Computer Methods and Programs in Biomedicine 175 (2019) 193–204

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Extraction of foetal ECG from abdominal ECG by nonlinear

transformation and estimations
Rolant Gini John a,∗, K I Ramachandran b
a
Department of Electronics and Communication Engineering, Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, India
b
Center for Computational Engineering & Networking (CEN), Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, India

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: This paper proposes a simple yet effective method for the extraction of foetal
Received 18 February 2019 ECG from abdominal ECG which is necessary due to similar spatial and temporal content of mother and
Revised 13 April 2019
foetal ECG.
Accepted 20 April 2019
Methods: The proposed algorithm for extraction of foetal ECG (fECG) from abdominal signal uses single
channel. Pre-processing of abdominal ECG (abdECG) has been done to eliminate noise and condition the
Keywords: signal. The maternal ECG R-peaks have been detected based on thresholding, first order Gaussian differ-
Foetal ECG extraction entiation and zero cross detection on pre-processed signal. Having identified R-peaks and pre-processed
Maternal ECG signal as base, using Maximum Likelihood Estimation, one beat including QRS complex morphology of
Abdominal ECG
maternal ECG (mECG) has been constructed. Extraction of maternal ECG from abdECG is done based on
Foetal heart rate
the constructed beat, R-peak locations and its corresponding QRS complex of abdECG. Extracted mECG
Maximum Likelihood Estimation
Non-invasive extraction has been cancelled from abdECG. This results in foetal ECG with residual noise. The noise has been re-
duced by Polynomial Approximation and Total Variation (PATV) to improve SNR. This approach ensures
no loss of partially or completely overlapped fECG signals due to mECG removal. The algorithm is tested
on three database namely daISy (DBI ), Physiobank challenge 2013 (DBII ) and abdominal and direct foetal
ECG database (adfecgdb) of Physiobank (DBIII ).
Results: The algorithm detected no false positives or false negatives with certain channel for DBI , DBII
and DBIII which shows that the proposed algorithm can achieve good performance. Overall accuracy and
sensitivity of the system is 98.53% and 100% for DBI . Best accuracy and sensitivity of 97.77% and 98.63%
are obtained for DBII . Best accuracy of 92.41% and sensitivity of 93.8% are obtained for DBIII . Correlation
coefficient between actual foetal heart rate (fHR) and estimated fHR of 0.66 for DBII and 0.59 for DBIII
is obtained. The method has obtained overall F1 score of 99.25% for DBI , 96.04% for DBII and 94.25% for
DBIII . It has obtained a best MSE of fHR and overall MSE of R-R interval which is 10.8bpm2 and 2.2 ms for
DBII , 12bpm2 and 2.14 ms for DBIII .
Conclusion: The results for different public databases show that the proposed method is capable of pro-
viding good results. The foetal QRS, R-peaks and R-R intervals have also been obtained in this method.
Thus, it gives a significant contribution in the required area of research.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction
Invasive methods are not recommended during prenatal exam-
ination despite the high accuracy it provides. Therefore, the most
resorted practice is the extraction of foetal ECG (fECG) using non-
invasive methods. Foetal ECG extraction has been an area of inten-
sive research. Since foetal ECG is acquired by non-invasive meth-
ods, it is found to be contaminated with noise such as maternal
ECG (mECG), uterine contraction, very weakly propagated signal
∗
Corresponding author.
E-mail address: j_rolantgini@cb.amrita.edu (R.G. John).
from the foetus due to obesity of mother in addition with com-
mon noises encountered by any other bio signals. Accurate detec-
tion of the fECG and foetal heart rate (fHR) is the need of the hour
to reduce the foetal mortality, which is mainly caused due to heart
related birth defects. Foetal heart rate is one of the most impor-
tant factors that need to be monitored precisely, to ensure that
the foetus remains healthy. General causes of Intra Uterine Foetal
Death (IUFD) include congenital anomaly, anaemia, cord accidents,
oligoamnios, etc., [1]. Congenital anomaly can be diagnosed and
treated with the aid of electro cardio gram (ECG) which shows
the manifestation in the morphology of cardiac electrical signal
when there is a cardiac defect. ECG is also believed to carry much

https://doi.org/10.1016/j.cmpb.2019.04.022
0169-2607/© 2019 Elsevier B.V. All rights reserved.
194 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
more information than conventional sonography methods to detect
congenital heart disease (CHD) [2,3]. Extraction of fECG by non-
invasive methods has been a field of intense study. Although dif-
ferent kinds of algorithms exist to obtain the fECG, the complex-
ity associated is high. The fECG detected during the early stages of
pregnancy is critical in detecting congenital diseases in the foetus.
It is an exacting task to obtain the fECG during the early stages
of pregnancy but the sooner the information has been extracted
and analysed, the better the chances of giving proper treatment
and delivering a healthy baby. ECG also gives information such as
the size and position of the heart chambers, the presence of any
impairment in conduction system or in the muscles of the heart
[4]. Invasive methods were implemented by placing electrodes on
the scalp of the foetus which may be risky for foetus and mother’s
health and also cannot be used during regular pregnancy tests [5].
In order to avoid such risks, the focus has been shifted to non-
invasive methods of fECG extraction which is a safer method. Al-
though several sophisticated algorithms exist to extract the foetal
ECG, obtaining a clear foetal ECG signal still remains a challenging
task since the mother and foetal heart beat possess similar tem-
poral and spatial content. The required ECG of mother and foe-
tus lies in the same frequency range where simple filtering cannot
extract the foetal ECG from the abdominal ECG. This paper pro-
poses a simple algorithm that extracts the fECG, even when there
is a possibility of partially or fully overlapped fECG with mECG.
The purpose of this algorithm is to maintain the accuracy of the
signal obtained and reduce the complexity to a sufficient extend.
Careful attention has been paid to avoid false positives and false
negatives.
The proposed method eliminates the maternal signal and ob-
tains the foetal signal by removing remaining noise signals. The
noise signals present in the abdominal ECG (abdECG) are ma-
ternal electrocardiogram, Electro magneto gram (EMG) from the
chest wall, signal weakening due to obesity of mother or in-
sufficient growth of the foetus or vernix caseosa (highly non-
conducting layer) [6] which formed around the foetus, electrode
contact noise, power line interference, baseline wandering, instru-
mentation noise, electro surgical noise, random electronic noise,
low signal to noise ratio of fECG compared to mECG [7] etc.,. One
of the major aspects to keep an eye out for is when the fECG is
buried under the mECG. Even in this case, the algorithm detects
the presence of fECG signal and efficiently extracts it out. The fHR
obtained by this method is accurate. Thus, the proposed method
can also be employed to diagnose abnormalities in the foetus dur-
ing prenatal examination. The rest of the paper is structured in the
following manner: Section II gives the literature survey; Section III
describes the proposed method; Section IV describes the database
used; Section V deals with results; Section VI deals with discus-
sions and it is followed by conclusion in Section VII.
2. Background of fECG extraction
Several methods that have been employed for the extraction of
fECG have been studied, some of which make use of single chan-
nel electrocardiogram and some employ multi-channel data. Some
of the methods that have been proposed include the event syn-
chronous canceller algorithm [8], independent component analy-
sis (ICA) [9], blind source separation [10], singular value decom-
position (SVD) [11] and adaptive noise canceller [12], extended
kalman filter [13], fast independent component analysis (fast ICA)
[14], echo state neural network (ESN) based filtering [15], SVD and
smooth window (SW) combined [16], Signal quality assessment
(SQA) and fine-tuning for maternal ECG (FTM) [17], synchronized
averaging and ICA [18], FUSE algorithm [19] (combination of subset
of ICA, template subtraction and Kalman filters), optimal principal
components [20], wavelet-based approach [21], Principal compo-
nent model [22], extended Kalman smoother (EKS) and template
adaption (TA) method [23]. Each of the above mentioned meth-
ods concentrates on different performance metrics. In the event
synchronous canceller algorithm, the repetitive disturbing noise is
removed by the subtraction of a “cycle template” in an adaptive
manner. An indicator signal helps to remove the mECG signal. Al-
though the event synchronous canceller algorithm cancels out the
noise effectively, the background noise is almost ignored [8]. Also,
it requires a separate reference channel. The ICA algorithm on the
other hand, is good but requires the use of higher order statistics
and high demanding in terms of computations [9]. Also, the se-
lection of channel plays an important role that impacts the fECG
extraction. The next algorithm, blind source separation is robust
and successful but at the cost of very high computational complex-
ity. It extracts the required information without any prior knowl-
edge of how the signals were mixed. A large number of samples
are required to reach accuracy, at a relatively high cost. Hence the
method is expensive [10,24].
The adaptive noise canceller algorithm, like the event syn-
chronous noise canceller requires an additional reference channel.
Also peak detection cannot be done directly; as a result of which,
an additional algorithm is required. Single channel methods such
as polynomial network techniques for fECG extraction have also
been studied [25], which can be used to non-linearly map the tho-
racic ECG to the abdominal ECG signal and subtracting one from
the other. This method proposes a way to separate both overlap-
ping as well as non-overlapping fECG and mECG beats in a non-
iterative manner. However, the results are not practically effective
in the sense that the background noise including muscle noise and
uterine contractions are completely ignored. Another method that
was studied includes a simple technique for mECG cancellation
with the help of R-peaks detection algorithms [26]. Although the
algorithm employed is based on simple cancellation and is effec-
tive, overlapping fECG peaks cannot be detected using this method.
Even though several methods have been suggested for non-invasive
foetal ECG detection, this field still remains an area where exten-
sive research needs to be done to become a part of regular clinic.
Either the singular value decomposition method or combined
with smooth window [16] has been tested to detect the foetal ECG
accurately even in a moderately high noise environment. The pro-
cessing speed for this algorithm is very low and it is computation-
ally complex [11]. Fast ICA [14] reduces computational complexity
and fast compared with ICA. But the ambiguity exist in deciding
the number of Eigen vector which corresponds to the mECG based
on the higher difference between the Eigen values of the data co-
variance matrix in the pre-processing, number of iterations in post
processing in addition with the usage of all eight channel’s data.
ICA is combined with other techniques like synchronized averag-
ing [18] for improved performance. There are hardware implemen-
tations of the certain methods such as STAN monitors and Mon-
ica AN24 [27], MindChild MERIDIAN Monitor [28] and Newer non
voltage foetal heart monitoring etc., [29,30].
STAN monitor improved the monitoring efficiency by incorpo-
rating a proxy of the foetal ST-segment and there by improved the
detection accuracy. The STAN monitor is clinically successful but is
a failure in business. This is mainly due to the fact that it is not
non-invasive and that more features of ECG waveform can be ex-
ploited [27].
Another type is the Monica AN24 monitor. The major challenge
faced by it, is overwhelming of maternal ECG on foetal ECG wave-
form, which has a voltage 100 times greater than its tiny foetal
counterpart. In addition, ambient electrical noise generated by ma-
ternal movement, other nearby electrical devices and the lights
overhead also overwhelm the foetal ECG signal [28]. Algorithms
to filter out the maternal ECG signal and ambient noise have
been developed over past 30 years. But in recent days only, the
 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
microprocessors have become capable of running these algorithms
in real time and in an affordable manner [29].
A simple approach to identify the fECG is being discussed in
this paper. Using this approach, the overlapped fECG were effec-
tively identified with lesser computational steps. This approach
also provides a degree of freedom in extracting the fECG from
abdECG wherever the electrode has been placed in the abdomen.
Although this method approaches the problem using a single
channel without the reference channel or data from additional
channels, this method yields very positive results in the extrac-
tion of fECG in the overlapped and non-overlapped cases. This
approach was tested on three different publicly available database
namely daISy database [31] (denoted as DBI in the discussions),
Physiobank challenge 2013 dataset [32] (denoted as DBII in the
discussions) on fourteen recordings and abdominal and direct
foetal ECG database (adfecgdb) of Physiobank [33] (denoted as
DBIII in the discussions) on all recordings.
3. Methodology
The complete flow of the proposed method has been repre-
sented with a block diagram as shown in Fig. 1.
3.1. Pre-processing
The channel under consideration for foetal ECG (fECG) extrac-
tion is corrupted with noise which has to be reduced for good
accuracy. Hence, pre-processing is done. This is the first step in
the algorithm where the considered single-channel abdominal ECG
(abdECG - represented as a[n]) to extract the required information
is conditioned by partially removing the noise.
A 15th order band pass filter f[n] (BPF) is designed with cut
off frequencies that match the signal frequencies [34]. This filter is
designed in order to eliminate a fair amount of baseline wander as
well as muscle noise which is in lower and/or higher frequencies.
It also eliminates the 50 Hz power-line noise [12]. The frequency
descriptions of the filter used based on the signal and the filtering
are as follows:
Fs (Sampling Frequency) =250 Hz for DBI , 1 KHz for DBII and
DBIII
f_cl (lower cut off frequency) = 5 Hz
f_cu (upper cut off frequency) = 20 Hz
x [n ] = a [n ] ∗ f [n ]
Fig. 1. Block diagram of fECG extraction. abdECG: abdominal ECG and represented as a[n], mECG: Maternal ECG, FOGD: First order Gaussian Differentiator, ZC: Zero Crossing,
MLE: Maximum Likelihood Estimation, PA: Polynomial Approximation, TV: Total Variation, fECG: foetal ECG, fHR: foetal Heart Rate.
195
Filtered signal x[n] is subjected to normalisation process. Nor-
malisation is done in order to limit the amplitudes of the signal at
unity in accordance with the highest amplitude. The transformed
signal will have the same energy at every scale. When the signal
is subjected to further filtering and processing, the signal tends to
change the amplitude which can be avoided by this. It also facili-
tates a standard way of comparison [26]. The filtered signal x[n] is
normalised as specified in (2) to get x˜ [n].
x [n ]
x˜ [n] = (2)
max N
n=1 (|x[n]| )
The abdECG of channel 1 in DBI database which is used as
the input signal is shown in Fig. 2(a) and the normalised signal is
shown in Fig. 2(b) after normalising. Basic noise removal has been
done with a BPF to remove the baseline wander and filter out the
high frequency noises as stated earlier. Fig. 2(a) and Fig. 2(b) are in
different amplitude range, which cannot be compared directly. The
noise is not completely removed at this stage since there is a pos-
sibility of fECG lies in the amplitude level of noise which may get
eliminated. In this method, Noise removal has been concentrated
to improve signal to noise ratio only after the removal of maternal
ECG (mECG).
3.2. R-peaks detection
The normalized signal ͠x [n] is subjected to thresholding and
first order Gaussian differentiation (FOGD) process for detecting
the maternal ECG R-peaks.
A threshold is set according to the nature of the signal because
the amplitude of mECG has very high magnitude compared to the
fECG. Therefore this technique consists of two different threshold
limits. One threshold limit for mother and the other is for foetus.
Part of the signal above the threshold is taken into account as such
and it is the nonzero information which forms the Region of Inter-
est (ROI). Remaining part of the signal has been nullified.
The threshold value has been assumed as a percentage of the
maximum value of the signal to make R-peak detection adaptive.
First threshold (th_m) can be set at 70% of the maximum value of
the normalised abdECG for maternal R-peak detection if the signal
is not too noisy. The general format of th_m is as given below in
(3):
N
(1) th_m = (0.6 ± 0.1 )xmax(x˜ [n] ) (3)
n=1
196 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
Fig. 2. (a) Channel 1 abdECG (input signal: a[n]) of DBI ; (b) Normalized and filtered abdECG: ͠x [n]. Y-axis represents the amplitude (μV).
The thresholding process generates ͠x [n] which consists of the
signal only above th_m as specified in (4).
 mum values of R lie with the noise. The detected R-Peaks and the
0, x˜[n] < th_m

x[ n ] =
x˜[n], otherwise
The ROI at the end of thresholding process gives the possible
locations of R-peaks of the mother from the abdECG and is given
by (5).
ROI = {xˆ [n]|xˆ [n] = 0}
The second threshold which is used in R-peak identification of
the foetus has to be set at an optimal percentage of the maximum
value of the fECG to avoid false negatives (FN) as well as false pos-
itives (FP) which influences the sensitivity and accuracy of fECG
detection. To find the R-peaks of the mother in the ROI, the local
maxima [35] can be found as shown in (6) and the same location
has been tweaked up based on the zero crossing point after FOGD
[34].
N1
m=1 = max {x
m_loc1M ˆ m [i]}
i=1
where m is the number of non-zero envelops in the signal after
thresholding and i is the number of samples consist in every non-
zero envelope. At every ROI, the local maximum has been found
which gives the indicative R-peak location. To find the exact R-peak
(denoted as r[n]) location in the narrowed area around m_loc1,
the signal is passed through a FOGD and the zero crossing (ZC) of
the outcome has been studied. The zero crossing point in the nar-
rowed ROI gives the exact location of the R-peaks of the mother.
The FOGD will change the peak point into a transition point from
positive going pulse to negative going pulse. This has been a part
of the nonlinear transformation.
The FOGD function {wd [m]; m = 1,2,3,…M1 -1} is gener-
ated using a differentiated M1 -point Gaussian window {w[m];
m = 1,2,3,…, M1 }. The convolution of signal near the region of
m_loc1 (i.e., r[n] and wd [m]) generates signal ͠wd [m] and is given
by (7).
∞
 responding beats in ͠x, constructed mECG beat and its peak loca-
˜ d [m] =
w wd [k]r[m − k]
k=−∞
This process retains the higher frequencies and attenuates the
lower frequencies. This results in retaining the QRS due to its high
frequency and thereby able to identify the R-peaks. The outcome
͠wd [m] of FOGD passes through the ZC detector to get the exact lo-
cation of R-peaks in abdECG and is denoted as m_loc. FOGD and ZC
detector after local maxima help in finding the exact R-peak loca-
tion even if it is a negative going QRS complexes where the maxi-
(4) signal surrounded have been used [36] for Maximum Likelihood
Estimation (MLE) of mECG to construct single beat of the mother.
3.3. Maximum Likelihood Estimation (MLE)
Maximum Likelihood Estimation (MLE) has been performed in
(5) order to generate a reference signal. The signal is subjected to MLE
in the region of each heart beats in order to obtain a reference
PQRST complex of the mECG. Single beat has been constructed us-
ing MLE, based on the available abdECG nature to obtain the ref-
erence signal. MLE is efficient for data having large samples and it
extricates all possible information from the data. This way of esti-
mation does not depend on any additional reference data. Essen-
tially, MLE maximizes the likelihood function. Respective samples
of every beat (denotes as y) have been consolidated as an array to
estimate each value of PQRST complex as shown in (8).
1+k2+1,M k2,m_loc M
yki=1 = x˜ [j + k]k=−k1, j=[ m]_loc[1] (8)
(6) ,m=1
where k1 and k2 are the number of samples to be considered be-
fore and after the R-peak to construct PQRST. MLE is obtained by
maximizing the log-likelihood function, lnL(y|x) [37]. This is be-
cause the two functions, lnL(y|x) and L(y|x) are monotonically re-
lated to each other. Thus the same MLE estimate has been obtained
by maximizing anyone.
From the consolidated arrays of y, the maximum likelihood yˆ
has been estimated for each values of PQRST as shown in (9). The
constructed beat has been shown in Fig. 3(c). In Fig. 3, x-axis is
Sample number and y-axis is Amplitude (in NU – No Unit).
 
yˆ ki=1
1+k2+1
= θˆ yi [m]M
m=1 = arg maxL (θ |yi ) (9)
θ
yˆ gives the constructed beat of the mother. Thus with the iden-
tified R-peaks, MLE has been used to form a PQRST complex by
considering the information around the R-peaks of every beat of
the signal and the likelihood exist between them.
Using m_loc, normalised signal ͠x, R-peak amplitude of the cor-
(7) tion, mECG has been constructed as shown in (10).
Y[i + m]km2=−k1 = yˆ [m + k2 + 1] where i
= {m_loc[1], . . . .., m_loc[M]} (10)
Thus MLE provides a suitable method to model mECG, so it is
a generic principle which is not patient centric. The constructed
mECG (Y) is shown in Fig. 3(a).
 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
Fig. 3. DBI channel 1:- (a) Constructed reference of mECG by MLE; (b) Extracted fECG with R peaks marked; (c) Constructed Single beat of mECG.
3.4. Cancellation of mECG from abdECG
The constructed mECG by ML estimate has been subtracted
from normalised abdECG signal as shown in (11). The resultant
signal is a combination of fECG and noise. Thus mECG has been
removed from normalised abdECG and is shown in Fig. 3(b). This
method culls out the fECG signal even when it is buried in ab-
dECG due to complete or partial overlap with the mECG and is
not clearly visible. Thus this algorithm efficiently and accurately
detects the presence of fECG and extracts it out by avoiding false
positives and false negatives.
zi = x˜ i − Yi ; i = 1, 2, 3, . . . N
3.5. Noise removal using polynomial approximation and total
variation
In polynomial approximation and total variation (PATV) algo-
rithm [38], piecewise constant signal and polynomial signal have
been estimated simultaneously from a noisy additive mixture. The
signal corrupted with noise is represented as follows
z(n ) = p(n ) + b(n ) + w1 (n ); n = 0, . . . , N − 1
In Eq. (12), the noisy signal z(n) has been expressed as a com-
bination of low order polynomial p(n) of order d  N, approx-
imately piecewise constant b(n) and stationary white Gaussian
noise w1 (n). The signal z consists of smooth information which
gives envelop of the signal, additive step discontinuity information
and noise. S which is cumulative sum matrix, u of length (N-1)
where u ∈ RN-1 and orthogonal basis (G) (also known as vander-
monde matrix) of degree d polynomials have been calculated in
an iterative process from z. Finally polynomial coefficients (c) and
the signal x1 have been found out by (13) and (14).
x1 = S u
c = GT ( z − x1 )
Using c, polynomial approximation of the signal z has been cal-
culated. Using the polynomial approximation and x1 , the noiseless
signal ͠z has been found out. The noise removed signal is shown in
Fig. 4(b) where Fig. 4(a) shows the extracted fECG signal. Vertical
axis in Fig. 4 gives the amplitude in relative scale.
3.6. Post-processing of the fECG, R-peaks detection and fHR
determination
The obtained noiseless signal ͠z from PATV gives the noiseless
fECG and can do some filtering if required. Obtained fECG is then
197
subjected to the same process of R-peaks detection which was fol-
lowed in the mECG as shown in Fig. 1. Thresholding, FOGD and ZC
detection are done and all the R-peaks of fECG are detected. The
threshold used for foetal ECG is given by th_f as in (15). Threshold-
ing process used in fECG R-peaks detection is given by the below
pseudo code.
N
th_f = (0.4 ± 0.2 )xmax(z˜ [n] ) (15)
n=1
Algorithm for thresholding in foetal ECG:
(11) Inputs:
fet_ECG(extracted ͠z) = mECG and noise removed abdECG
N = Length of fet_ECG
Outputs:
fet_peak_Region // Select region of interest to identify foetal R peak
Begin
threshold_foetal ≈ 40% of ͠z maximum
fet_peak_region =1 to N
for i = 1 to N, increment by 1
If fet_ECG > threshold_foetal
fet_peak_region = fet_ECG // Marking the region of interest
end
(12) End
As stated earlier, the threshold value to be used in mECG and
fECG may vary since the fECG can be a signal of magnitude up
to 100 times smaller than the mECG. A threshold value of ≈ 40%
is used here. If required, further conditioning and elimination of
noise in the resultant fECG can be done by another band pass fil-
ter based on the signal. The signal obtained is subjected to R-peaks
detection and then foetal heart rate (fHR) and R-R intervals are de-
termined. It is observed that a foetal R-peak is detected even in
regions where fECG was buried in noise and was unable to be no-
ticed by the naked eye. Fig. 3(b) shows the extracted fECG signals
(13) with R-peaks marked. Extracted fECG signal, noise removed signal
using polynomial approximation and total variation with R-peaks
marks has been shown in Fig. 4. The foetal heart rate is calculated
(14) as [10] in (16).
Number of peaks detected
f HR = ∗ 60 (16)
Duration of the signal
4. Database
Three databases have been used in this work. The first database
is daISy database which is an eight channel recording with
250 Hz sampling rate. It has a recording from mother’s thorax,
abdomen and its related information. First five channel record-
ings are around the abdomen of the mother (denoted as channel
1 - 5) and the next three channel recordings are in the thorax
198 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
Fig. 4. (a) Extracted fECG of channel 1; (b) Noise removed fECG using PATV with R-peaks marked of channel 1 from DBI .
(channel 6 – 8). Out of five abdomen recording, only first three
abdomen channels provide clear information about the foetus. It
shows that those channel electrodes are near the foetus and re-
maining channel electrodes are far from the foetus to pick up the
lower strength signal. It has been denoted as DBI in the discus-
sions.
The second database used is Physionet challenge 2013 data set-
a. It has initial record of 25 and supplementary record of 50. Each
recoding is done for one minute at the sampling rate of 10 0 0 Hz. It
consists of time, four channel recordings (AECG1 – AECG4) of ECG
in the abdomen of the mother and annotations for the same. Some
fourteen recordings of this have been used in the proposed work to
check its efficiency. It has been denoted as DBII in the discussions.
The third database used is Abdominal and Direct Foetal ECG
Database (adfecgdb) of physiobank. It has 5 recordings of women
in labour during 38 to 41 weeks of gestation. Each recoding is done
for five minute at the sampling rate of 10 0 0 Hz with resolution of
15bits. It has four channel recordings (denoted as Abdomen_1 to
Abdomen_4) of ECG around the naval, one direct ECG recording of
foetus (scalp electrode which is represented as Direct_1) and anno-
tations for the same. All five recordings of this have been used in
the proposed work to check the efficiency. This database has been
denoted as DBIII in the discussions.
5. Results
The foetal ECG and its related information have been ex-
tracted effectively from an abdominal signal by using the proposed
method. This proposed method has been implemented using MAT-
LAB and the results are as follow.
A detailed analysis of channel 1, 2 and 3 of DBI (in Fig. 5) have
been done since only those electrodes have clear presence of fECG
out of 8 electrode signals. Channel 3 of DBI has been displayed
first since it is the most corrupted signal by noise compared to the
other two channels. Fig. 5(a) shows the input signal (the abdECG of
channel 3) with extracted foetal beats (from the respective channel
inputs) numbered and R-peaks locations indicated with circle (in
red). Similarly Fig. 5(b) is for channel 2 and Fig. 5(c) is for Channel
1. x-axis represents sample number and y-axis represents ampli-
tude (μV) in Fig. 5.
Table 1 gives the condensed summary of effectiveness of the
technique for DBI . This method can achieve 100% accurate detec-
tion and sensitivity (can be observed from Fig. 5) by critically op-
timising the algorithm to find the R-Peaks of the fECG which may
change in dynamic estimation. The actual R-peaks have been com-
pared with the R-peaks detected by the algorithm. False positive
(FP) peaks and false negative (FN) peaks have been categorised as
defects. FP peaks are the peaks that are not pertain to the foetal
and detected as foetal R-peak. FN peaks are the peaks that are per-
tain to the foetal R-peaks and not detected as foetal R-peak. Truly
detected (TD) peaks are the foetal R-peaks that are present and de-
tected. The accuracy calculation has been done with a false positive
consideration for channel 3 input which was avoided when worked
by professional and may not be possible in real time extraction.
Accuracy, sensitivity and other performance metrics are calcu-
lated using formulas given below in Eqs. (17)–(20).
TD
Accuracy = ∗ 100 (17)
(TD + FN + FP)
TD
Sensitivity = ∗ 100 (18)
(TD + FN )
TD
PPV = ∗ 100 (19)
TD + FP
PPV ∗ Sensitivity 2 ∗ TD
F1 = 2 ∗ = ∗ 100 (20)
PPV + Sensitivity 2 ∗ TD + FN + FP
Positive Predictive Value (PPV) and F1 score are American Na-
tional Standards Institute/ The Association for the Advancement
of Medical Instrumentation (ANSI/AAMI) guideline for performance
measurement [16]. F1 gives the overall probability of correct fECG
detection and its QRS complex. It is used as an alternate for accu-
racy measurement.
The foetal heart rate has also been calculated as 132 bpm (beats
per minute) for DBI and has been found to lie within the range of
120 to 160 bpm [36], which is the normal average range of foetal
heart. Noise analysis for the same database is shown in Fig. 6.
A detailed analysis of noise reduction has been shown in
Fig. 6 for all channels of DBI to show that polynomial approxima-
tion and total variation (PATV) improves the fECG signal as well as
its detection. PATV depends on various parameters like number of
iterations (N1 ), λ, μ0 and μ1 which can influences the convergence
of the algorithm to the specified solution [39]. The scalar param-
eter λ has to be initialised to control the trade-off between signal
distortion and noise reduction. μ0 and μ1 are positive scalar pa-
rameters which affect the convergence speed of the solution but
do not affect the solution of the problem. The effect of various pa-
rameters on the signal to noise ratio (SNR) for DBI can be studied
from Fig. 6. Best SNR achieved is 14.7 dB. Peak detection and fECG
extraction of other two databases are shown in Figs. 7 and 8.
Fig. 7 shows the input abdECG of a channel from a01 and
a04 recordings with its detected R-peaks and annotated R-peaks
 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204 199

Table 1
Accuracy and sensitivity for DBI .

Ch. no Total peaks Peaks detected FP FN TD Accuracy (%) Sensitivity (%)

1 22 22 0 0 22 100 100
2 22 22 0 0 22 100 100
3 22 22 1 0 22 95.6 100

Fig. 5. DBI with identified foetal R-peaks marked (by red circle):- (a) Channel 3 abdECG; (b) Channel 2 abdECG; (c) channel 1 abdECG.

Fig. 6. Number of iteration (N1 ) Vs. SNR: (a) Channel 1; (b) Channel 2; (c) Channel 3; (d) λ Vs. SNR for N1 =30.
200 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204

Fig. 7. DBII signal with identified foetal R-peaks and annotated foetal R-peaks marked: (a) AECG1 channel of a01 signal; (b) AECG4 channel of a04 signal.

Fig. 8. DBIII database: - a) Direct foetal ECG of r01 with identified R-peaks from abdomen_4 marked; b) Abdomen_4 channel of r01 with identified foetal R-peaks marked.

marked in it. It shows that the identification of the foetal peaks
is correctly done for DBII . Fig. 8(a) shows the scalp electrode sig-
nal of r01 recording with the identified foetal R-peaks from an ab-
domen electrode (Abdomen_4 channel) marked on it. It shows that
the R-peaks have been identified correct from the abdomen signal.
Fig. 8(b) shows Abdomen_4 channel recording of the same data
from which the extraction has been done. The identified R-peaks
from the extraction also marked on the same signal. It proves that
the algorithm is capable of extracting the fECG as well as its re-
lated information for DBIII .
The accuracy, sensitivity and other parameters of DBII and DBIII
for some duration have been listed in Tables 2 and 3. It can be

Table 2
Accuracy, sensitivity, fHR and R-R interval for DBIII .

Data set Name Abdomen Lead Name TD FP FN Accuracy (%) Sensitivity (%) fHR (in bpm) R-R Interval (in seconds)

r01 Abdomen_1 20 3 2 80.00 90.91 120 0.5

Abdomen_2 19 2 3 79.17 86.36 114 0.53
Abdomen_3 22 2 0 91.67 100 132 0.45
Abdomen_4 21 0 1 95.45 95.45 126 0.48
r04 Abdomen_4 19 0 2 90.48 90.48 114 0.53
r07 Abdomen_1 19 3 1 82.61 95.00 114 0.53
r08 Abdomen_1 19 0 3 86.36 86.36 114 0.53
r10 Abdomen_1 21 1 1 91.30 95.45 126 0.48
Abdomen_2 21 2 1 87.50 95.45 126 0.48
Abdomen_4 20 3 2 80.00 90.91 120 0.5
 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204 201

Table 3
Accuracy, sensitivity, fHR and R-R interval for some recordings of DBII .

Data set name Abdomen lead name TD FP FN Accuracy (%) Sensitivity (%) fHR (in bpm) R-R Interval (in sec)

a01 AECG 1 22 0 0 100 100 132 0.45

a02 AECG 1 22 2 5 75.86 81.48 132 0.45
a03 AECG 4 21 0 1 95.45 95.45 126 0.48
a04 AECG 4 21 0 0 100 100 126 0.48
a05 AECG 4 20 1 1 90.91 95.24 120 0.50
a06 AECG 1 23 1 2 88.46 92.00 138 0.43
a07 AECG 1 20 2 2 83.33 90.91 120 0.50
a08 AECG 1 21 0 0 100 100 126 0.48
AECG 4 21 0 0 100 100 126 0.48
a09 AECG 1 17 1 5 73.91 77.27 102 0.59
a10 AECG 1 27 0 0 100 100 162 0.37
a11 AECG 1 23 0 0 100 100 138 0.43
a12 AECG 1 23 0 0 100 100 138 0.43
AECG 4 23 0 0 100 100 138 0.43
a14 AECG 1 21 0 0 100 100 126 0.48
AECG 2 21 0 0 100 100 126 0.48
a15 AECG 1 22 0 1 95.65 95.65 132 0.45
AECG 4 23 1 0 95.83 100 138 0.43

Table 4
Occurrence of false positives/ negatives and occurrence of complete and partial overlap of DBI :- Analysis.

Channel Possible Error / Overlap Occurrences Avoided False Detection of fECG Detection despite Partial/
number Positive/False Negative (Yes/No) Complete overlap
Beat FP/ FN avoided

3 1 Yes False Positive – –

3 2 Yes False Negative – –
3 5 Yes False Positive Yes Partial Overlap
3 10 – – Yes Partial Overlap
3 12 Yes False Positive – –
3 17 Yes False Positive Yes Complete Overlap
3 19 Yes False Positive – –
3 20 – – Yes Partial Overlap
3 22 – – Yes Partial Overlap
2 3 – – Yes Partial Overlap
2 6 – – Yes Partial Overlap
2 15 Yes False Positive –
2 16 Yes False Positive –
2 22 – – Yes Partial Overlap
1 2 – – Yes Partial Overlap
1 5 Yes False Positive – –
1 10 – – Yes Partial Overlap
1 15 – – Yes Complete Overlap
1 17 – – Yes Partial Overlap
1 20 Yes False Negative – –
1 22 – – Yes Partial Overlap

observed from the tables that a very good sensitivity and accuracy
is possible from different channels of a recording at a time. r07
performance in Table 2 is listed for different time period than the
others.
6. Discussion
Table 4 shows the areas where there is a posibility of error
which has been overcome by this algorithm in channels 1, 2 and
3 of DBI .
Analysis of channel 3 is as follows: In Fig. 5, false positives are
possible and avoided at beats 1, 5, 12, 17 and 19. A false negative is
possible at beat 2 but the algorithm detects the fECG even in this
case. There is a partial overlap of mECG and fECG around beats
5, 10, 20 and 22, yet no detection error occurs. Even in the case of
complete overlap at beat 17, the fECG R-peak has been detected. All
cases which influence the detection, accuracy, efficiency and other
performance parameters for remaining channel’s output of DBI in
Fig. 5 are tabulated in Table 4.
The performances of different methods for different signals
have been analysed in Tables 5 and 6. The proposed method has
been efficient in certain aspects compared with the other. Best
performance mentioned in the table means only the good perfor-
mance (i.e., fHR > 110 bpm). Overall performances listed out in the
tables include the signal which does not even give minimum fHR
of 110 bpm. One such performance has been shown for a09 record-
ing of DBII in Table 3.
There are noises like baseline wander, respiration noise etc.,
which can affect the fECG extraction most. One such signal and
its identified fECG R-peaks with its annotated R-peaks have been
shown in Fig. 9.
It can be noted from Fig. 9 that there is 1 FP (around 1 s) and
1 FN (around 1.5 s) in the result which leads to an accuracy of
90.9% and sensitivity of 95.24%. Even with those noises, the ob-
tained accuracy and sensitivity is good. There are also possibilities
of FP and FN around time periods 1.7 s and 1.85 s in this signal
which are neighboured with noise and have been smoothened out
due to effective noise removal done by PATV. Else noise (peak af-
ter the fECG R-peak) occurs around time period 1.7 s and a noise
(peak around before the actual fECG R-peak) occurred around time
period 1.85 s may get detected as fECG R-peaks which is not actual
foetal R-peaks and can lead to false negatives also. But sometime,
202 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204

Table 5
Overall statistical analysis for DBI , DBII and DBIII .

MSE_HR (in MSE_RR (in

Methods F1 score (in%) Sensitivity (%) PPV (%) Accuracy (%) bpm2 ) ms)

ESN [15] 90.20 91.4 88.9 – – –

Least mean square (LMS) [15] 87.9 89.3 86.5 – – –
Recursive least square (RLS) [15] 88.2 89.7 86.8 – – –
TS [15] 89.3 89.9 88.8 – – –
FUSE [19] 95 95.6 94.3 – – –
FUSE – SMOOTH [19] 96 95.9 96 – – –
Optimal principle component selection [20] Only Best 89.8 89.3 90.5 – – –
SVD and SW combined [16] r01 signals 99.61 99.53 99.69 99.22 – –
(DBIII )
SVD and SW combined [16] r07 signals 99.28 99.2 99.36 98.57 – –
(DBIII )
Template adaption [23] TA 97.3 ± 10.8 97.4 ± 11.0 97.2 ± 10.7 96.0 ± 13.4 – –
Extended Kalman smoother [23] EKS 93.0 ± 20.3 93.1 ± 20.3 92.8 ± 20.3 91.2 ± 23.2 – –
FTM [17] without SQA 84.9 – – – 185.6 19.4
FTM [17] with SAQ 93.9 – – – 47.5 7.6
Principal component model [22] – 95 – – – 4.84
Proposed Method DBI (Average 99.25 100 98.55 98.53 0 0
of Channel 1,
2 & 3)
Proposed Method DBII (Best) 98.85 98.63 99.52 97.77 10.8 0.17
DBII (Overall) 96.04 94.86 97.41 92.8 156.85 2.22
Proposed Method DBIII (Best) 96.04 93.8 98.48 92.41 12 0.3
DBIII 94.25 92.55 96.36 88.48 108 2.14
(overall)

Table 6
Additional statistical analysis for DBII and DBIII .

Methods Correlation between FHR Median sensitivity Median positive predictivity

Synchronized averaging [18] – 0.982 0.976

Wavelet-based approach [21] 0.48 – –
Wavelet-based approach [21] fHR >110bpm 0.73 – –
Proposed Method DBII (Best) 0.97 0.99 0.989
DBII (Overall) 0.66 0.956 0.97
Proposed Method DBIII (Best) 0.94 0.954 0.99
DBIII (overall) 0.59 0.95 0.97

Fig. 9. AECG2 channel of a08 signal from DBII with identified foetal R-peaks and annotated foetal R-peaks marked.

it may require additional operations for removal of some noises
before extraction [40,41] i.e., in the pre-processing stage.
Most of the methods like efficient wavelet based ECG process-
ing for single lead fHR extraction [42] have acquired only the foetal
heart rate and other parameters but no fECG signal. Some methods
also require multiple channels to achieve good results. So the pro-
posed work proves to be a required development in this area of
research, since it is capable of extracting the fHR as well as fECG
with good accuracy and less computation time by using single lead
abdECG where the same can be observed from the results and dis-
cussions. Concentration in extraction of mECG, noise removal, ex-
traction of fECG and its related information has been carried out
with equality to ensure the performance of the proposed algo-
rithm. Quality of mECG signal extraction shown in Fig. 10 ensures
the availability of remaining information for good extraction of the
required. Fig. 10(a) shows the abdomen electrode signal of a06
recording from DBII which is used as the input signal and Fig. 10(b)
shows the constructed maternal ECG from the abdomen electrode
signal a06.
7. Conclusion
A simple approach has been used in this paper for extracting
a distinct foetal ECG from the mother. It is obtained by eliminat-
ing the mECG from the abdominal ECG. The algorithm performs
well even for extracting the fECG beat which was partially or com-
pletely overlapped. The foetus position and orientation with re-
spect to the abdomen’s surface of the mother is dynamic in time
which gets reflected in the acquired signal by change in shape,
amplitude and pulse generated. The same can be observed in the
 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
Fig. 10. (a) AECG1 of a06 (input abdomen signal after normalization stage) from DBII ; (b) Extracted mECG from the abdomen signal AECG1 of a06. Vertical axis gives the
amplitude (in NU).
signals discussed for different databases. This algorithm proves
good for various recordings of DBI , DBII and DBIII databases. In cer-
tain cases, good performance achieved from different channels of a
recording at a time which can be noted from Tables 1–3. The algo-
rithm has been tested with an i5 processor desktop which gener-
ates results in a second and this shows that the algorithm is capa-
ble of doing good in real time extraction.
This approach needs only single lead input and provides very
good accuracy in extracting the foetal information. Single lead in-
put requirement also reduces computational complexity and sys-
tem becomes more compact. The accuracy achieved is dynamic,
adaptive in nature and proved to be independent since the refer-
ence signal has been estimated from the abdECG itself to remove
the mECG. This shows that the extraction of fECG from abdECG by
the proposed method becomes very effective with its dynamic es-
timation.
Twin foetal ECG extraction is an area to which this work can be
extended. It can be clinically verified with very good SNR. These
advancements will definitely lead to a compact and real time sys-
tem realization.
Conflict of interest
The authors declare that they have no conflict of interest.
References
[1] S. Patel, R. Thaker, P. Shah, S. Majumder, Study of causes and complications of
intra uterine foetal death (IUFD), Int. J. Reprod. Contracept. Obstet. Gynaecol. 3
(4) (2014) 931–935, doi:10.5455/2320-1770.ijrcog20141211.
[2] R. Sameni, G.D. Clifford, A review of foetal ECG signal processing: issues and
promising directions, Open Pacing Electrophysiol. Ther. 3 (1) (2010) 4–20,
doi:10.2174/1876536X01003010 0 04.
[3] R. Mogra, Simplifying ultrasound assessment of the fetal heart: incorporating
the complete three vessel view into routine screening, Australas J. Ultrasound
Med. 16 (4) (2015) 168–175, doi:10.1002/j.2205-0140.2013.tb00243.x.
[4] R.O. Bonow, D. Mann, D. Zipes, P. Libby, E. Braunwald, in: Heart Disease: A
Textbook of Cardiovascular Medicine, fifth ed., Saunders, Philadelphia, 2012,
p. 108.
[5] Z. Luo, Fetal electrocardiogram extraction using blind source separation and
empirical mode decomposition, J. Comput. Inf. Syst. 8 (12) (2012) 4825–4833.
[6] G. Singh, G. Archana, Unraveling the mystery of vernix caseosa, Indian J. Der-
matol. 53 (2) (2008) 54–60, doi:10.4103/0019-5154.41645.
[7] J. Behar, J. Oster, G.D. Clifford, Non-invasive FECG extraction from a set of ab-
dominal sensors, Comput. Cardiol. 40 (2013) 297–300.
[8] G. Mihaela Ungureanu, J.W.M. Bergmans, The event synchronous canceller al-
gorithm removes maternal ECG from abdominal signal without affecting the
foetal ECG, Comput. Biol. Med. 39 (2009) 562–567, doi:10.1016/j.compbiomed.
2009.03.013.
203
[9] A. Cichocki, S.I. Amari, Adaptive Blind Signal and Image Processing: Learning
Algorithms and Applications, Wiley, New York, 2002.
[10] Z.-L. Zhang, Z. Yi, Extraction of temporally correlated sources with its appli-
cation to non-invasive fetal electrocardiogram extraction, Neurocomputing 69
(7–9) (2006) 894–899, doi:10.1016/j.neucom.2005.08.004.
[11] P.P. Kanjilal, S. Palit, G. Saha, Fetal ECG extraction from single channel mater-
nal ECG using singular value decomposition, IEEE Trans. Biomed. Eng. 44 (1)
(1997) 51–59, doi:10.1109/10.553712.
[12] W. Zheng, H. Liu, A. He, X. Ning, J. Cheng, Single-lead foetal electrocardiogram
estimation by means of combining R peak detection, resampling and comb fil-
ter, J. Med. Eng. Phys. 32 (2010) 708–719. https://doi.org/10.1016/j.medengphy.
2010.04.012.
[13] M. Niknazar, B. Rivet, C. Jutten, Fetal ECG extraction by extended state Kalman
filtering based on single - channel recordings, IEEE Trans. Biomed. Eng. 60 (5)
(2013) 1345–1352, doi:10.1109/TBME.2012.2234456.
[14] R. Martín-Clemente, J.L. Camargo-Olivares, S. Hornillo-Mellado, M. Elena,
I. Román, Fast technique for noninvasive fetal ECG extraction, IEEE Trans.
Biomed. Eng. 58 (2) (2011) 227–230, doi:10.1109/TBME.2010.2059703.
[15] J. Behar, A. Johnson, G.D. Clifford, J. Oster, A comparison of single channel fetal
ECG extraction methods, Ann. Biomed. Eng. 42 (6) (2014) 1340–1353, doi:10.
1007/s10439- 014- 0993- 9.
[16] N. Zhang, J. Zhang, H. Li, O.O. Mumini, O.W. Samuel, K. Ivanov, L. Wang, A novel
technique for fetal ECG extraction using single - channel abdominal recording,
Sensors 17 (3) (2017) 1–6 (457), doi:10.3390/s17030457.
[17] C. Liu, P. Li, C.D. Maria, L. Zhao, H.Z.Z. Chen, A multi-step method with signal
quality assessment and fine-tuning procedure to locate maternal and fetal QRS
complexes from abdominal ECG recordings, Physiol. Meas. 35 (8) (2014) 1665–
1683, doi:10.1088/0967-3334/35/8/1665.
[18] A. Dessì, D. Pani, L. Raffo, An advanced algorithm for fetal heart rate estima-
tion from non-invasive low electrode density recordings, Physiol. Meas. 35 (8)
(2014) 1621–1636, doi:10.1088/0967-3334/35/8/1621.
[19] J. Behar, J. Oster, G.D. Clifford, Combining and benchmarking methods of foetal
ECG extraction without maternal or scalp electrode data, Physiol. Meas. 35 (8)
(2014) 1569–1589, doi:10.1088/0967-3334/35/8/1569.
[20] C.D. Maria1, C. Liu, D. Zheng, A. Murray, P. Langley, Extracting fetal heart beats
from maternal abdominal recordings: selection of the optimal principal com-
ponents, Physiol. Meas. 35 (8) (2014) 1649–1664, doi:10.1088/0967-3334/35/8/
1649.
[21] R. Almeida, H. Gonçalves, J. Bernardes, A.P. Rocha, Fetal QRS detection and
heart rate estimation: a wavelet-based approach, Physiol. Meas. 35 (8) (2014)
1723–1735, doi:10.1088/0967-3334/35/8/1723.
[22] J.A. Lipponen, M.P. Tarvainen, Principal component model for maternal ECG
extraction in fetal QRS detection, Physiol. Meas. 35 (8) (2014) 1637–1648,
doi:10.1088/0967-3334/35/8/1637.
[23] F. Andreotti, et al., Robust fetal ECG extraction and detection from abdomi-
nal leads, Physiol. Meas. 35 (8) (2014) 1551–1567, doi:10.1088/0967-3334/35/
8/1551.
[24] H. Zhang, Semi - blind source extraction algorithm for foetal electrocardiogram
based on generalized autocorrelations and reference signals, J. Comput. Appl.
Math. 223 (2009) 409–420, doi:10.1016/j.cam.2008.01.03.
[25] K. Assaleh, H. Al-Nashash, A novel technique for the extraction of fetal ECG
using polynomial networks, IEEE Trans. Biomed. Eng. 52 (6) (2005) 1149–1152,
doi:10.1109/TBME.2005.844046.
[26] R.H. Nair, J. Rolant Gini, K.I. Ramachandran, A Simplified approach to identify
the fetal ECG from abdECG and to measure the fHR, IFMBE Proc. 52 (2015)
23–26, doi:10.1007/978- 3- 319- 19452- 3_7.
204 R.G. John and K.I. Ramachandran / Computer Methods and Programs in Biomedicine 175 (2019) 193–204
[27] K.G. GreeneM. Whittle (Ed.), The ECG waveform, Baillière’s Clin. Obstet. Gy-
naecol. (1987) 131–155.
[28] T. Stampalija, M. Signaroldi, C. Mastroianni, Fetal and maternal heart rate con-
fusion during intrapartum monitoring: comparison of trans-abdominal fetal
electrocardiogram and Doppler telemetry, J. Matern. Fetal Neonatal 25 (2012)
1517–1520, doi:10.3109/14767058.2011.636090.
[29] J. Reinhard, B.R. Hayes-Gill, S. Schiermeier, Intrapartum signal quality with ex-
ternal fetal heart rate monitoring: a two way trial of external Doppler CTG
ultrasound and the abdominal fetal electrocardiogram, Arch. Gynecol. Obstet.
286 (2012) 1103–1107, doi:10.1007/s00404- 012- 2413- 4.
[30] A.J. Wolfberg, The future of foetal monitoring, Rev. Obstet. Gynecol. 5 (3/4)
(2012) e132–e136, doi:10.3909/riog0197.
[31] Cutaneous potential recordings of a pregnant woman, daISy database (2010).
http://homes.esat.kuleuven.be/∼smc/daisy/daisydata.html.
[32] A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, P.Ch. Ivanov, R.G. Mark,
J.E. Mietus, G.B. Moody, C.-K. Peng, H.E. Stanley, PhysioBank, PhysioToolkit, and
PhysioNet: components of a new research resource for complex physiologic
Signals, Circulation 101 (23) (20 0 0) e215–e220. Circulation Electronic Pages
http://circ.ahajournals.org/content/101/23/e215.full . June 13.
[33] M. Kotas, J. Jezewski, L. Horoba, A. Matonia, Application of spatio-temporal
filtering to fetal electrocardiogram enhancement, Comput. Methods Programs
Biomed. 104 (1) (2011) 1–9, doi:10.1016/j.cmpb.2010.07.004.
[34] P. Kathirvel, M.S. Manikandan, S.R.M. Prasanna, K.P. Soman, An efficient R
peak detection based on new nonlinear transformation and first-order Gaus-
sian differentiator, Cardiovasc. Eng. Technol. 2 (4) (2011) 408–425, doi:10.1007/
s13239- 011- 0065- 3.
[35] Stack overflow, QRS detection of a raw ECG signal in matlab stackover-
flow.com/questions/qrs detection of a raw ECG signal in matlab, 2013 (Ac-
cessed 24 December 2018)
[36] Dr. H. Knipe, Dr. Y. Weerakkody, et al., http://radiopaedia.org/articles/
fetal- heart- rate, 2015 (Accessed 12 January 2019).
[37] I.J. Myung, Tutorial on maximum likelihood estimation, J. Math. Psych. 47
(20 03) 90–10 0, doi:10.1016/S0 022-2496(02)0 0 028-7.
[38] I.W. Selesnick, Simultaneous polynomial approximation and total variation de-
noising, in: IEEE International Conference on Acoustics, Speech and Signal Pro-
cessing, 2013, pp. 5944–5948, doi:10.1109/ICASSP.2013.6638805.
[39] I.W. Selesnick, S. Arnold, V.R. Dantham, Polynomial smoothing of time series
with additive step discontinuities, (2012), http://eeweb.poly.edu/iselesni/patv/
PSTV2012.pdf.
[40] J. Joseph, J. Rolant Gini, K.I. Ramachandran, Removal of BW and respiration
noise in abdECG for fECG extraction, Adv. Signal Process. Intell. Recogn. Syst.
678 (2018) 3–14, doi:10.1007/978- 3- 319- 67934- 1_1.
[41] N. Mohan, S. Kumar S, P. Poornachandran, K.P. Soman, Modified variational
mode decomposition for power line interference removal in ECG signals, Int. J.
Electr. Comput. Eng. (IJECE) 6 (1) (2016) 151–159, doi:10.11591/ijece.v6i1.8592.
[42] Castilloa, et al., Efficient wavelet-based ECG processing for single-lead FHR ex-
traction, Digit. Signal Process. 23 (6) (2013) 1897–1909, doi:10.1016/j.dsp.2013.
07.010.