Professional Documents
Culture Documents
net/publication/319175920
Non-invasive Fetal ECG Extraction from Maternal Abdominal ECG Using LMS
and RLS Adaptive Algorithms
CITATIONS READS
16 351
3 authors:
Petr Bilik
VŠB-Technical University of Ostrava
107 PUBLICATIONS 860 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Device for measurement and processing of biosignals with the use of fiber optic sensors View project
TA ČR GAMA: PRE SEED VŠB-Technical University of Ostrava fund - BroadbandLIGHT View project
All content following this page was uploaded by Radek Martinek on 03 May 2019.
kahankova.radana@vsb.cz, radek.martinek@vsb.cz,
petr.bilik@vsb.cz
1 Introduction
adfa, p. 1, 2011.
© Springer-Verlag Berlin Heidelberg 2011
more suitable, convenient, and can be used not only during the labor but before it, too.
This method is based on recording the electrical signals by multichannel skin electrodes
placed on mother’s abdomen. Its drawback is that the measured signal is characterized
by a significant amount of noise and artifacts such as fetal and maternal movement
activity, respiration, and mainly by the maternal heart activity [3], [4].
Maternal Electrocardiogram (mECG) overlaps with fECG in time and in frequency do-
main as well. Due to that, reducing the strong maternal component from a composite
abdominal signal is a challenging task, where conventional filtration techniques fail.
There are some scientific teams that have been working on solving this problem so far.
However, compared to the research associated with the standard ECG, understanding
and solving fECG problematics is yet in the beginning. Comprehensive overviews of
the current state of the problem can be found in [1], [2]. Some methods available for
fECG elicitation are described in the following section.
Several methods have been used the extraction of the fECG with various degrees of
success. These methods can be Non-adaptive, for example Single Value Decomposition
(SVD), Blind Subspace Separation (BSS), Averaging techniques, Finite Impulse Re-
sponse (FIR) and Infinite Impulse Response (IIR), Principal Component Analysis
(PCA), Periodic Component analysis (πCA) and so on. Fig. 1, shows an overview of
methods that are associated with fECG [3], [5]. This paper focuses on Adaptive meth-
ods, which are going to be explained in more detail in the following section.
There are many different adaptive algorithms that can be used to obtain fECG (as shown
in Fig. 1). These methodologies include linear techniques, such as Least Mean Square
(LMS) algorithms, Recursive Least Square (RLS) algorithms. Adaptive methods for
fECG elicitation include also so-called Non-linear methods, such as Particle Swarm
Optimization (PSO) and Quantum behaved Particle Swarm Optimization (QPSO), ar-
tificial intelligence techniques, Adaptive Neuro-fuzzy Inference system (ANFIS, more
in [1], [6], [7], and [8]), genetic algorithms (GA), Hybrid neural networks (HNN), Ar-
tificial Neural Networks (ANN) and Bayesian adaptive filtering frame works which
comprise Kalman filters.
Authors have been testing several linear adaptive algorithms in the previous work (see
[9]). According to these results, the authors have decided to design and implement own
adaptive system based on the two algorithms that were the most effective in fECG ex-
traction. Additionally, the experiments in the recent work, which is presented in this
paper, were performed on different signals than in [9]. The main reason was that the
signals used before were very simple and did not adjust the influence of the human
body. Thus, the signals were only linearly disturbed (by 50 th order FUR filter), which
would affect the results. This paper used the signals generated by a Novel generator
(see [13]), which is able to create signals with the characteristics of real signals (in time
and frequency domain, too). This ensures objectivity of the experiments.
Based on previous research, this paper focuses primarily on the Least Mean Squares
(LMS) and Recursive Least Squares (RLS) based FIR Adaptive Filters methods. In the
chapter below there are mathematical descriptions of the chosen representatives of
these two groups – standard LMS and RLS algorithm.
The choice of the Standard LMS and RLS algorithms is obvious from the figure below
(see Fig. 3), which summarizes part of the results from [9]. The quality is evaluated by
the parameter Percentage Root Difference (PRD), which should approach the zero
value for reaching ideal extraction (more details in section 4.1). The Standard LMS
shows the best results in the case of LMS based methods, while the Standard RLS al-
gorithm outperforms the RLS based methods. In addition, both of them are very stable
and have relatively low computational time. For more information, please, see [9].
3 Theoretical background
This section focuses on the description of the algorithms used for the experimental part.
The authors choice of LMS and RLS algorithms was based on their previous research
(see [9]).
3.1 LMS based algorithms
Least mean squares (LMS) algorithms are adaptive filters with the ability to change the
filter coefficients to produce the least mean squares of the error signal (the difference
between the desired and the actual signal). It is a stochastic gradient descent method in
that the filter is only adapted based on the error at the current time [7].
The standard LMS algorithm performs the following operations to update the coeffi-
cients of an adaptive filter:
1. Calculates the output signal y(n) from the adaptive filter.
2. Calculates the error signal e(n) by using the following equation:
where w(n) is the filter coefficients vector, e(n) is the error signal and x(n) is the filter
input vector, and μ is the convergence constant or also the step size of the adaptive
filter. Step size significantly influences the convergence speed and performance of the
adaptive filter [6.].
The standard RLS algorithm performs the following operations to update the coeffi-
cients of an adaptive filter:
1. Calculates the output signal y(n) from the adaptive filter by using the following equa-
tion:
where w(k) is the filter coefficients vector, 𝐊(𝑘) is a gain vector and is defined as:
𝑃(𝑘)𝐮(𝑘)
𝐊(𝑘) = . (6)
𝜆+𝐮T (𝑘)𝑃(𝑘)𝐮(𝑘)
𝑃(𝑘) is a correlation matrix of the input signal and its ignition value 𝑃(0) is defined
as:
𝛿 −1 0 … 0
𝑃(0) = [ 0 𝛿 −1 … 0 ], (7)
… … … …
0 0 … 𝛿 −1
where 𝛿 is a regulation factor. Standard RLS algorithm uses following equation for the
correlation matrix actualization:
The filtration quality evaluation is based on absence of the noise and the estimation of
the similarity of the recovered and the ideal (reference) fECG signals. In this paper, we
used two parameters to control the effectiveness of the fECG extraction, namely Per-
cent Root-Mean-Square Difference (PRD) and Signal to Noise Ratio (SNR). In addi-
tion, Root mean square error (RMSE) is introduced as accompanying parameter. Note
this is possible only with synthetic data due to the existence of reference signal. In case
of using real data, the objective quality assessment is much more challenging task.
The PRD is a common parameter used in ECG data compression (see [10]) and it is
determined by following equation:
2
∑𝑁
𝑖=1[𝑥𝑜𝑟𝑔 (𝑖)−𝑥𝑟𝑒𝑐 (𝑖)]
𝑃𝑅𝐷 = √ ∑𝑁 2 ∙ 100, (9)
𝑖=1 𝑥𝑜𝑟𝑔 (𝑖)
where xorg denotes the original signal (ideal fECG) and xrec the signal recovered by the
algorithm. The closer this value is to zero the similar are both signals [10], [11].
4.2 Signal to noise ratio (SNR)
SNR is a parameter for describing the relation between the signal and the noise. The
input and output values of the parameter are defined by Eq. 10 and 11, respectively.
2
∑𝑁−1
𝑖=1 [𝑥𝑜𝑟𝑔 (𝑖)]
SNRIN = 10∙ log10 ( 𝑁−1 2 ), (10)
∑𝑖=1 [𝑥𝑛𝑜𝑖𝑠𝑒 (𝑖)−𝑥𝑜𝑟𝑔 (𝑖)]
2
∑𝑁−1
𝑖=1 [𝑥𝑜𝑟𝑔 (𝑖)]
SNROUT = 10∙ log10 ( 𝑁−1 2 ), (11)
∑𝑖=1 [𝑥𝑟𝑒𝑐 (𝑖)−𝑥𝑜𝑟𝑔 (𝑖)]
where xorg is the reference signal (ideal fECG), xnoise is the undesired signal produced
by maternal heart (mECG), and xrec the recovered signal.
The SNR quantifies the relation between the fetal signal and the rest of the unwanted
components (mECG). In case of real measurements of fECG, this approach cannot be
used, since for calculating the values of both PRD and SNR the contribution of the fetal
signal and the noise must me known. In this initial research, we use synthetic signals
and therefore this information is available [11].
RMSE is the last parameter used in this paper and it is defined as follows:
1 2
𝑅𝑀𝑆𝐸 = √ ∑𝑁
𝑖=1(𝑠𝑖𝑔𝑜𝑟𝑔 − 𝑠𝑖𝑔𝑟𝑒𝑐 ) , (12)
𝑛
Sigorg is again the reference signal (ideal fECG) and sigrec is the recovered signal.
RMSE measures the differences between values predicted by an estimator and the
values observed. The closer this value is to zero the more accurate is the system [11].
5 Experiments
In this section, we present the experimental results. Firstly, it is vital to find optimal
settings for the designed algorithms. Optimization of the LMS parameters, namely step
size μ and filter length M, was performed using a stage-wise grid search as follows:
1. Filter length value was set on the value of 40 (considered [9]) and stayed fixed, while
the value of the step size μ varied in order to find the optimal value.
2. With the fixed optimal value μ from the previous step, the optimal value M was
searched.
3. Table 1 summarizes the above mentioned step 1. The best results, which are charac-
terized by lowest value of the parameter PRD and the highest SNR, are marked out
and the corresponding value of μ is used in the step 2.
Table 1. The results for the LMS algorithm with filter length M = 40.
Table 2 summarizes the above mentioned step 2. The optimal value of the step size (μ
= 0,017) was set, while the filter length varied.
Table 2. The results for the LMS algorithm with optimal value of μ.
Similar process was repeated in case of the implemented RLS algorithm. Optimization
of the RLS parameter (filter length M), was performed using a stage-wise grid search
as follows:
In the previous research, the authors tested the filters that were already included in the
MATLAB® library (function ‘adaptfilt’). On the following Figures, there is a compar-
ison of these filters and the algorithms implemented by the authors.
Figure 4 shows the influence of the filter length value on the filtration efficiency. Con-
sidering the ideal filtration with the value of PRD = 0%, the implemented RLS algo-
rithm shows the same results in the stable area (M ∈ 〈20,100〉), while the implemented
LMS algorithm slightly outperforms the MATLAB® LMS algorithm.
Fig. 4. : Performance comparison of the filters included in the MATLAB® library and imple-
mented by the authors.
Figure 5 shows the dependence of the computational time on the filter length M. Results
of the implemented LMS algorithm considerably outperformed the algorithm in
MATLAB®. The reason was that the algorithm was optimized for the purpose of the
fECG extraction, whereas the second algorithm was designed for the general use.
However, in case of the RLS algorithm it was quite opposite. The algorithm is much
more complicated in nature, therefore even its design and implementation is not a sim-
ple task.
Figures 6 and 7 show the abdominal compound signal used for the experiments before
after using the adaptive system. The waveform of the ideal (reference) output signal,
which was used for the performance evaluation, is also included.
The result show that using linear adaptive filters autonomously is not efficient enough
for the elimination of the maternal component from the abdominal signal. Linear meth-
ods are able to reduce amplitude of the mECG, especially P and T waves, but never
accomplish its complete elimination from real recordings. The shape of maternal QRS
complex was only distorted due to the linear filters’ ability of reducing high frequen-
cies. However, most of the QRS complex energy in spectral domain belongs to the
frequencies below 50 Hz.
Fig. 6. : Waveforms of the LMS adaptive system input and output and the ideal fECG
Fig. 7. Waveforms of the RLS adaptive system input and output and the ideal fECG
Non-linear methods are able to work with the nonlinearity of the human body. Regard-
ing this, nonlinear methods can show better results in fetal ECG extraction in case of
the real recordings, which is confirmed by a lot of research focused on this topic (see
for example [14], [15], [16], and [17]).
However, using nonlinear methods does not ensure successful fECG reduction. In case
of linearly distorted signals it can even show worse results, which was proved experi-
mentally in simultaneous work (more in [16.]). The future of the fECG extraction be-
longs to the combination of the linear and nonlinear methods.
Note that the experimental results were influenced by the type of data used. These re-
cordings did not take into account the position of the fetus with regard to the electrodes,
which is a crucial condition that affects the signal properties. These data were also used
in case of a research taken place in Oxford and only on a small amount of the data was
fHR detection successfully accomplished. This problem should therefore should be
considered in the future research. Authors are currently working on improving the sig-
nal generator that would create synthetic data considering above mentioned conditions
[18], [19], and [20].
7 Conclusion
This paper introduced an extension of the previous research based on simulations with
LMS based FIR Adaptive Filters contained in MATLAB® function ‘adaptfilt’.
Authors designed and implemented RLS and LMS algorithms, used different type of
data and again suggested the ideal combination of the chosen settings for these func-
tions based on the results of Percentage Root-Mean-Square Difference (PRD), input
and output Signal to Noise Ratios (SNRs), and Root Mean Square Error (RMSE).
In comparison with the results in previous research, where the elimination of the ma-
ternal components was successful, the findings show that the linear adaptive methods
are not as efficient with the fECG extraction when used autonomously. The reason is
the choice of data used for the research. Data in the previous research are publicly avail-
able and used for the experiments. However, they are very simple and do not take into
account the nonlinear influence of the human body. The distortion is modelled using
50th order FIR filter, which causes some delay, but do not affect the signal in any dif-
ferent way, unlike the human body. Therefore, the linear methods worked very well.
In this extended research, authors used signals generated by a Novel signal generator,
which involve the nonlinearities, thus have the characteristics more similar with the real
signals and therefore the validity of the experiments is higher.
Although the maternal signal was not eliminated completely, there was a successful
reduction achieved, mainly in case of the P and T waves. Therefore, authors suggest
further research in this field. The combination of the linear and nonlinear adaptive
methods should be used to reduce both, linear and nonlinear distortion.
Acknowledgements
This paper has been elaborated within the framework of the project SP2016/146 of the
Student Grant System, VSB-TU Ostrava, Czech Republic.
References
1. Jagannath, D. J., and Selvakumar, A. I. (2014). Issues and research on foetal electrocardio-
gram signal elicitation. Biomedical signal processing and control, 10, 224-244.
2. Agostinelli, A., Grillo, M., Biagini, A., Giuliani, C., Burattini, L., Fioretti, S., ... and Burat-
tini, L. (2015). Noninvasive fetal electrocardiography: an overview of the signal electro-
physiological meaning, recording procedures, and processing techniques. Annals of Nonin-
vasive Electrocardiology, 20(4), 303-313.
3. Alexander, T. S. (2012). Adaptive signal processing: theory and applications. Springer Sci-
ence & Business Media.
4. Martinek, R., Kelnar, M., Koudelka, P., Vanus, J., Bilík, P., Janku, P., ... and Zidek, J.
(2015). Enhanced processing and analysis of multi-channel non-invasive abdominal foetal
ECG signals during labor and delivery. Electronics Letters, 51(22), 1744-1746.
5. Martinek, R., and Zidek, J. (2012). Refining the diagnostic quality of the abdominal fetal
electrocardiogram using the techniques of artificial intelligence. Journal: Przeglad Elektrot-
chniczny (Electrical Review), 88(12B), 155-160.
6. Least Mean Square (LMS) Adaptive Filter - Developer Zone - National Instruments." Na-
tional Instruments. N.p., 10 June 2009. Web. 26 Feb. 2016.
7. Poularikas, A. D., and Ramadan, Z. M. (2006). Adaptive filtering primer with MATLAB.
CRC Press.
8. Mahfuz, E., Wang, C., and Ahmad, M. O. (2005, May). A high-throughput DLMS adaptive
algorithm. In 2005 IEEE International Symposium on Circuits and Systems (pp. 3753-3756).
IEEE.
9. Martinek, R., Kahankova, R., Skutova, H., Koudelka, P., Zidek, J., and Koziorek, J. (2016,
July). Adaptive signal processing techniques for extracting abdominal fetal electrocardio-
gram. In Communication Systems, Networks and Digital Signal Processing (CSNDSP),
2016 10th International Symposium on (pp. 1-6). IEEE.
10. Jalaleddine, S. M., Hutchens, C. G., Strattan, R. D., and Coberly, W. A. (1990). ECG data
compression techniques-a unified approach. IEEE transactions on Biomedical Engineering,
37(4), 329-343.
11. Martinez, M., Calpe, J., Soria, E., Guerrero, J. F., Camps, G., and Gómez, L. (2001). Meth-
ods to evaluate the performance of fetal electrocardiogram extraction algorithms. In Com-
puters in Cardiology 2001 (pp. 253-256). IEEE.
12. Martinek, R., Kelnar, M., Vojcinak, P., Koudelka, P., Vanus, J., Bilík, P., and Zidek, J.
(2015). Virtual simulator for the generation of patho-physiological foetal ECGs during the
prenatal period. Electronics Letters, 51(22), 1738-1740.
13. Martinek, R., Kelnar, M., Koudelka, P., Vanus, J., Bilik, P., Janku, P., and Zidek, J. (2016).
A novel LabVIEW-based multi-channel non-invasive abdominal maternal-fetal electrocar-
diogram signal generator. Physiological measurement, 37(2), 238.
14. Camps-Valls, G., Martınez-Sober, M., Soria-Olivas, E., Magdalena-Benedito, R., Calpe-
Maravilla, J., and Guerrero-Martınez, J. (2004). Foetal ECG recovery using dynamic neural
networks. Artificial Intelligence in Medicine, 31(3), 197-209.
15. Al-Zaben, A., and Al-Smadi, A. (2005). Extraction of foetal ECG by combination of singu-
lar value decomposition and neuro-fuzzy inference system. Physics in medicine and biology,
51(1), 137.
16. Martinek, R., Kahankova, R., Skukova, H., Konecny, J., Bilik, P., Zidek, J., and Nazeran,
H. (2016). Nonlinear Adaptive Signal Processing Improves the Diagnostic Quality of Trans-
abdominal Fetal Electrocardiography.
17. Czabański, R., Jeżewski, J., Horoba, K., & Jeżewski, M. (2013). Fetal state assessment using
fuzzy analysis of fetal heart rate signals—Agreement with the neonatal outcome. Biocyber-
netics and Biomedical Engineering, 33(3), 145-155.
18. Jezewski, Janusz, Janusz Wrobel, and Krzysztof Horoba. "Comparison of Doppler ultra-
sound and direct electrocardiography acquisition techniques for quantification of fetal heart
rate variability." IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING BME 53.5
(2006): 855.
19. Jezewski, Janusz, et al. "Determination of fetal heart rate from abdominal signals: evaluation
of beat-to-beat accuracy in relation to the direct fetal electrocardiogram." Biomedizinische
Technik/Biomedical Engineering 57.5 (2012): 383-394.
20. Jezewski, J., Horoba, K., Roj, D., Wrobel, J., Kupka, T., & Matonia, A. (2016). Evaluating
the fetal heart rate baseline estimation algorithms by their influence on detection of clinically
important patterns. Biocybernetics and Biomedical Engineering, 36(4), 562-573.