Non-invasive fetal ECG extraction using discrete wavelet transform recursive inverse adaptive algorithm

Abstract

BACKGROUND:

Fetal heart activity adds significant information about the status of the fetus health. Early diagnosis of issues in the heart before delivery allows early intervention and significantly improves the treatment.

OBJECTIVE:

This paper presents a new adaptive filtering algorithm for fetal electrocardiogram (FECG) extraction from the maternal abdominal signal, known in literature as abdominal electrocardiogram (AECG) signal. Fetal QRS complex waves will be identified and extracted accurately for fetal health care and monitoring purposes.

METHODS:

We use discrete wavelet transform recursive inverse (DWT-RI) adaptive filtering algorithm for this objective. Thoracic maternal electrocardiogram (MECG) is used as a reference in the proposed algorithm and FECG components are extracted from AECG signal after suppressing the MECG projections. The proposed algorithm is compared to other typical adaptive filtering algorithms, least mean squares (LMS), recursive least squares (RLS), and recursive inverse (RI).

RESULTS:

Fetal QRS waveforms successful identification and extraction from AECG signal is evaluated objectively and visually and compared to other algorithms. We validated the proposed algorithm using both synthetic data and real clinical data.

CONCLUSIONS:

The proposed algorithm is capable of extracting fetal QRS waveforms successfully from AECG and outperforms other adaptive filtering algorithms in terms of accuracy and positive predictivity.

Keywords

Fetal electrocardiogram (FECG)fetal heart rate non-invasive FECG extraction adaptive filters recursive inverse discrete wavelet transform

1. Introduction

Fetal heart activity, reflected by FECG, is essential to evaluate the status of the fetus heart. It provides information about the health of the fetus, accordingly, early diagnosis of any issues in the heart before delivery allows early intervention and improves the treatment [1]. Doppler ultrasound, direct invasive FECG monitoring, and non-invasive FECG monitoring are some of the most popular techniques to monitor the fetal heart. The former is considered the most common technique for recording fetus heart rate, however, it has some disadvantages. Launching ultrasound waves toward the fetus is classified as an invasive method which is not recommended, especially for extended exposure periods.

In addition, monitoring using the currently available ultrasound sensors, attached to the mother for extended periods, is uncomfortable. Direct invasive FECG monitoring is usually performed during delivery where the electrode is attached to the fetus scalp. Non-invasive FECG techniques depend on collecting signals using surface electrodes from the mother’s body at locations that have information about the FECG signal. Research has been carried out for FECG extraction based on different processing techniques, such as singular value decomposition [2], wavelet analysis [3], independent component analysis or blind source separation [4, 5], adaptive filtering techniques [6, 7], neural networks [8, 9], adaptive neuro-fuzzy inference systems [10], fuzzy logic [11], and time domain analysis [12].

There are various noise sources that affect the noninvasive FECG signal extraction from the AECG signal which makes the process challenging because of the low signal power compared to the other components and noise sources included in the AECG such as MECG component, maternal electromyograms components, respiration, motion artifacts, power line interference and ambient noise. Many of these noise sources can be removed effectively by using classical low pass, high pass and notch filters. FECG signals exhibits a bandwidth of 0.05–120 Hz with fetal heart rate normally between 120 and 160 beats-per-minute (bpm). Using conventional selective filtering techniques to separate FECG from MECG is impossible because of the significant frequency range overlap between FECG and MECG signals.

FECG extraction using adaptive filtering techniques, which is considered as simple and fast approach, usually uses thoracic single channel as a reference signal and one maternal abdominal signal as processed signal. Adaptive filters according to optimization techniques are trained to remove the components related to the thoracic MECG signal in the abdominal signal. Propagation of the signal from the maternal heart to the location of the acquired abdominal signal is nonlinear and the patterns of the abdominal and thoracic ECG waveforms depend on the location of the electrode placement [6].

In spite of the fact that adaptive filter algorithm implemented here is effective in fetal ECG extraction, the major disadvantage is that it requires additional electrodes on the maternal thorax to provide the reference maternal ECG signal. Research has been also done of FECG extraction from the maternal abdominal without a required thoracic signal being used as a reference, e.g. [2, 13]. However, most of the time, both maternal abdominal and thoracic recordings are used for FECG extraction, e.g. [3, 14, 15].

Recursive inverse (RI) algorithm has shown a robust performance compared to other adaptive algorithms that have less order of complexity; i.e. LMS algorithm ( $O(N)$ ), same order of complexity, i.e. RLS algorithm ( $O(N^{2})$ ), or even higher complexity order. When the input signal is highly correlated, most of the adaptive filtering algorithms either fail to converge or converge slowly. Although the RI algorithm had shown relatively high performance under such conditions [16], its performance can be improved further by transforming the input signal. For example, discrete-wavelet transform enhances the performance of the RI algorithm in terms of convergence rate and mean-square-error (mse) value. Nevertheless, this is not the case of the RLS algorithm because it is not a steepest-descent-based method.

In this paper, we use discrete wavelet transform recursive inverse (DWT-RI) adaptive filter algorithm to eliminate the abdominal MECG projection from the maternal AECG signal and estimate the FECG components. Thoracic MECG is considered as the reference signal in the adaptive noise canceller algorithm. The method requires one abdominal recording and one maternal thoracic recording. We show results and analysis for the performance of the proposed algorithm on both synthetic and clinical real data. Performance of the proposed algorithm is compared to those of three adaptive filtering FECG extraction algorithms: the LMS, RLS and RI.

This paper is organized as follows: Section 2 presents the structure of the used adaptive filter in addition to DWT-RI algorithm. Databases and evaluation techniques used in the current study are described in the same section. Section 3 presents the results and the performance of the proposed and compared algorithms on both clinical real and synthetic ECG signals. Discussion of results is carried out in Section 4 and finally conclusions are presented in Section 5.

2. Materials and methods

2.1 ECG data

The proposed algorithm improves the capability of adaptive filtering in extracting the fetal ECG data form maternal abdominal signal while maternal thoracic ECG signal is available as a reference. In order to verify the reliability and feasibility of the proposed algorithm, synthesized data and real clinical ECG data are used.

2.1.1 Synthetic ECG data

Synthetic ECG data for thoracic ECG and abdomen ECG have been generated in Matlab environment according to [17], taking into consideration the differences between FECG and MECG signals [18]. Fetus heart beats are noticeably faster and has a much weaker amplitude than those of the mother. MECG is usually higher in intensity than the fetal ECG by around five to ten times [19]. Maternal heart rate is chosen to be around 90 bpm, while fetal heart rate is around 140 bpm. Peak voltage for the maternal thoracic ECG signal is taken as 3.5 mV compared to 0.25 mV for the fetus ECG. Both signals were generated with a sampling rate of 4 kHz and 40 s period. Thoracic mom’s signal, MECG, has been generated with additive broadband noise with SNR equals 30 dB. Mother abdomen signal has been emulated by the addition of mother thoracic signal with the effect of propagation from the heart to the abdomen represented by an FIR filter with 10 randomized coefficients. Generated fetal ECG signal and uncorrelated gaussian noise to resemble noise sources in the data is also added to the previous component to consist the overall emulated abdomen signal, AECG. Accordingly, abdominal signal is a mixture of the noisy propagated MECG, FECG signal, and noise. Generated MECG, FECG and synthesized mother’s abdomen ECG (AECG) signal are shown in Fig. 1. As noticed, the abdomen signal is dominated by the propagated mother’s ECG signal. Fetal (FECG) to propagated-maternal ECG (pMECG) signal-to-noise ratio (SNR) equals $-$ 24.5 dB according to Eq. (1). This very low SNR is due to the fact that FECG is usually few $\mu$ Vs while the interfering signals are in the range of hundreds of $\mu$ Vs at the surface electrodes used for data acquisition.

$\displaystyle\textit{SNR}=10\log_{10}\frac{\sum\limits_{n}{\left({\textit{FECG% }\left(k\right)}\right)^{2}}}{\sum\limits_{n}{\left({\textit{pMECG}\left(k% \right)}\right)^{2}}}$ (1)

Figure 1.

Synthesized (a) noisy thoracic MECG signal, (b) noisy AECG signal, and (c) FECG reference signal.

2.1.2 Real ECG data

Real ECG data used in our paper was downloaded from a well-known database prepared by De Moor, Database for the Identification of Systems (DaISY Database) [20]. It is used to evaluate the performance of the proposed algorithm and consists of a single dataset of non-invasive potential recording of a pregnant woman. A total of eight channels are included in this dataset, five are abdominal and three are thoracic, with a sample rate of 250 Hz and duration of 10 s each. Thoracic channels measure the same mother’s ECG from the chest but at different locations instantaneously. Abdominal channels measure the same cutaneous abdominal AECG data instantaneously but from different locations on the abdomen. AECG abdominal signal has been tested against all thoracic ECG data, namely, Thoracic1, Thoracic2, and Thoracic3 to test the performance of the proposed algorithm compared to the other algorithms in extracting FECG signal.

2.2 Proposed algorithm

An adaptive filtering algorithm is built to adaptively remove or suppress the maternal ECG signal in the abdominal signal according to the estimated thoracic reference signal, MECG here. The block diagram that represents a noise cancelling adaptive filter for this purpose is shown in Fig. 2. Desired signal here is the AECG, $d(k)$ , reference signal, MECG, is the mother’s thoracic ECG signal. Accordingly, noisy FECG signal is the resulting error and the output of the adaptive filter in this diagram, $e(k)$ . MECG projections are removed adaptively from AECG signal.

Figure 2.

Adaptive noise canceller block diagram used for the proposed algorithm.

Discrete-wavelet transform enhances the performance of the RI algorithm in terms of convergence rate and mse value. The update equation can be written as:

$\displaystyle\textbf{h}(k)=[\textbf{I}-\mu(k)\textbf{R}_{zz}(k)]\textbf{h}(k-1% )+\mu(k)\textbf{p}_{zd}(k)$ (2)

where

$\displaystyle\textbf{R}_{zz}(k)=\beta\textbf{R}_{zz}(k-1)+\textbf{z}(k)\textbf% {z}^{T}(k)$ (3)

and

$\displaystyle\textbf{p}_{zd}(k)=\beta\textbf{p}_{zd}(k-1)+d(k)\textbf{z}(k)$ (4)

are the recursive estimates of the autocorrelation matrix and cross-correlation vector of $z(k)$ , respectively. $\textbf{z}(k)=\textbf{Wx}(k)$ is the transformed input signal, $\textbf{x}(k)$ , and W is the discrete-wavelet transform matrix. $\beta$ is the forgetting factor, $0\ll\beta<1$ , and $\mu(k)$ is the variable step-size given as:

$\displaystyle\mu(k)=\frac{\mu_{o}}{1-\beta^{k}}$ (5)

In Eq. (5), $\mu_{o}$ is a constant constrained as:

$\displaystyle\mu_{0}<\mu_{\textit{max}}=\frac{2(1-\beta)}{\lambda_{\textit{max% }}(\textbf{R}_{zz})}\text{ and }\textbf{R}_{zz}=E[\textbf{z}(k)\textbf{z}^{T}(% k)]$ (6)

The posteriori-error $e(k)$ is given as:

$\displaystyle e(k)=d(k)-y(k)$ (7)

where $y(k)$ is the transformed output of the adaptive filter,

$\displaystyle y(k)=\sum\limits_{i=0}^{N-1}\sum\limits_{i}h_{i}(k)w_{i}(l-k)x(l)$ (8)

where $N$ is the filter order. A summary of the DWT-RI algorithm is provided in Table 1. The DWT process reduces the eigenvalue spread of the autocorrelation matrix of the input signal,

$\displaystyle\chi(R)=\frac{\lambda_{\textit{max}}(\textbf{R}_{zz})}{\lambda_{% \textit{min}}(\textbf{R}_{zz})}$ (9)

Table 1

Summary of DWT-RI algorithm

DWT-RI algorithm

initialize

\mu_{0}

\beta

, h,

\textbf{R}_{zz}

\textbf{p}_{zd}

estimate

\textbf{R}_{zz}(k)=\beta\textbf{R}_{zz}(k-1)+\textbf{z}(k)\textbf{z}^{T}(k)

\textbf{p}_{zd}(k)=\beta\textbf{p}_{zd}(k-1)+d(k)\textbf{z}(k)

e(k)=d(k)-y(k)

\mu(k)=\frac{\mu_{0}}{1-\beta^{k}}

\textbf{h}(k)=[\textbf{I}-\mu(k)\textbf{R}_{zz}(k)]\textbf{h}(k-1)+\mu(k)% \textbf{p}_{zd}(k)

Although LMS algorithm has an advantage over other algorithms in its computational complexity ( $O(N)$ ), it provides poor performance compared to the other algorithms in most applications. Comparing the performances of the RI and RLS algorithms, even they provide the same performance in terms of error and signal detection, and the same order of complexity ( $O(N^{2})$ ), the RI algorithm has less number of Mults./Divs. and Adds./Subs. than the RLS algorithm, see Table 2. On the other hand, the proposed algorithm outperforms the other algorithms in terms of accuracy and signal detection. In terms of computational complexity, DWT needs ( $O(N)$ ) number of computations [21]; hence, the proposed algorithm will have the same computational complexity order of the RLS and RI algorithms ( $O(N^{2})$ ). DWT-RI, where “db5” wavelet is used here, shows higher computational complexity than both RI and LMS, but it shows less computational complexity compared to RLS algorithm as shown in Table 2 and Fig. 3.

Table 2

Computational complexity of the LMS, RLS, RI and DWT-RI algorithms

Algorithm	Mults./Divs.	Adds./Subs.
LMS	2N $+$ 1	2N
RLS	3N ${}^{2}$ $+$ 11N $+$ 9	3N ${}^{2}$ $+$ 7N $+$ 4
RI	2.5N ${}^{2}$ $+$ 3.5N	2N ${}^{2}$ $+$ N
DWT-RI	2.5N ${}^{2}$ $+$ 13.5N	2N ${}^{2}$ $+$ 10N

Figure 3.

Computational complexity of the LMS, RLS, RI, and DWT-RI adaptive algorithms.

Matlab ${}^{\@setsize{\scriptsize}{9.5pt}{\viiipt}{\@viiipt}\textregistered}$ R2014a, specifically adaptive filters toolbox and wavelet toolbox, is used to implement the algorithms and assess their performance. An FIR filter is used to remove high frequency noise accompanying the resulting FECG signal. FIR filters are used because of their linear phase response compared to the infinite impulse response filters and accordingly they do not produce phase distortion for the denoised signal. QRS complex of fetal ECG are detected as the peaks above an adaptive threshold for the filtered fetal ECG signal.

2.3 Performance evaluation

The proposed approach capability in extracting fetal ECG from maternal abdomen signal is evaluated and compared to the performance of other adaptive filtering techniques. Beat-to-beat comparison between the fetal reference QRS complex and the detected QRS complex is done using performance metrics parameters, namely, sensitivity (S), positive predictive value (PPV) and accuracy (Acc) [22], given by Eqs (10)–(12), respectively.

$\displaystyle S=\frac{TP}{TP+FN}\times 100\%$ (10) $\displaystyle\textit{PPV}=\frac{TP}{TP+FP}\times 100\%$ (11) $\displaystyle\textit{Acc}=\frac{TP}{TP+FP+FN}\times 100\%$ (12)

Table 3

Statistical assessment of detected QRS complex waves using adaptive filters of different algorithms compared to the proposed algorithm, RI-DWT using synthetic data

Technique		TP	FN	S (%)	PPV (%)	Accuracy (%)
LMS	( $\mu=$ 0.00001)	83	5	94.3	100	94.3
RI	( $\beta=$ 0.99, $\mu_{o}=$ 0.00001)	85	3	96.6	100	96.6
RLS	( $\beta=$ 0.999)	85	3	96.6	100	96.6
RI-DWT	( $\beta=$ 0.99, $\mu_{o}=$ 0.00001)	88	0	100	100	100

Figure 4.

Recovered noisy FECG signal after applying DWT-RI adaptive filter in addition to the reference fetal ECG signal.

Figure 5.

Recovered low pass filtered FECG using DWT-RI in addition to reference fetal ECG.

Figure 6.

Automatic detection of the fetal QRS complex. QRS complex is automatically registered whenever a peak of the filtered FECG signal is detected above the dynamic threshold.

Figure 7.

FECG extraction: Recovered fetal ECG signals using the LMS algorithm (a) and using DWT-RI algorithm (b) compared to the reference FECG signal.

Figure 8.

DaISY Database: Abdomen2 (a), and Abdomen5 (b), as examples of the abdomen signals and Thoracic3 (c), and Thotracic1 (d) as examples of maternal ECG signals.

where $T P$ , $F P$ and $F N$ represent number of true positives (correctly identified fetal QRS wave), number of false positives (wrongly detected fetal QRS wave) and number of false negatives (missed detected fetal QRS wave), respectively.

3. Results

Three typical adaptive filtering techniques are implemented and tested on both, simulated and real data, in order to evaluate the performance of the proposed method compared to them. The LMS, RLS, and RI algorithms are used for this comparison with number of taps or coefficients set to 15 for all of them. Using DWT-RI adaptive filter algorithm on FECG extraction from AECG signal of SNR $=-$ 24.5 dB with ( $\beta=$ 0.99 and $\mu_{o}=$ 0.00001), the resulting extracted FECG signal in addition to the reference fetal ECG signal are shown in Fig. 4. FIR low pass filter at 200 Hz is applied to remove the high frequency noise from the estimated fetal ECG signal. The calculated denoised fetal ECG signal in addition to the reference fetal ECG signal are presented in Fig. 5. As noticed, the DWT-RI is able to recover the FECG information from AECG signal according to the comparison with the reference FECG. Since QRS complex is considered as the most important wave in the ECG signal, especially in heart rate estimation, the automatic detection of the QRS wave is evaluated by setting a threshold such that whenever a peak is detected above this threshold, a heartbeat is registered. Dynamic threshold has been chosen as four multiples of the absolute average value of the recovered fetal signal. Figure 6 shows the automatically detected peaks according to the dynamic threshold. DWT-RI was able to extract QRS complex waves successfully even if the fetal QRS is synchronized with the maternal QRS complex. To assess the performance of DWT-RI adaptive filter in fetal ECG recovery and automatic detection of the fetal QRS complex wave, TPs, FPs, and FNs are calculated and presented in Table 3. Performance of DWT-RI algorithm has been compared to other adaptive filter algorithms, namely, LMS, RLS and RI. In general, all tested algorithms were able to extract FECG signals and can correctly separate the mixed signal, though, sometimes, QRS wave is missed by some techniques. The LMS adaptive filter algorithm has been compared to the DWT-RI algorithm performance. Figure 7 shows one of the cases where some QRS complex waves were missed by the LMS algorithm but still recovered by DWT-RI, e.g. at time 16.83 s and 22.87 s, when compared to the reference FECG signal.

Table 4
Statistical assessment of detected QRS complex waves using adaptive filters of different algorithms compared to the proposed algorithm, RI-DWT using real clinical data

Technique		TP	FP	S (%)	PPV (%)	Accuracy (%)
LMS	( $\mu=$ 0.00001)	63	39	100	61.8	61.8
RI	( $\beta=$ 0.99, $\mu_{o}=$ 0.00001)	63	9	100	87.5	87.5
RLS	( $\beta=$ 0.999)	63	9	100	87.5	87.5
RI-DWT	( $\beta=$ 0.99, $\mu_{o}=$ 0.00001)	63	6	100	91.3	91.3

Figure 9.

Extracted fetal ECG from DaISY clinical data using adaptive filter based on DWT-RI algorithm.

Figure 10.

DWT-RI performance in FECG extraction from AECG by detecting fetal QRS complex even when occurring almost at the same time of the maternal QRS.

Figure 11.

Performance of DWT-RI on real clinical data (a), compared to the LMS (b), and RLS (c) algorithms. Estimated FECG signal is multiplied by 4 to show both AECG and FECG signals clearly on the same graph. “R” labeled arrows determine the instants of the fetal QRS complex waves.

Sensitivity and PPV values for the proposed algorithm are calculated and presented in Table 3 in addition to the performance of the other adaptive filtering algorithms. Statistical evaluation is done for the entire period of the data starting after 2 s, i.e. from 2 s to 40 s, which contains 88 fetal QRS complex waves. Clinical data from DaISY database has been used to evaluate the performance of the proposed adaptive filter algorithm compared to other algorithms. Examples of the maternal abdominal and thoracic signals are shown in Fig. 8. As noticed, abdomen signals are significantly smaller than thoracic ones and slightly delayed compared to them by around few milliseconds. This is because the thoracic measurements are the direct measurement of the maternal ECG while abdomen measurement is the propagated version of the maternal ECG and accompanied by the weak noisy FECG from the abdomen. AECG abdominal signal has been tested against all thoracic ECG data, namely, Thoracic1, Thoracic2, and Thoracic3, in order to evaluate the performance of the proposed adaptive filter algorithm on real clinical data compared to other algorithms. Statistical assessment of the performance of the detected QRS complex waves using adaptive filters algorithms under comparison on clinical real data is presented in Table 4. Figure 9 shows the results of applying DWT-RI adaptive algorithm on thoracic and abdomen signals, where the projection of MECG has been significantly suppressed. All QRS components of FECG have been extracted well using this algorithm even when it is occurring almost at the same time of the MECG as shown in Fig. 10. Figure 11 shows a comparison between magnified parts of the resulting extracted fetal ECG signal from the abdominal signal using DWT-RI algorithm compared to the LMS and RLS algorithms. In spite that all of them are able to keep the fetal ECG, DWT-RI is more successful in suppressing the projections of MECG signal and accordingly isolate and extract FECG more successfully as noticed in Fig. 11 which shows one of the examples for these cases. This is very clear especially when the performance is compared among the algorithms if the fetal QRS is very close to the maternal QRS complex. Statistical analysis on real clinical data shows that, while all tested algorithms gave sensitivity of 100% in detecting the fetal QRS waves, positive predictivity values for the LMS, RLS, RI, and DWT-RI algorithms are 61.8%, 87.5%, 87.5%, and 91.3%, respectively.

4. Discussion

Non-invasive FECG extraction from AECG signal is a challenging task because of the very low SNR due to the fact that FECG is usually few $\mu$ V compared to the interfering abdomen signal that is in the range of hundreds of $\mu$ V. Furthermore, many sources of noise are included in the abdomen signal in addition to the MECG and FECG information which makes the task more difficult. In order to suppress the accompanying high frequency unwanted components and artifacts, FIR low pass filter at 200 Hz is applied to the estimated fetal ECG signal. Performance evaluation of DWT-RI algorithm and comparison to other typical adaptive filter algorithms, namely, the LMS, RLS, and RI, have been done on two kinds of data, synthetic data and real clinical data. The DWT process reduces the eigenvalue spread of the autocorrelation matrix of the input signal. This, in turn, enables the proposed algorithm to outperform the RI algorithm in terms of error and signal detection. Both RI and the proposed algorithm have advantage over the LMS algorithm; which is the variable step-size, as stated in Eq. (5). The forgetting factor, $\beta$ , usually is very close to unity, hence, initially the step-size of the RI algorithm is relatively large which provides fast convergence of the algorithm. On the other hand, when the number of iterations increases, the step-size decreases, which guarantees convergence to a low mse value. Comparing the proposed and RI algorithms to the RLS algorithm, they use an update estimate of the autocorrelation matrix, not its inverse (the case in RLS), which provides a guaranteed stable performance for the RI algorithms. In addition, the RLS algorithm is sensitive to the initialization of the inverse autocorrelation matrix estimate (depending on the problem configuration) [23], where the initialization of the autocorrelation matrix for the proposed and RI algorithms is independent of the problem configuration.

Many of the previously proposed methods to extract FECG signal from the AECG signal in literature have been evaluated using synthetic or simulated signals, e.g. [24, 25, 26], or/and used real clinical datasets of very small recordings, e.g. [14, 27, 28]. The effectiveness of the algorithms is tested objectively and visually in the synthetic data. Based on the statistical evaluation on synthetic ECG data, only DWT-RI recorded no false positives nor false negatives among the total of 88 fetal heart beats among the algorithms tested in this paper. Table 3 shows the statistical evaluations for the studied algorithms.

Performance of the proposed algorithm has been also tested visually and objectively and results were presented for real clinical. All tested algorithms: LMS, RLS, RI, and DWT-RI were able to detect most of the fetal QRS complex waves but DWT-RI outperformed the other tested algorithms in both kinds of data visually and objectively. However, based on the resulting FECG signals from the real data, none of the tested methods could remove the MECG QRS component completely from AECG signal but they have been suppressed significantly. DWT-RI was more successful in suppressing the maternal QRSs as was shown in Fig. 10. Statistical evaluation using sensitivity, positive predictivity and accuracy in addition to visual inspection are used to quantitatively prove those results. According to the evaluation done in this paper on the clinical data, DWT-RI was able to extract all fetal QRS complex patterns even when they are very close to the maternal QRS waves and outperforms the other tested methods in suppressing the maternal QRS from AECG recording and extracting fetal QRS waves successfully.

5. Conclusions

In this paper, DWT-RI algorithm has been presented and evaluated to extract FECG from AECG signal. Objective and visual evaluation of the proposed algorithm shows promising performance in such kinds of applications. DWT-RI is capable of extracting main components of FECG signal, namely, fetal QRS complex waves, from AECG. Datasets consists of thoracic, MECG signal, and abdominal, AECG signal, were used to evaluate the performance of the proposed algorithm where one of the datasets is synthetic data and the other is real clinical data. DWT-RI shows an overall accuracy in identifying fetal QRS waveforms of 100% on synthetic and real data and significant suppression of MECG projections in AECG signal and accordingly successful extraction and identification of all fetal QRS in both kinds of data. The major drawback in the proposed algorithm, which is a common drawback in almost all algorithms using a reference signal, is the difficulty in identifying FECG components from AECG signal in very noisy environments.

Footnotes

Conflict of interest

The authors declare that they have no conflict of interest in relation to this work.

References

Deam

. The use of fetal electrocardiogram. J. Obstet. Gynecol. 1994; 101: 9-17.

Liu

Chen

Sun

. Detection of fetal ECG R wave from single-lead abdominal ECG using a combination of RR time-series smoothing and template-matching approach. IEEE Access. 2019; 7: 66633-43.

Sutha

Jayanthi

. Fetal electrocardiogram extraction and analysis using adaptive noise cancellation and wavelet transformation techniques. J. Med Syst. 2018; 42: 21.

Lieven

Bart

Joos

. Fetal electrocardiogram extraction by blind source subspace separation. IEEE Trans Biomed Eng. 2000; 47: 567-72.

Karimi Rahmati

Setarehdan

Araabi

. A PCA/ICA based fetal ECG extraction from mother abdominal recordings by means of a novel data-driven approach to fetal ECG quality assessment. J Biomed Phys Eng. 2017; 7: 37-50.

Zhang

Mumini

Samuel

Ivanov

Wang

. A novel technique for fetal ECG extraction using single-channel abdominal recording. Sensors. 2017; 17: 457.

Singh

Kaur

. Adaptive filter design for extraction of fetus ECG signal. Proc. of the Second Int. Conf. on Computer and Communication Technologies. 2016; pp. 109-17.

Camps-Valls

Martnez-Sober

Soria-Olivas

Magdalena-Benedito

Calpe-Maravilla

Guerrero-Martnez

. Foetal ECG recovery using dynamic neural networks. Artif. Intell. Med. 2004; 31: 197-209.

Parameshwari

Ponrani

Devi

. Foetal ECG extraction using BPN and UWT. IJBET. 2016; 22: 1-13.

10.

Assaleh

. Extraction of fetal electrocardiogram using adaptive neuro-fuzzy inference systems. IEEE Trans. on Biomed. Eng. 2007; 54: 59-67.

11.

Azad

KAK

. Fetal QRS complex detection from abdominal ECG: a fuzzy approach. Proc. IEEE Nordic Signal Processing Symposium. 2000; pp. 275-8.

12.

Ibrahimy

Ahmed

Mohd Ali

Zahedi

. Real-time signal processing for fetal heart rate monitoring. IEEE Trans. Biomed. Eng. 2003; 50: 258-62.

13.

Karvounis

Tsipouras

Fotiadis

Naka

. An automated methodology for fetal heart rate extraction from the abdominal electrocardiogram. IEEE Trans. Biomed. Eng. 2007; 11: 628-38.

14.

Zarzoso

Nandi

. Noninvasive fetal electrocardiogram extraction: Blind separation versus adaptive noise cancellation. IEEE Trans. Biomed. Eng. 2001; 48: 12-18.

15.

Assaleh

Al-Nashash

. A novel technique for the extraction of fetal ECG using polynomial networks. IEEE Trans. Biomed. Eng. 2005; 52: 1148-52.

16.

Salman

Kukrer

Hocanin

. Recursive inverse adaptive algorithm: A second-order version, a fast implementation technique, and further results. Signal, Image and Video Processing. 2015; 9: 665-73.

17.

“Mathworks”. Simulink model for fetal ECG extraction (HDL compatible algorithm). 2012. https://www.mahworks.com/matlabcentral/fileexchange/35328.

18.

Kumar

Dewan

Singh

Kurukshetra

. Realtime monitoring system for ECG signal using virtual instrumentation. IEEE Transactions on Biology and Biomedicine. 2006; 3: 638-43.

19.

Sameni

Shamsollahi

Jutten

Clifford

. A nonlinear bayesian filtering framework for ECG denoising. IEEE Trans. Biomed. Eng. 2007; 54: 2172-85.

20.

De Moor

De Gersem

De Schutter

Favoreel

. DAISY: A database for identification of systems. Journal A, Spec. Issue Comput. Aided Control Systems Design. 1997; 38: 4-5.

21.

Rioul

Duhamel

. Fast algorithms for discrete and continuous wavelet transforms. IEEE Transactions on Information Theory. 1992; 38: 569-86.

22.

Samuel

Asogbon

Sangaiah

Fang

. Integrated decision support system based on ANN and Fuzzy_AHP for heart failure risk prediction. Expert Syst. Appl. 2016; 68: 163-72.

23.

Haykin

. Adaptive filter theory. 4th ed. Prentice-Hall: Upper Saddle River, 2002.

24.

Barros

. Extracting the fetal heart rate variability using a frequency tracking algorithm. Neurocomputing. 2002; 49: 279-88.

25.

Zhang

. Extraction of a source signal whose kurtosis value lies in a specific range. Neurocomputing. 2006; 69: 900-4.

26.

Khamene

Negahdaripour

. A new method for the extraction of fetal ECG from the composite abdominal signal. IEEE Trans. Biomed. Eng. 2000; 47: 507-16.

27.

Camps-Valls

Martinez-Sober

Soria-Olivas

MagdalenaBenedito