondly, the QRS-complex is associated
with the contraction of the ventricles. Due to the magni-
tude of the R-wave, it is extremely reliable. Finally, the
T-wave, which corresponds to the repolarisation phase
which follows each heart contraction. The delay associ-
ated to the R-R interval leads to the heartbeats frequency.
4. Methodology: Neural Network Architecture
The architecture of the neural network is mainly de-
signed by using the adaptive filtering approach that is the
combination of ADALINE (adaptive linear network) and
TDL (Tapped Delay Line). According to the concept of
TDL, the input signal (maternal ECG) enters and passes
through the N-1 delays and the output of the TDL is an
N-dimensional vector, made up of the input signal at the
current time, the previous signal, that is fed to the ADA
LINE shown in Figure 2. For the less complexity, the
value of N is considered 2. By combination of the TDL
and ADALINE network the adaptive filter network
shown in Figure 3. The maternal ECG, which is pre-
dicted and closely to the abdominal ECG, passes through
the 1 tapped delay line and the delayed output was multi-
plied by the two corresponding initial weights. After ad-
dition of the weighted output, it passes through the linear

Fetal ECG Extraction from Maternal Abdominal ECG Using Neural Network 333
Figure 3. Adaptive filter network
Figure 4. Neural network architecture with weight adjust-
ment
activation function. Finally, the output of the network
was detracted from the target input (abdominal ECG) and
to reduce the difference between input and target signal
the weight has been updated every step. Therefore, the
difference is considered the Fetal ECG as the abdominal
ECG contains the maternal and fetal ECG, and the ma-
ternal ECG has been suppressed from the abdominal
ECG. The overall architecture of the neural network is
shown Figure 4, where, the difference between the target
and input is error that is considered the expected outcome
and by using the error, the weight has been updated.
5. Result and Discussion
The initial weight was considered w1,1 = 0, and w1,2 =
-2. For input signal, 1000 data was fed into the network
that is considered the maternal signal and for the target
signal also 1000 data has been used as abdominal signal.
Initially, the learning rate and momentum has been taken
arbitrary. The changing of the learning rate and the mo-
mentum also affect the output of the network. According
to the out put it has been observed that the learning rate
is low that time the fetal signal was not reasonable but
the increased value of the learning rate, the suppressed
output from the target signal that is the fetal ECG was
very much satisfactory. Again, the effects of moments
also observed. If the low momentum value used in the
network that time the fetal signal contains some un-
wanted signal. Therefore, the high learning rate = 1, and
the low momentum value = 0.2 has been considered to
get the maximum satisfactory fetal ECG output that is
shown in Figure 5. According to the figure, at around the
Figure 5. Suppressed fetal ECG from abdominal ECG to maternal ECG

Fetal ECG Extraction from Maternal Abdominal ECG Using Neural Network 334
Figure 6. Overall system architecture
100 sample data, the QRS of the FECG signal is not clear
like the ordinary FECG QRS; sometimes it can be also
reasonable by changing the appropriate value of the
learning rate and the value of the momentum. Because, it
has been observed that the FECG signal after extracting
from the abdominal ECG to maternal ECG is greatly
affected by the value of learning constant and momentum.
6. Future Works
The work described here is currently on-going and is part
of a larger approach incorporating this utility. The de-
veloped algorithm is with the objective of implementing
in hardware. Therefore, the future work will involve
modeling the algorithm with VHDL and perform simula-
tion and synthesis for hardware implementation and
download into FPGA (Field Programmable Gate Array).
The overall system architecture for hardware implemen-
tation has shown in Figure 6.
7. Conclusions
FECG signal contains the valuable information that could
assist clinicians in making more appropriate and timely
decisions. The technique, adaptive neural network filter-
ing approach has been used to extract the FECG signal
from the abdominal ECG that is the agreeable output. In
brief, the accuracy of output depends on how many
variations of signals are used as input and the target in
the network. Furthermore, in this approach, learning rate
and momentum is also an important factor to affect the
desired FECG signal. By the observation the technique is
working high learning rate and low momentum value.
8. Acknowledgments
The authors would like to express sincere gratitude to the
Ministry of Science, Technology and Innovation of Ma-
laysia for providing fund for the research under eScien-
ceFund grant (Project No.01-01-08-SF0029).
REFERENCES
[1] M. Ungureanu, J. W. M. Bergmans, M. Mischi, S. G. Oei,
and R. Strungaru, “Improved method for fetal heart rate
monitoring,” Proceedings of the 2005 IEEE Engineering
in Medicine and Biology 27th Annual International Con-
ference, Shanghai, China, pp. 5916–5919, January 2006.
[2] P. P. Kanjilal, S. Palit, and G. Saha, “Fetal ECG extrac-
tion from single-channel maternal ECG using singular
value decomposition,” IEEE Transactions on Biomedical
Engineering, Vol. 44, pp. 51–59, 1997.
[3] T. Solum, I. Ingermarsson, and A. Nygren, “The accuracy
of abdominal ECG for fetal electronic monitoring,” Jour-
nal of Perinatal Medicine, Vol. 8, No 3, pp. 142–149,
1980.
[4] K. Karlsso, H. Lilja, K. Lindecrantz, and K. G. Rosen,
“Microprocessor based waveform analysis of the fetal
electrocardiogram during labor, ” International Journal of
Gynaecol and Obstetrics, Vol. 30, No. 2, pp. 109–16,
1989.
[5] M. B. I. Reaz and L. S. Wie, “Adaptive linear neural net-
work filter for fetal ECG extraction,” Proceedings of In-
ternational Conference on Intelligent Sensing and Infor-
mation Processing, Chennai, India, pp. 321–324, January
2004.
[6] G. Vasios, A. Prentza, D. Blana, E. Salamalekis, and P.
Thomopoulos, “Classification of fetal heart rate tracings
based on wavelet transform and self organizing map neu-
ral networks,” Proceedings of the 23rd Annual Interna-
tional conference of the IEEE, Istanbul, Turkey, Vol. 2,
pp. 1633–1636, 2001.
[7] G. Camps, M. Martinez, and E. Soria, “Fetal ECG extrac-
tion using an FIR neural network,” Computers in Cardi-
ology, Rotterdam, Netherlands, pp. 249–252, 2001.
[8] K. A. K. Azad, Z. M. Darus, and M. A. M. Ali, “Devel-
opment of a fuzzy rule-based QRS detection algorithm for
fetal and maternal heart rate monitoring,” Proceedings of
the 20th Annual International Conference of the IEEE En-
gineering in Medicine and Biology Society, Vol. 1, pp.
170–173, 1998.
[9] S. Selvan and R. Srinivasan, “A novel adaptive filtering
technique for the processing of abdominal fetal electro-
cardiogram using neural network,” Adaptive Systems for
Signal Processing, Communications, and Control Sympo-
sium 2000. (AS-SPCC), Louise, Alta., Canada, pp.
289–292, 2000.
[10] A. K. Barros and A. Cichocki, “Extraction of specific
signals with temporal structure,” Neural Computation,
MIT Press, Vol. 13, pp. 1995–2003, 2001.
[11] V. Zarzoso, A. K. Nandi, and E. Bacharakis, “Maternal
and foetal ECG seperation using blind source separation
methods,” IMA Journal of Mathematics Applied in Medi-
cine and Biology, Vol. 14, pp. 207–225, 1997.
[12] F. Vrins, C. Jutten, and M. Verleysen, “Sensor array and
electrode selection for non-invasive fetal electrocardio-
gram extraction by independent component analysis,”
Proceedings 5
th International Conference Independent
Component Analysis, Granada, Spain, pp. 1017–1024,
2004.