INVESTIGATION OF BLIND SOURCE SEPARATION METHODS

FOR EXTRACTION OF FETAL ECG

K. V. K. Ananthanag J. S. Sahambi
Dept. of Electronics & Comm. Engg. Dept. of Electronics & Comm. Engg.
1.I.T. Guwahati, Assam, India I.I.T. Guwahati,Assam, India
ananthanag78@yahoo.com jsahambi@ieee.org

Abstract acquired which contains a strong Maternal Electrocardio-

Extraction offetal electrocardiogram (FECG)from ma- gram (MECG). An early diagnosis before delivery using
ternal skin electmde measurements needs a robust technique. non-invasive techniques increases the effectiveness of the
This problem can be modeledfiom the perspective of Blind appropriate treatmcnt.
Source Separation (BSS) and so the problem reduces to the The detection of the antepartum E C G from electrodes
estimation of the independent sources offetal and maternal on the mother’s skin, is fundamentally equivalent to the
cardiac activiiy. In this paper; analysis and study of some adult problem, hut some specific difficulties arise in this
major algorithms like Bell and Sejnowski’s infomar algo- case. The omnipresent MECG, which can he 5-1000 times
rithm, Cardoso’sjoint appmximate diagonalization of eigen larger in its intensity, forms the largest interference. Elec-
matrices (JADE) algorithm, Hyvarinen ’sfued-point algo- tromyographic (EMG) activity, potential distributions gen-
rithm and Comon’s algorithm, was madefor this important erated by respiration and stomach activity can become very
biomedical application. For mbustness, two scenarios, i.e, important and the positioning of thc electrodes is crucial,
( a )different amplitude ratios of simulated maternal and fe- since fetal movements can completely change the physi-
tal ECG and (b)different values of additive white Gaussian cal configuration. Non-biological interference sources, such
noise, were investigated. It was observed that ifthe ratio of as power-line coupling and thermal noise due to the elec-
the amplitude of maternal to fetal ECG is IO:] with an in- tronic equipment, conupt the recordings as well. There-
put SNR of 2 dB, allfour algorithms were able to extract the fore an FECG-extraction method requires the elimination of
fetal ECG. The signal-to-error (SER) ratios of the extracted the MECG as well as an “optimal” detection of the E C G .
maternal and fetal ECG were around 3 dB and I dB, re- E C G can be used to classify arrhythmias, to study con-
spectively. genital heart disease and to ohserve fetal well being during
growth retardation or (abnormal) twin pregnancy.
Keywords: Fetal Electrocardiogram, Blind Source Separa- Fetal ECG extraction problem is not easily solved by
tion, Independent Component Analysis, JADE. conventional filtering techniques. Linear filtering in the Fou-
rier domain essentially fails since the differences among the
three components: matemal, fetal, and noise can’t be de-
1. INTRODUCTION
fined in the spectral domain. Also, the spectral content of
the three components is rather similar and contains strong
The extraction of FECG is very important from clinical point
broad-hand contributions. Several different approaches like
of view because sometimes it is the only information source
coherent averaging, matched filtering, auto- and cross - cor-
in early stage diagnostics of the fetal health. Most methods
relation based methods, adaptive filtering, sequenced adap-
for acquiring the FECG are invasive and are done by plac-
tive filtering, singular value decomposition (SVD), multi-
ing an electrode on the fetal scalp. These methods are accu-
reference adaptive noise cancellation (MRANC) etc., have
rate but they are feasible only during delivery. Non-invasive
been proposed to address this problem [ I ] [2]. All these
methods must be employed if the antepartum FECG is to be
methods have one of the following drawbacks: 1) A single
obtained. The E C G may hc acquired by placing a number
channel of E C G is extracted whereas better characteriza-
of electrodes on the maternal abdomen, as close as possible
tion can be accomplished by two or more channels. 2) The
to the fetal heart, hoping that at least one of the electrodes
signal acquisition is highly sensitive to electrode placement,
will have the E C G with high enough Signal-to-Noise Ra-
human interaction, stage of pregnancy, position of fetus, etc.
tio (SNR). In addition to this, thoratic signals can also be
3) Extraction of P and T waves is not satisfactory [3] [41 [ 5 ] .
CCECE 2003 - CCGEI 2003, Montr&l, May/& 2003 ?he different approaches which have been developed for
0-7803-778 I -8/03/I 17.00 @ZOO3 IEEE extracting the fetal ECG from a single lead are all nonlinear,

- 2021 -
reflecting the difficulty of distinguishing the MECG, FECC the computational load, only the n most significant eigen
and noise in the frequency domain. In the Multi-Reference pairs of fourth order cumulants obtained from the whiterled
Adaptive Noise Cancellation (MRANC) [2] only second or- process arc joint diagonalized by unitary matrix U. The
der statistics (SOS) are used. A very strong point of BSS separated matrix can he estimated as Ut 2,where t repre-
is its robustness with respect to the number of electrode sents pseudo inverse. The JADE contrast function [8]is the
and their placement. This level of robustness cannot be ob- sum of squared fourth order cross cumulants
tained easily with SOS. Further, FECG sources were clearly
revealed by exploiting a few abdominal leads exclusively, q J A D E=
(y) (Qij~)' (2)
with favorable results from up to only 3 electrodes [ 11. This ijkljiikl
robustness arises as a major advantage of BSS techniques. As this algorithm uses cross-cumulants, there is no need to
go for gradient descent and hence there is no chance of di-
2. METHODOLOGY vergence. Also there is no problem of updating the weights
and tuning the parameters as in Bell and Sejnowski's info-
Independent component analysis (ICA) is the most widely max algorithm.
used technique in BSS and many authors often use the terms
interchangeably. The basic assumption of ICA is "statistical 3.2. Fixed-point Algorithms
independence" of the sources. Different bioelectric current
sources correspond to different bioelectric mechanisms, and The original fixed-point algorithm [9]uses kurtosis and com-
so without loss of generality we can assume them to be sta- putations can he performed either in hatch mode or in a
tistically independent. semi-adaptive manner. It uses deflation approach to update
Letx(k) = ( ~ ( k ) , . .,zq(k))*
. t Rq (wherekisthe the columns of separating matrix W and to find the inde-
discrete time index) be the unknown independent bioelectric pendent components one at a time. More recent versions are
currentsourcesandy(k) = (yl(k),..., ~ ~ ( kE ) Rp, ) ~ using hyperbolic tangent, exponential or cubic functions as
be the signals recorded by the electrodes, then the FECG contrast function.
extraction problem can be formulated as The update rule for the deflation method is given by I 101

y ( k )= Mx(k) +n(k) (1) w*(k) = C-'E{x g(w(k - l ) T X ) }

-E{g'(w(k - 1))Tx)}w(k - 1) (3)
where matrix M E RPXq is the mixing matrix whose svuc-
ture is determined by the body geometry, the electrodeplace-
ment and source positions and the conductivity of the body w(k) = w * ( k , / J ~ (4)
tissues, and n(k) = (nl(k), . . . ,np(k))T E E P is the
noise. Fortunately, (1) is the familiar instantaneous linear where g can be any suitable non-quadratic contrast function,
mixtures and thus the problem reduces to the estimation of with derivative g', and C is the covariance matrix of the
the independent sources of fetal and maternal cardiac activ- mixtures, x. ~ ( k ) ~ x ( t t) ,= 1 , 2 , ... equals one of the
ity. sources.

3.3. Bell and Sejnowski's algorithm

3. ICA ALGORITHMS
Bell and Sejnowski have developed an unsupervised leam-
Jutten and Hdrault provided one of the first significant ap- ing algorithm (Infomax algorithm) [ l l ] based on entropy
proaches to the problem of blind separation of instantaneous maximization in a single layer feed-fonvard neural network.
linear mixtures [6]. Since then, many different approaches The main idea is that maximizing the joint entropy H(g) of
have been attempted by numerous researches using neural the outputs of a neural processor can approximately mini-
networks, artificial learning, higher order statistics, mini- mize the mutual information among the output components.
mum mutual information, beam-forming and adaptive noise It is proved that infomax is equivalent to maximum likeli-
cancellation, each claiming various degrees of success. This hood [ 121.
section presents a brief description of the respective appro- The joint entropy of n variables, y1, yz, ..., ynr which
aches of the compared algorithms. are the outputs of the neural network, may be written as:

3.1. JADE algorithm H ( Y I , . . ,yn)

. = H ( Y I ) +... +H(yn)
--I(Yl,"' ,Yn) (5)
The JADE algorithm [71 relies on second and fourth-order
cumulants to separate the sources. SOS are used to obtain a If the nonlinear transfer function of a neural network matches
whitening matrix Z from the sample covariances. To reduce the probability density function of the inputs, and the joint

- 2022 -
entropy H(y1,. . ,yn) or the outputs is maximized, the
mutual information I(yl, ' . ,yn) among the outputs is then
minimized. Thc output signals are assumed to be indepen-
dent. The learning rule for a single laycr feed-forward neu- SNR JADE Fastica Bcll Comon
Mat Fet Mat Fet Mat Fe1 Mat Fet
ral network to implemcnt the separation is
2 2.96 2.63 2.94 2.64 2.96 2.60 2.71 2.23
AW a [WT]-i + (I - 2y)xT (6) 4 4.67 4.38 4.67 4.37 4.67 4.35 4.35 3.90
6 6.44 6.17 6.43 6.20 6.44 6.1 1 6.04 5.56
Awo c( 1-2y (7)
8 8.36 7.97 8.34 8.04 8.35 7.88 7.83 7.17
where y = f(u), U = Wx +
wo and f(u)is a sigmoid IO 10.29 9.78 10.27 9.67 10.27 9.63 9.57 8.70
+
contrast function. usually f(u)= 1 &U-') or f(u)=
tanh(u).Here W is the weight matrix and WO is the bias
vector.

3.4. Comon's algorithm

A specific contrast function is proposed in [131, based on
minimization of mutual information between the compo-
nents at the output of separator (which is directly related
to Kullback-Leibler divergcnce between the output vector
probability density function (pdf) and it's pdf if it was made
of independent components). After some manipulations on
the Edgeworth expansion [14j of the source joint pdf, the
contrast function simplifies into sum of the output squared SNR JADE Fastica Bell Comon
rtt-order marginal cumulants and for r = 4 it becomes out- Mat Fet Mat Fet Mat Fet Mat Fet
put squared kurtosis: 2 3.35 1.10 3.31 1.11 3.35 1.11 2.91 0.92
4 5.04 2.77 5.00 2.77 5.03 2.77 4.60 2.58
2, 6 6.76 4.49 6.76 4.49 6.76 4.47 6.33 4.26
WQ)= C(k:a,J2 (8) 8 8.63 6.24 8.63 6.24 8.62 6.19 8.22 5.98
i=l 10 10.52 8.02 10.51 8.00 10.50 7.93 10.12 7.70

4. RESULTS & DISCUSSIONS

Thc observed signals at the electrodes were simulated by

taking two different ECG signals. To simulate real condi-
tions, the first signal (assumed as E C G ) was madc 5 or IO
times less in strength and with double the number of cycles
(or peaks) compared to the second (assumed as MECG). Af-
ter mixing the two signals by a random matrix, the strength
of the signals was calculated and random white Gaussian
noise was added accordingly so as to make the desired SNR
as shown in the Tables (1) and (2). In order to compare the
four algorithms each was given the same set of input sig-
nals, random mixing matrix and random noise. Due to lack -- 0

of space, only the SER of the extracted signals and number

of floatingpoint operations (FLOPS) required are presented.
The SER was obtained by using the following relation [I31

Fig. 1. Input signals(z1, $2) and mixed signals ( y ~y, ~ )

where s(tj is the desired signal and e(t) = i(t) - s ( t ) is
the error (or noise to be more accurate). Here i ( t ) is the The JADE, Infomax and fixed point algorithms perform-
estimated source signal and i(t) and s ( t ) should be at the ed slightly better than the Comon's algorithm. The accuracy
same energy level and phase while calculating e ( t ) . of Infomax algorithm is highly dependent upon the sweeps,

- 2023 -
 algorithms were able to detect the R wave.

6. REFERENCES
[ I ] V. Zarzoso and A. K. Nandi, “Noninvasivc fetal elec-
trocardiogram extraction:blind source separation ver-
sus adaptive noise cancellation,” IEEE Trans. Biomed.
Eng., vol. 48, pp. 12-18, Jan. 2001.
121 B. Widrow er al., “Adaptive noise cancelling: Prin-
ciples and applications,” in Proc. IEEE, vol. 63,
L, 4 l.J&&-Jl 1 pp. 1692-1716, Dec. 1975.
5m m
Samples
moa- 2000 2500 J. Vanderschoot et al., “Two methods for opli-
mal MECG elimination and FECG detection from
skin electrode signals,” IEEE Trans. Bio-Med. Eng.,
Fig. 2. Extracted maternal signal by a) JADE b) Fastica c) vol. 34, pp. 233-243, March 1987.
Infomax d) Comon’s algorithm
D. Callaerts et al., “Comparison of SVD methodj to
extract the foetal electrocardiogram from cutaneous
electrode signals,” Med. Biol. Eng. Comput., vol. 28,
pp. 211-224, May 1990.
A. Kam and A. Cohen, “Matemal ECG Elimination
and Foetal ECG Detection - Comparison of Several
Algorithms,” in Proc. of 20th Ann. Int. Con& IEEE
EMBS, vol. 20, pp. 174-177, 1998.
C. Jutten and J. Hkrault, “Blind separation of sources,
part I:An adaptive algorithm based on neuromimatic
architecture,” Signal Processing, vol. 24, pp. 1--10,
July 1991.

samp,ss
J.-F. Cardoso and A. Souloumiac, ‘‘Blind heamforn-
ing for non-Gaussian signals,” in Pmc. IEE -F,
vol. 140, pp. 362-370, Dec. 1993.
Fig. 3. Extracted fetal signal by a) JADE b)Fastica c) Info-
max d) Comon’s algorithm J.-F. Cardoso, ‘‘Higher order contrasts for independent
component analysis,” Neural computation, vol. 11,
or the iterations, to update the weights. This can give better pp, 157-192, Jan. 1997.
results hut with a higher computational cos!. Fixed-point d- [9] A. Hyvirinen and E. Oja, “A rast fixed-point algorithm
gorithm is fast but the ordering of the extracted sources vary for independent component analysis,” Neural compu-
for each run even with the same input data and same accu- tation, vol. 9, no. 7, pp. 1483-1492, 1997.
racy in results. JADE algorithm has the least computational
[IO] A. Hyvhnen, “A family of fixed-point algorithms
cost and also no parameters to tune. for indeDendent comvonent analvsis.” IEEE Confer-
~~ ~

ence on‘ Acoustics, kpeech and- Signal Process&,

pp. 3917-3920, 1997.
5. CONCLUSIONS
[ I l l A. J. Bell and T. J. Sejnowski, “An information-
All the BSS algorithms can extract FECG even in the pres- maximization amroach to blind seuaration and blind
ence of a third or a fourth source. To he more clear, if noise deconvolution,”’ Neural computation, vol. 7, no. 6,
is taken as a third source and mixed with ECG signals, all pp. 1129-1159, 1995.
the BSS algorithms can extract FECG with very good accu- [12] J.-E Cardoso, “Infomax and Maximum Likelihood
racy. This is very useful especially in multi fetal cases. The in Blind Source Separation:’ IEEE Signal Processing
noise that is added after mixing can’t be considered as a Lett., vol. 4, pp. 112-1 14, April 1997.
source and hence can’t he separated completely. The more [13] P. Comon, “Independent component analysis, a new
the number of electrodes or observed signals, the more is concept?,” Signal Processing (Special Issue Hi,gher
the accuracy. BSS methods based on higher order statis- Order Statistics), vol. 36, pp. 287-314, April 1994.
tics are not influenced by the electrode placement. All the
1141 P. Comon, “Contrasts for Multichannel Blind De-
algorithms were able to extract E C G considerably if the convolution,” IEEE Signal Processing Lett., vol. 3,
amount of the input SNR was high. As the SNR was de- pp. 209-21 1, July 1996.
creased, the P and T waves were lost in noise, but still the

- 2024 -