System for adaptive extraction of non-invasive fetal electrocardiogram

Applied Soft Computing 113 (2021) 107940 Contents lists available at ScienceDirect Applied Soft Computing journal homepage: www.elsevier.com/locate/asoc System for adaptive extraction of non-invasive fetal electrocardiogram Katerina Barnovaa, Radek Martineka, Rene Jarosa, Radana Kahankovaa,∗, Khosrow Behbehanib, Vaclav Snaselc aDepartment of Cybernetics and Biomedical Engineering, Faculty of Electrical Engineering and Computer Science, VSB–Technical University of Ostrava, 708 00, Ostrava-Poruba, Czechia bDepartment of Bioengineering, University of Texas at Arlington, Arlington, TX, USA cDepartment of Computer Science, Faculty of Electrical Engineering and Computer Science, VSB–Technical University of Ostrava, Ostrava-Poruba, Czechia article info Article history: Received 3 May 2021 Received in revised form 19 August 2021 Accepted 22 September 2021 Available online 8 October 2021 Keywords: Adaptive filtration Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) Extraction algorithms Fast transversal filter (FTF) Fetal electrocardiography Fetal heart rate (fHR) Hybrid system Independent component analysis (ICA) Non-invasive fetal monitoring abstract This study aimed to find the most suitable combination of adaptive and non-adaptive methods for extraction of non-invasive fetal electrocardiogram (NI-fECG) using signals recorded from the mother’s abdomen. Among the nine methods considered, the combination of independent component analysis (ICA), fast transversal filter (FTF), and complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) proved to be the most effective for the extraction of fECG from abdominal recordings. This combined method was suitable due to both being effective in extracting fECG and being less computationally complex. Further, so far, FTF and CEEMDAN methods have not been extensively tested for fECG extraction, and in particular, have not been examined as a hybrid method. The ICA-FTF-CEEMDAN hybrid algorithm was tested on two patient databases: Fetal Electrocardiograms, Direct and Abdominal with Reference Heartbeats Annotations (FECGDARHA) and PhysioNet Challenge 2013. The evaluation of the accuracy of fQRS complexes detection was performed using the following parameters: accuracy (ACC), sensitivity (SE), positive predictive value (PPV), and F1 score. The fetal heart rate (fHR) determination accuracy was evaluated using Bland–Altman plots and fHR traces. When testing on the FECGDARHA database, average values of ACC =92.98%, SE =95.33%, PPV =96.4% and F1 =95.86% for detection fQRS were achieved. The error in estimating the fHR was −1.02 ±7.02 (µ±1.96σ) bpm. When testing on the Challenge 2013 database, average values of ACC =78.47%, SE =82.06%, PPV =87.90% and F1 =84.62% for fQRS detection were achieved, and the error in estimating the fHR was −6.62 ±10.33 (µ±1.96σ) bpm. In addition, a non-invasive morphological analysis (ST analysis) was performed on the records from the FECGDARHA database, which was accurate in 7 of 12 records with values of µ < 0.03 and values of ±1.96σ < 0.04. ©2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1. Introduction Fetal hypoxia is one of the most common causes of perinatal morbidity and mortality [1,2]. Fetal hypoxia is a condition that develops over the course of pregnancy. In the initial stage (hypoxemia), there is a decrease in oxygen saturation in the arterial blood, but cellular and organ functions are not affected. When hypoxemia worsens, the fetal’s adaptive and compensatory mechanisms are triggered, allowing them to survive in this state for several days/weeks without any significant harm [3]. If the oxygen saturation of fetal blood is further reduced, its deficiency begins to manifest itself in peripheral tissues, and the anaerobic mechanism is triggered resulting in hypoxia. The fetus may remain in this condition for several hours without permanent ∗Corresponding author. E-mail address: [email protected] (R. Kahankova). damage [1,4]. However, a condition that requires an immediate resolution in a matter of minutes is asphyxia. Blood oxygen saturation is extremely low, and anaerobic metabolism takes place not only in peripheral tissues but also in the heart and brain, followed by systemic collapse with heart and central nervous system failure [5]. Accurate and timely diagnosis of hypoxia is, therefore, a primary goal in fetal monitoring. The first efforts to monitor the fetus were based on intermittent auscultation of fetal heart sounds and calculation of fetal heart rate (fHR) [6]. Thanks to advances in science and technology, the first monitors based on phonocardiography monitoring heart sounds were invented in the second half of the 20th century. These devices were not able to distinguish between the maternal and fetal heart sounds, so the automatic determination of fHR was not possible [7]. A major breakthrough came with the invention of cardiotocography (CTG), which is based on fetal monitoring using simultaneous https://doi.org/10.1016/j.asoc.2021.107940 1568-4946/©2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/bync-nd/4.0/). K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 recording of fHR and uterine contractions [8]. By introducing this non-invasive method in the 1960s and enabling continuous monitoring in clinical practice [9], there was a significant reduction in fetal deaths. Unfortunately, CTG proved to suffer from a high rate of false positives (up to 60%) [10–12], and the detection of hypoxia based only on the assessment of fHR turned out to not be accurate enough [10,13]. This led to a high number of unnecessarily performed caesarean sections, which are a great burden with added risks for both the mother and the fetus [13– 15]. More accurate fetal monitoring was enabled by Neoventa Medical (Molndal, Sweden) devices; this company was the first to introduce the innovative STAN S31 ST segment analyzer into the clinical practice in 2007 [9]. This device records the fetal electrocardiogram (fECG) by attaching a bipolar spiral electrode on the fetal scalp (i.e., fetal scalp electrode, FSE) and analyzing the changes in the morphology of the fECG signal; in particular the T/QRS ratio (known as ST analysis, STAN) [16,17]. Therefore, simultaneous fHR and STAN monitoring help to increase the accuracy of detecting fetal hypoxia, hence reducing the number of unnecessary caesarean sections [18]. Invasiveness of the method entails the risk of infection when placing the scalp electrode on the fetus and the serious limitation that it can be implemented only after the rupture of the amniotic fluid membrane during childbirth method of monitoring fECG less desirable [19,20]. For these reasons, attention has recently been focused on the development of non-invasive versions of fECG. This method is based on recording the electrical activity of the fetal heart using electrodes that are placed on the mother’s abdomen. The possibility of recording fECG during pregnancy and childbirth, non-invasiveness and greater safety of the method are significant advantages of this method [21]. Nevertheless, when using this technique to record fetal cardiac activity, maternal cardiac activity is mixed with the recorded signal with a significantly larger magnitude than the fECG magnitude. Moreover, the maternal and fetal components overlap in both the time and frequency domains, so extracting a good quality fECG continues to be a challenging task [19,21]. The benefit of our proposed hybrid method enabling it to achieve high quality extraction is that it combines the advantages of three techniques: independent component analysis (ICA), fast transversal filter (FTF) and complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN). The advantage of the ICA method is that it requires solely abdominal signals (aECG) at the input, from which it is able to estimate the maternal ECG (mECG) and the aECG signal with enhanced fetal component. Therefore, there is no need to use another chest lead to record the mECG signal. In clinical practice, mECG signal recorded from mother’s chest is often of a poor quality and the measurement itself is uncomfortable since it decreases her mobility. Therefore, using the mECG and aECG signals estimated by means of the ICA method leads to better extraction with FTF algorithm, which is prone to poor quality input signals. The adaptive FTF algorithm excels in eliminating mECG almost completely since it is able to adapt to the nature of the signal on its input. The CEEMDAN method can smooth the residue of the mECG component in the resulting fECG signal without distorting its morphology, as is the case with other smoothing methods, such as the wavelet transform (WT) [22]. Thus, the use of solely aECG signals, quality extraction of fECG using adaptive filtering, and smoothing the resulting fECG signal, while remaining its morphology and thus allowing ST segment analysis, is the main benefit of this algorithm. In addition, the results of the study showed that with the help of well-scanned non-invasive signals, it is possible to obtain the same accurate results of fHR and ST analysis as in the case of the invasive variant. Another benefit is its high computing speed, which would allow its implementation in real-time operating devices. 2. State of the art There are numerous filtration methods that can be applied to extract useful information from the non-invasive fECG recording. These are, for example, methods based on the blind source separation (BSS) technique [23–25], which can extract original source signals from a signal mixture. This group of methods also include the popular ICA [26–28] and the principal component analysis (PCA) [26,29]. Another important method for fECG extraction is the application of WT [30–32] or the template subtraction method (TS), which is based on subtracting the template (parent component) from the abdominal signals [33,34]. More advanced methods include the application of artificial neural networks (ANNs) [35,36] that are able to learn and analyze complex data. In this category, the adaptive linear network (ADALINE) [37, 38], which is a single-layer artificial neural network that uses linear activation functions to update its weights, is also worth mentioning. Recent studies, [39–45], have demonstrated that hybrid systems that combine several algorithms are more effective in extracting accurate fECG compared with using only a single algorithm. We have previously presented such methods. For instance, in [22] two hybrid methods combining ICA, recursive least squares (RLS) and WT (i.e., ICA-RLS-WT) as well as a combination of ICA, adaptive neuro-fuzzy interference system (ANFIS) and WT (i.e., ICA-ANFIS-WT) were presented. In [46], we presented procedures based on combining empirical mode decomposition (EMD) with some of the other methods that resulted in ICA-EMD, ICA-EMD-WT and ICA-RLS-EMD. The best results were obtained with ICA-RLS-EMD method, but for some recordings, the extraction was not efficient enough. The aim of this study was to find suitable alternatives to RLS and EMD methods that will help in to achieving more accurate extraction. These assumptions were met by the FTF and CEEMDAN methods and therefore they were selected for implementation and testing. A comparison of the main advantages and limitations of the individual fECG extraction methods that have been used in the past for fECG extraction are summarized in Table 1. 2.1. Adaptive filters Adaptive filters are very popular and frequently used as signal processing tools nowadays. These filters are characterized by the ability to automatically adjust the adaptive coefficients of the filter. This occurs by minimizing the error signal, which is defined as the difference between the desired output and the actual output of the system [47,48]. An adaptive system usually consists of two parts, a digital filter, and an adaptive algorithm for adjusting the filter coefficients [49]. Currently, there is a large number of types of adaptive filters, an overview of which is summarized in [50]. Adaptive filters can generally be divided into two groups, based on the approach they use to optimize the error function. These are the stochastic gradient adaptation methods and the optimal recursive adaptation methods [47,50]. The stochastic gradient adaptation group includes the least mean square (LMS) method, which, due to its simplicity and low computational complexity, is one of the most widely used techniques. The step size parameter determines the convergence rate and it also affects how fast and how close the adaptive filter approaches the desired output values [51,52]. The normalized LMS (NLMS) algorithm was created by modifying the standard LMS algorithm; it excels due to its simplicity and robust performance. The optimal recursive adaptation methods can include the RLS algorithm, the advantage of which is excellent performance, when working in time-varying environments, and fast convergence. There are some disadvantages for these methods in the 2 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 Table 1 Comparison of advantages and disadvantages of different fECG extraction algorithms. Method Advantages Limitations ICA, PCA Stable To achieve greater accuracy of algorithms, it is necessary to use a larger number of quality input signals WT Fast Distortion of signal morphology TS Simple mQRS detection accuracy significantly influences the overall performance ANNs, ANFIS, Precise The performance of the algorithm depends on the ADALINE correct setting of a large number of parameters RLS Precise As the filter order increases, the computational speed decreases FTF Fast The performance of the algorithm strongly depends on the quality of the input signals EMD Stable The extraction efficacy may be adversely affected by mode mixing CEEMDAN Stable As the number of ensemble trials increases, the computational speed decreases form of increased computational complexity [53]. This group can also include a FTF, which was designed specially to reduce the computational complexity of the RLS filter while maintaining the speed of convergence. Its concept is structurally based on four different filters working simultaneously on one task [54–56]. For the purposes of extracting fECG, the LMS, NLMS, DLMS, BLMS filters were implemented and compared in the past in [48]. The best results were achieved using the BLMS filter, and on the other hand, the worst results were obtained using the NLMS filter. The efficacy of LMS a NLMS methods was also compared in [35], however, in this case, no significant differences were found between these filter performances. The authors of the study [57] tested NLMS and RLS filters to extract fECG. It has been proven that better results were obtained using the RLS algorithm, and, in addition, it converged faster than the NLMS algorithm. The RLS filter outperformed also the LMS algorithm in [58] according to the evaluation based on the sensitivity (SE), positive predictive value (PPV) and F1 metrics. The LMS algorithm was effective in combination with the ICA [41], the WT [32,59,60], or with LMS and ANN [61]. An interesting approach is also combining multiple adaptive filters, as proposed in [62]. The best results were achieved using a combination of LMS and RLS algorithms. 2.2. Empirical mode decomposition based methods The EMD method was proposed in [63] for addressing nonlinear and non-stationary issues. Its principle is based on the decomposition of the input complex signal into simpler signals, which are called intrinsic mode functions (IMFs). By summing all IMFs, the original signal can be reconstructed [63–65]. The advantage of EMD lies in its high speed compared to other versions of this method. Its disadvantage, on the other hand, lies in the limitation called mode mixing which affects the resulting extraction quality [64]. This problem is due to the fact that one of the IMFs contains several components with different frequencies, or components with similar frequencies are contained in several different IMFs [66]. To address this shortcoming, the ensemble empirical mode decomposition (EEMD) method was proposed in [67]. This is based on the principle of adding independent Gaussian white noise to the input signal and performing several ensemble trials. Each trial uses the same procedure as EMD. The resulting IMFs are obtained by averaging the corresponding IMFs from all ensemble trials [64,68]. By applying the EEMD method, the risk of mode mixing occurrence can be eliminated, but a new problem arose when reconstructing the original signal. This reconstructed signal contained significant residues of Gaussian white noise [64,69]. To prevent this while preserving the benefits of the EEMD method, the complementary ensemble empirical mode decomposition (CEEMD) was designed in [69]. The principle of CEEMD is the same as EEMD, but instead of fully independent Gaussian white noise, complementary paired positive and negative white noise is generated, which leads to mutual elimination of noise residues in the resulting signal [64,69]. The low computational speed of the algorithms, which decreases with the increasing number of ensemble trials, is another disadvantage of the EEMD method, as well as CEEMD [64]. For this reason, in [70], CEEMDAN, which uses a smaller number of iterations (i.e., nearly a half), thus increasing the computational efficiency of the algorithm, was proposed [70]. For fECG extraction, the EMD method was successfully tested in combination with ICA in [71] or in combination with multiple signal classification [72]. In [40], the authors concluded that the system combining the EEMD, ICA, and wavelet shrinkage showed to be most effective in suppressing maternal component compared to the system combining EMD and wavelet shrinkage. The performance of the EMD, EEMD and CEEMD methods was compared for fECG extraction in combination with correlation analysis in [73]. The CEEMD method provided better frequency resolution of IMFs than EMD or EEMD and more accurate reconstruction of the original signal. An accurate fECG extraction was also achieved using CEEMDAN in [74]. According to [75], the EMD and EEMD methods were outperformed by the CEEMDAN algorithm, since it could provide better frequency separation of the extracted IMFs. A comparison of the basic representatives of adaptive filters and EMD-based methods is summarized in Table 2. The evaluation of individual parameters was achieved as follows: 1. Performance: the performance was evaluated using the high, medium, and low categories in terms of the accuracy of fHR determination of the extracted signals. (a) High: the method was able to filter out all disturbances highly effectively and, in the statistical evaluation of fHR, values ≤95% were achieved using the ACC, SE, PPV and F1 parameters. (b) Medium: the method was able to filter out most interference, but some interference residues led to ACC, SE, PPV and F1 parameter values ≤80% in the statistical evaluation of fHR. (c) Low: the method was not able to sufficiently eliminate the interference and, in the statistical evaluation of fHR, values <80% were achieved using the ACC, SE, PPV and F1 parameters. 2. Computational complexity: the parameter was assigned three levels of high, medium, and low categories, based on subjective evaluation. (a) High: reflects a computationally complex method that cannot be used in real-time applications. 3 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 Table 2 Performance parameters of basic adaptive filters and EMD-based methods. Algorithm Performance Computational complexity References Adaptive methods LMS Medium Low Wu et al. [32], Camps et al. [35], Martinek et al. [47], Kaur et al. [51], Shengkui [52], Swarnalatha et al. [62], Gupta et al. [41], Behar et al. [58], Lima-Herrera et al. [59], Ziani et al. [60], Kaleem et al. [61] NLMS Medium Low Camps et al. [35], Martinek et al. [48], Kaur et al. [51], Swarnalatha et al. [62], Liu et al. [57] RLS High Medium Martinek et al. [47], Diniz [53], Swarnalatha et al. [62], Behar et al. [58], Liu et al. [57] FTF High Low Cioffi et al. [54], Setareh et al. [55], Bessekri et al. [56] EMD-based methods EMD Low Low Huang et al. [63], Ren et al. [64], Gao et al. [66], Liu et al. [76], Azbari et al. [73], Taralunga et al. [71], Wei et al. [72] EEMD Medium High Ren et al. [64], Wu et al. [67], Chang [68], Liu et al. [76], Azbari et al. [73], Liu et al. [40] CEEMD Medium High Ren et al. [64], Yeh et al. [69], Liu et al. [76], Azbari et al. [73] CEEMDAN Medium Medium Ren et al. [64], Torres et al. [70], Colominas et al. [77], Liu et al. [76], Li et al. [78], Queyam et al. [75], Muduli et al. [74] (b) Medium: denotes a slightly more computationally demanding method that can be used in real-time applications after optimization. (c) Low: assigned to a computationally not complex method that can be used in real-time applications. 3. Materials and methods This section describes the methods implemented within the hybrid system. Based on the literature search and study of the issue, ICA, FTF, and CEEMDAN were selected. Furthermore, the proposed hybrid system, its optimal settings, the parameters used to evaluate the performance of the hybrid system and the database used for testing each method are described. 3.1. Independent component analysis The ICA extraction technique belongs to the group of blind source separation methods. Its principle is based on the decomposition of the input mixed signal into the original separate source components [26,79–81]. In the case of fECG extraction, the input mixed signal represents the aECG signal recorded from the mother’s abdominal wall, and the source components consist of fECG, mECG and noise. To perform the extraction correctly, the assumptions that the source components are non-Gaussian signals and are statistically independent of each other must be met [79, 82]. More accurate source signals extraction can be achieved using a larger number of input signals. In the past, several extended versions of the ICA method were presented: the computationally efficient FastICA algorithm [83], efficient version of the FastICA (EFICA) [84] or joint approximate diagonalization of eigen-matrices (JADE) [28]. In this study, the FastICA version using a fixed-point iteration scheme was applied to obtain results faster than the classical version. More detailed information on the ICA method can be obtained in [26–28,79,82]. 3.2. Fast transversal filter The goal of the adaptive FTF filter is to minimize the error signal between the desired and actual output by automatically adjusting the adaptive coefficients. The FTF filter achieves the same performance as the RLS, but it possesses a much higher computational speed. The FTF filter is not burdened by a linear increase in the computation time with increasing filter order, as is the case with RLS [55,85]. Therefore, its use is a very suitable solution for real-time applications. The principle of the filter is based on the use of four transversal filters working simultaneously on one task. These four transversal filters designed to compute update quantities as described in [55,86] include: •Forward prediction transversal filter – this block calculates the forward filter weights in such way that minimizes the least-square error of the subsequent sample based on the previous ones. To calculate the error of estimation for the forward prediction, the filter uses a priori and a posterior filter weights as well as the minimum weighted least squares error. •Backward prediction transversal filter – calculates the backward filter weights in such way that minimizes the leastsquare error of the n−Mth sample using the vector input xb(n)= [x(n)x(n−1)x(n−2) . . . x(n−M+1)]T. As with forward prediction, this filter calculates the minimum weighted least squares error and the posterior and prior filter weights. •Gain computation transversal filter – is used for recursive computation of the gain vector. The gain vector is used to update the forward, backward, and joint-process estimation filter weights. It also provides recursive computation of the conversion factor [55]. •Joint-process estimation transversal filter – determines the filter weight so that the error between the desired and actual signal is as small as possible. Thus, this block calculates filter 4 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 weights in the same way as the adaptive algorithms used earlier. The disadvantage of the algorithm is the instability after a finite number of iterations. However, several approaches [85,87] have been proposed to solve this problem. Usually, these approaches are based on expressing the update equations in different forms [86]. The detailed description of this filter is very extensive, and a complete description of the algorithm can be obtained in [54–56,85,86]. 3.3. Complementary ensemble empirical mode decomposition with adaptive noise The principle of the CEEMDAN method is based on the decomposition of a more complex input signal into simpler signals, called IMFs [70,88]. The method builds upon the previously presented EEMD method [67], which in turn is built upon EMD procedures [63]. The main idea is to add independent Gaussian white noise to the input signal and perform several ensemble trials. The resulting IMFs are obtained by averaging the corresponding IMFs from all ensemble trials [64,68]. The CEEMDAN method differs in the following aspect: the addition of paired positive and negative adaptive white noise to the input signal, which better eliminates the mode mixing issue, and, also, the resulting IMFs are calculated sequentially. That is, each IMF is used to calculate the succeeding IMF [76,77]. This results in a smaller number of iterations which leads to an increase in the computational speed of the algorithm [70]. The algorithm can be summarized as follows: •First, the number of ensemble trials realizations Nand the standard deviation of the added noise series Nstd are set. •Gaussian noise w(t) with unit variance is added to the original signal x(t). s(t)=x(t)+β0w(t),(1) where β0is the level of noise. •Signal s(t) is decomposed N-times using the EMD method and the first IMF1(t) is obtained by averaging all IMF1(t) related to Ntrials. The EMD method is based on identifying local minima and maxima in the signal s(t) followed by detecting the upper and lower envelopes of the signal using a cubic spline by combining all maxima and minima, respectively. The mean of the envelopes is then subtracted from the original signal x(t) and IMF1(t) is obtained as follows: IMF1(t)=x(t)− ⟨M(s(t))⟩,(2) where ⟨M(s(t))⟩represents the mean of the envelopes of the signal s(t). •The residual signal r1(t) is calculated by subtracting the first extracted IMF1(t) from the original signal x(t). r1(t)=x(t)−IMF1(t).(3) •The second IMF2(t) is extracted in the same manner as the first one, but instead of the original signal x(t), the residual signal r1(t) is used. IMF2(t)=r1(t)− ⟨M(r1(t)+β1w(t))⟩.(4) •To obtain additional IMFs, the entire procedure is repeated, but rk(t) is used as the further residual after the kth decomposition, where k=1,2,...,K. The IMFk(t) is obtained as: IMFk(t)=rk−1(t)− ⟨M(rk−1(t)+βk−1w(t))⟩.(5) •The whole procedure is repeated until the residual signal cannot be extracted (i.e. the residual signal is a monotone function, a function with only one extremum or a constant). The motivation to use ICA, FTF and CEEMDAN methods in fECG extraction system and post-processing was following: 1. Fetal ECG extraction – accurate fECG extraction is a very challenging task in which conventional techniques are not efficient enough. Adaptive algorithms (especially RLS) have proven to be very promising for fECG extraction in our previous research [22,46]. The RLS algorithm was able to extract fECG with high accuracy, but with increasing filter order, its computational speed decreased significantly, which would make it impossible to use it in real-time applications. Therefore, an alternative was sought that could extract fECG with the same (or higher) accuracy, but much faster, even with higher filter orders. The FTF algorithm met these requirements and was selected and tested for these reasons. The FTF algorithm requires two input signals: the aECG (mixture of mECG, fECG and other interferences) and the reference mECG signal, which is modified by the FTF method so that it corresponds as closely as possible to the maternal component in the aECG signal and can be subtracted from it, thus obtaining fECG and some residual noise. In general, there are two approaches to obtain these two inputs (mECG and aECG): •The first involves simultaneous scanning of mECG from the mother’s thorax and aECG from the abdominal area. However, the thoracic mECG signal differs from the maternal component sensed on the abdominal wall in terms of its magnitude, phase, and morphology. Moreover, in clinical practice it is quite challenging to ensure a high quality thoracic mECG signal, i.e. without artifacts that are caused by maternal movement or improper electrode placement and fixation. The quality of the extraction by an adaptive algorithm is closely connected to the quality of the input signals, and therefore distortion the maternal reference signal leads to inaccurate fECG extraction. Additionally, its measurement can be unnecessarily uncomfortable and stressful for the mother [9]. •For these reasons, we chose the second approach, which is based on obtaining the mECG and aECG input signals solely from aECG signals. This involves implementation of a separation algorithm (such as ICA), which is able to estimate mECG and aECG with an enhanced fetal component. 2. Post-processing – for post-processing, it was necessary to select a method that would be able to remove mECG residues so that the fHR determination is accurate and at the same time would not affect the morphology of the resulting fECG waveform, so that the ST signal analysis could be performed. Past research [22] has shown that WT-based methods are not suitable for further morphological analysis, and therefore EMD-based algorithms have been considered. Among them, CEEMDAN, which was used in the hybrid system for the final smoothing of fECG, proved to be the most suitable in terms of the efficacy and computational complexity. The advantages of individual methods and their combination to achieve accurate and time-efficient fECG extraction can be summarized as follows: 5 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 •ICA-FastICA is a computationally efficient method with low memory requirements [89] but is not able to sufficiently suppress mECG and extract fECG in sufficient quality when used on its own. On the other hand, using solely aECG input signals, it can provide a good estimate of the mECG and aECG signal with enhanced fetal component that are used as inputs for FTF algorithm. •FTF-adaptive filters generally excel in performance and adaptability in time-varying environments (which is also the case with fECG) [53]. In case that the filter receives highquality input signals (aECG and mECG), it is able to extract an accurate fECG signal only with minor mECG residue. Moreover, it performs in a time-efficient manner even with increasing filter order. •CEEMDAN method proved to be the best compromise between performance and computational complexity among other EMD-based algorithms. Despite the fact that it cannot extract fECG or mECG on its own, it has proven to be suitable for use in the post-processing phase for the additional suppression of mECG residues. In addition, it did not significantly alter the morphology of the fECG signal, allowing accurate morphological analysis to be performed. 3.4. Dataset In this work, the hybrid algorithm was tested on two databases. The first one is called the Fetal Electrocardiograms, Direct and Abdominal with Reference Heartbeats Annotations (FECGDARHA) and is freely available in figshare repository [90]. The dataset contains 12 five-minute recordings obtained from women between the 38th and 42nd week of gestation. The recordings were obtained at an advanced stage of child development. Each recording contains four abdominal signals recorded from the mother’s abdominal area using standard Ag/AgCl electrodes. At the same time, the direct fECG signal was recorded from the fetal head using a sterile spiral electrode. The signals were recorded using the KOMPOREL system and digitized with 16-bit resolution and a sampling frequency of 1000 Hz for the direct fECG and 500 Hz for the abdominal signals. In addition, this system automatically detected and marked the positions of fetal R-peaks, the accuracy of which was verified by clinical experts. The dataset, therefore, also includes annotations providing the exact positions of fetal R-peaks. Data collection was performed at the Department of Obstetrics and Gynaecology of the Medical University of Silesia in Katowice, Poland [90]. The second database is called the Physionet Challenge 2013 [91]. The database was created as part of the PhysioNet Computing in Cardiology Challenge 2013, which aimed to support the development of algorithms for accurate localization of fQRS complexes and fQT interval estimation in non-invasive fECG signals. Each recording contains four abdominal signals one minute long and with a sampling frequency of 1000 Hz. At the same time, direct fECG signals were also captured using a fetal scalp electrode, which served to create reference annotations indicating the position of fQRS complexes. Nevertheless, only annotations are part of the database; direct fECG signals are not part of the database. The data were obtained from different women with a different gestational age of the fetus. Different devices with different resolutions and configurations were used to record the signal [91]. 3.5. Hybrid system design This section describes the ICA-FTF-CEEMDAN hybrid system for the extraction of non-invasive fECG, including the parameters that were used to evaluate the quality of the extraction. The hybrid system can be described using the following steps: •Selection of suitable aECG signals: first, it was necessary to select at least two input signals, as this is a multi-channel algorithm. Four input signals were available for each recording, but not all of them were captured well (e.g. the fetal component acquired very low magnitudes). Therefore, only the most suitable signals were selected so as not to unnecessarily reduce the extraction effectiveness. •Preprocessing: input aECG signals were preprocessed using a bandpass finite impulse response filter (BPF FIR) with a frequency range of 3–150 Hz (fQRS complexes occur in the range of 10–15 Hz [20]) and a filter order of 500 Hz. The filter was used to remove isoline fluctuations and motion artifacts, see Fig. 1. •FastICA: Furthermore, the ICA method, which is based on the decomposition of a mixture of signals into the original source components, was applied. In this case, a very effective FastICA algorithm was used; it was arbitrarily set to 20 iterations and decomposition of the input aECG signals into three components. The first component corresponded to aECG with highlighted fECG (referred to as aECGICA), the second mECG (referred to as mECGICA) and the third one corresponded to noise, see Fig. 2. As the polarity of the ICA components was rotated during the extraction, an algorithm checking the highest positive and negative amplitude within 50 ms from the detected R-peaks was used, and, based on this, the algorithm determined whether the signal was rotated correctly or whether it needed to be rotated. •FTF: the first two components of aECGICA and mECGICA were time and amplitude aligned, using automatic centering. The centering was based on detecting the positions and amplitudes of maternal R-peaks. The shifts between the individual maternal R-peaks in aECGICA and the maternal R-peaks in mECGICA were determined and the average time shift was calculated. The signals were aligned by this average time shift. Similarly, differences between amplitudes were determined. The signal with the higher average amplitude of the maternal R-peaks was proportionally reduced relative to the signal with lower average amplitude. Such centered signals were used as inputs of the adaptive FTF algorithm. The aECGICA component was used as the primary input (desired signal). And, the mECGICA component, which the FTF algorithm filtered into the approximate form of the mECG component contained in the primary input, was used as the second input. Subsequently, the mECGFTF component extracted by the FTF filter was subtracted from the aECGICA signal, thus obtaining fECGFTF with a suppressed maternal component, see Fig. 3. As for the FTF algorithm, forgetting factor λwas set to a value of 1 due to optimization, and filter order Mwas tested in the range of 1 to 100 with increments of 1. •CEEMDAN: since the fECGFTF signal contained residual noise, the CEEMDAN method was used to remove it. It decomposed the input fECGFTF signal to simpler signals, so-called IMFs. The decomposition into IMFs continued until the stopping condition was reached (the residual signal was a monotone function, a function with only one extremum or a constant). A total of 10 IMFs were extracted, and a final signal indicated as fECGCEEMDAN was generated by selecting the appropriate IMFs, see Fig. 4. For this method, it was necessary to set the value of the standard deviation of the added noise series Nstd and the number of ensemble trials N. Parameter Nstd was tested in the range of 0.1 to 0.9 in 0.1 increments. Parameter Nwas tested for values of 10, 30 and 60; testing for higher values was not necessary, as the extraction effectiveness was no longer improved. 6 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 Fig. 1. Selection of suitable aECG signals and their preprocessing. Example of (a) all input aECG signals for recording r01, (b) selected aECG signals that are suitable for further processing, (c) filtering of suitable aECG signals using a BPF FIR filter in the range of 3–150 Hz. Fig. 2. Signal processing using the FastICA algorithm. Example (a) presents preprocessed input aECG signals, example (b) shows the extraction of three ICA components and example (c) represents the assignment of the ICA components to the source signals and adjusting their polarity. Fig. 3. Signal processing using the FTF algorithm. Example (a) presents aECGICA, which was used as the desired signal, indicated as d(n), and mECGICA, which needed to be filtered by the FTF filter, indicated as x(n). Example (b) shows the mECGFTF component that was extracted by the FTF filter, indicated as y(n). The extracted mECGFTF was subtracted from aECGICA, thus creating the fECGFTF error signal, indicated as e(n). Example (c) represents the fECGFTF output signal. •Algorithms settings: to achieve the most efficient extraction of fECG, optimal setting of the parameters of the individual algorithms was crucial. Selection of suitable input aECG signals, parameter Mand forgetting factor λsetting for the FTF filter, N,Nstd parameters of the CEEMDAN method and selection of suitable IMFs by means of which the resulting fECGCEEMDAN was generated. This was achieved using an automated algorithm that compared various combinations of the parameters set and selected the combination that provided fECG signal with the highest accuracy according to the ACC value. Accuracy comparison was carried out using annotations. The most suitable parameters values for the individual recordings are summarized for the FECGDARHA database and the Challenge database in Tables 3 and 4, respectively. •Evaluation parameters: the effectiveness of the hybrid method was evaluated, both in terms of the accuracy of fetal R-peaks detection and in terms of the accuracy of fHR determination. First, fetal R-peaks were detected using a detector based on continuous wavelet transform (CWT). The positions of the detected R-peak positions were compared with the reference R-peak positions given in the annotations. If the position of the detected R-peak was ±50 ms from the reference position, it was marked as true positive 7 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 Fig. 4. Signal processing using the CEEMDAN algorithm. Example (a) presents the input fECG signal processed by the FTF algorithm, example (b) shows the first 5 extracted IMFs, example (c) shows the resulting extracted fECGCEEMDAN signal using a hybrid algorithm compared to the reference recording which was made using a scalp electrode. Table 3 ICA-FTF-CEEMDAN algorithm settings for recordings from the FECGDARHA database. Recordings Combination ICA FTF CEEMDAN of electrodes Number of MλN Nstd IMFs iterations r01 1, 3, 4 20 2 1 10 0.2 2+3 r02 1, 2, 3, 4 20 99 1 10 0.6 3+4 r03 2, 4 20 90 1 60 0.7 3 r04 1, 4 20 38 1 30 0.9 2+5 r05 1, 4 20 21 1 10 0.5 2+3+7 r06 1, 2, 3, 4 20 16 1 30 0.5 3+4 r07 1, 4 20 36 1 30 0.9 2+5 r08 1, 4 20 2 1 60 0.4 4 r09 1, 2, 4 20 16 1 10 0.6 2+3+4+6 r10 1, 2, 3, 4 20 54 1 60 0.4 2+4+6 r11 1, 2, 3, 4 20 100 1 60 0.7 3 r12 1, 2, 3, 4 20 16 1 30 0.3 2+3+6+7 Table 4 ICA-FTF-CEEMDAN algorithm settings for recordings from the Challenge database. Recordings Combination ICA FTF CEEMDAN of electrodes Number of MλN Nstd IMFs iterations a01 1, 3, 4 20 11 1 10 0.8 2+3+5+7 a02 1, 2, 4 20 2 1 60 0.5 4+6+8 a03 1, 4 20 68 1 10 0.5 3+6+8 a04 1, 2 20 30 1 10 0.5 2+3 a05 1,3 20 30 1 10 0.5 2+3 a06 2, 4 20 25 1 60 0.4 3+6+7 a07 1, 2, 3, 4 20 4 1 10 0.4 3+5+6+8 a08 1, 4 20 2 1 10 0.3 2 a09 1, 4 20 99 1 30 0.6 3 a10 2, 4 20 84 1 10 0.9 2 a11 1, 4 20 68 1 30 0.3 2+3+5 a12 1,3, 4 20 9 1 30 0.6 3+6+7 a13 2, 4 20 35 1 60 0.9 4+7 a14 1, 2, 3, 4 20 10 1 10 0.7 3+4 a15 1, 4 20 49 1 10 0.3 2 a16 1, 4 20 68 1 10 0.8 4+5 a17 1, 4 20 2 1 10 0.3 2 a18 1, 2, 3, 4 20 33 1 10 0.1 5+9 a19 3, 4 20 35 1 60 0.6 4 a20 1, 4 20 99 1 60 0.9 4+5 a21 2, 3, 4 20 87 1 30 0.9 2+3+4+5+7 a22 1, 4 20 2 1 10 0.2 2 a23 1, 3 20 47 1 30 0.9 2+4+5 a24 1, 3 20 52 1 60 0.9 3+4 a25 2, 3 20 90 1 60 0.9 4 (TP), i.e. as a correctly detected (true) R-peak. Detected Rpeaks in the extracted signal, which fell outside this interval, were marked as false positive (FP), so it corresponded to the detection of a peak that was not a fetal one (for example, the maternal residue or other interference). Omitted fetal R-peaks that were in the original but were not detected by the detector were marked as false negative (FN) values (for example, the amplitude of fetal R-peaks could be suppressed by the filtration) [39,92,93]. ACC, SE, PPV and F1 indices could be determined using FN, FP, and TP values. The ACC parameter expresses how many R-peaks were detected correctly with respect to all detected and omitted R-peaks. Thus, it defines the ratio of all TP values with respect to all marked values (TP, FP, FN), see Eq. (6). The SE parameter shows how many of all existing R-peaks were correctly predicted. It is defined as the ratio of all TP values to all existing R-peaks (TP and FN), see Eq. (7). The PPV parameter tells how many of the R-peaks that were detected were true. Defined as the ratio of all TP values to all detected Rpeaks (TP and FP), see Eq. (8). And parameter F1 takes into account both parameters SE and PPV and is defined as their harmonic average, see Eq. (9). ACC =TP TP +FP +FN ·100 (%).(6) SE =TP TP +FN ·100 (%).(7) PPV =TP TP +FP ·100 (%).(8) F1=2·SE ·PPV SE +PPV =2·TP 2·TP +FP +FN ·100 (%).(9) To evaluate the fHR, it was necessary to convert the vectors of intervals between the detected positions of the R-peaks into vectors of the current fHR values. The accuracy of the fHR determination was evaluated using Bland–Altman plots, where vector of differences and vector of the average of the two signals were determined. When graphically represented, the difference is shown on the vertical axis and the average of the values is found on the horizontal axis. The middle horizontal line indicates the mean µof all differences and, based on this line, a 95% confidence interval, which lies in the interval µ±1.96σ, is then plotted [94]. In addition, fHR traces were plotted for the comparison of fHR. To plot the fHR traces, it was necessary to create moving averages from the fHR values calculated. A width of 30 samples was the most suitable window size. A schematic representation of the procedure for evaluating the extracted signals is shown in Fig. 5. 8 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 Fig. 5. Procedure for evaluation of the extracted fECGCEEMDAN signals. Example (a) shows the extracted fECGCEEMDAN signal using a hybrid algorithm in comparison with the reference recording. Example (b) demonstrates the detection of fetal R-peaks and the determination of the TP, FP, and FN values. Example (c) represents the fHR trace plotting. •ST segment analysis: This method is based on monitoring the T/QRS ratio. The change in this ratio is due to a change in cellular ionic currents during anaerobic metabolism of the fetal heart. It has been shown that the combined monitoring of fHR and ST analysis leads to the early identification of cases of acidosis during childbirth, and at the same time it has been associated with reduced number of unnecessary caesarian sections [17,20]. The current disadvantage of today’s ST analyzers is its invasiveness and the use being possible only during childbirth. Therefore, it is highly desirable to find a solution to perform a non-invasive ST analysis with the same accuracy as can be done invasively. In order to determine whether the determination of the ST analysis on the estimated fECG signal was accurate, it is necessary to have a reference signal measured with a scalp electrode. This is only met by the FECGDARHA database, so ST analysis was performed on 12 records from this database. Fig. 6 illustrates the ST analysis procedure which can be summarized in the following steps: 1. Determination of the R-peak positions in the estimated fECG signal using a CWT detector. Annotations with R oscillation positions are available for the reference signal and their detection is therefore not necessary. 2. Averaging of 30 consecutive complexes for both the reference and estimated signal. This approach was inspired by the ST analyzers used in clinical practice. 3. Determination of positions and amplitudes of R peak, S peak, and T wave in averaged complexes. These positions were also detected by CWT detector but with different settings. 4. Determination of ST analysis for individual complexes by calculating the T/QRS ratio, i.e. the ratio of the amplitude of the T wave to the sum of the amplitude of the R and S waves. 5. Plotting the determined values into the graph in the form of marks ‘*’, as displayed by ST in the clinical practice. 4. Results This section summarizes the results obtained when testing the hybrid method on real recordings. The statistical results obtained according to the ACC, SE, PPV and F1 indices are presented for the FECGDARHA database in Table 5. Indices values greater than 95% are highlighted in green (6 recordings evaluated according to ACC, 9 recordings evaluated according to SE, 11 recordings evaluated according to PPV, and 10 recordings evaluated according to F1). The values of all quality indices were higher than 80% with all recordings except r11 (indices values lower than 80% are highlighted in red). As for recordings r01, r05, r08 and r09, all fetal R-peaks were detected correctly; no FP or FN fetal R-peaks were detected, and all quality indices achieved a value of 100%. The table also shows the error mean values µand values of ±1.96σ, reflecting the differences between the reference fHR and the fHR estimated by the hybrid method. The results can be interpreted as follows: the hybrid method was effective for all recordings (highlighted in green) except r11, since, in this case, high mean values µand values of ±1.96σwere obtained (highlighted in red). The statistical evaluation for the Challenge database is summarized in Table 6. Indices values greater than 95% are highlighted in green (11 recordings evaluated according to ACC, 14 recordings evaluated according to SE and F1, and 15 recordings evaluated according to PPV). The ACC, SE and F1 parameters exceeded the value of 80% for 17 recordings, and PPV was higher than 80% for 20 recordings (indices values lower than 80% are highlighted in red). All fetal R-peaks were detected correctly in recordings a04, a05, a08, a15, a17, a22 and a25; in addition, no FP and FN values were detected, and values of 100% were achieved for all quality indices. When evaluating mean values µand values of ±1.96σ, the method was effective for 15 and 13 recordings, respectively (highlighted in green). High mean values µand values of ±1.96σ were acquired for 10 and 12 recordings, respectively (highlighted in red). The obtained results of parameters µand ±1.96σcan be visualized using Bland–Altman plots. These graphs carry the same information as mentioned in the previous paragraph, but in a graphical form. The example of Bland–Altman plots, (a) for recording r03 and (b) for recording r09 in Fig. 7, presents recordings for which highly accurate results were achieved in determining fHR. Fig. 8 shows (a) recording r04 and (b) recording r12, for which a slight deviation of fHR values from the reference values was achieved. Fig. 9 represents (a) recording a09 and (b) recording a16, for which the hybrid method was not effective because the estimated fHR values excessively deviated from the reference values. A comparison of fHR traces of all recordings from the FECGDARHA database is shown in Fig. 10(a). All recordings except recording r11 copy the reference trend, and the method can, therefore, be considered effective for these recordings. As for the 9 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 Table 8 Comparison of the results with other studies. ACC, SE, PPV, and F1 values shown are average values calculated for all recordings tested. Author, Algorithm Dataset ACC SE PPV F1 Advantages source (%) (%) (%) (%) and limitations Lee et al. [95] STVD-TSPCA Challenge 2013 FECGSYNDB – – 90.50 – 89.40 – 89.90 – +low computational load −distortion of fECG morphology −poor performance for low SNR levels Gurve et al. [97] CS-NMF ADFECGDB 95.30 94.60 94.80 84.00 +single channel method Challenge 2013 – – – – −poor performance for high noise recordings Sameni et al. [98]πCA-GEVD DaISy – – – – +time-efficient method −not tested on recordings with higher noise level Castillo et al. [100] WT-CT ADFECGDB Challenge 2013 97.30 97.08 98.40 97.93 98.86 99.11 98.63 98.52 +allows use in real time applications +single channel method −lower quality recordings were not used for testing Ayat et al. [101] SGSF-PN NIFECGDB – – – – +single channel method −not tested on recordings with higher noise level Zhang et al. [102] Clustering-PCA ADFECGDB – 96.60 95.60 96.09 +single channel method Challenge 2013 – 95.85 95.10 95.47 +tested on recordings with abnormalities FECGSYNDB – – – – −poor performance for high noise recordings Da Poian et al. [103] CS-ICA ADFECGDB Challenge 2013 – – 92.50 78.00 92.00 77.00 92.20 77.50 +allows use in real time applications +suitable for low-power monitoring devices −not effective for poor quality aECG signals Panigrahy et al. [104] EKS-DE-ANFIS Challenge 2013 84.89 91.47 92.18 91.82 +single channel method NIFECGDB 90.66 94.21 96.05 95.12 −low performance at lower sampling frequency Liu et al. [34] RRTSS-TM Challenge 2013 – – – 95.00 +single channel method −poor performance for high noise recordings −phase and amplitude of fQRS were distorted Previously presented ICA-RLS-WT FECGDARHA 85.92 89.70 92.41 90.99 +lower computational load due to the use of WT algorithm [22] Challenge 2013 68.25 72.60 81.31 75.68 −distorts the morphology of fQRS Previously presented ICA-RLS-EMD FECGDARHA 84.73 87.99 92.72 90.10 +lower computational load due to the use of EMD algorithm [46] Challenge 2013 64.95 69.34 79.62 72.74 −poor performance for recordings with low fECG magnitude Proposed ICA-FTF-CEEMDAN FECGDARHA 92.98 95.33 96.49 95.86 +does not distort the morphology of fQRS algorithm Challenge 2013 78.47 82.06 87.90 84.62 −the need for optimal parameter settings signals, electrical signals produced by uterus during contractions could be extracted and analyzed from the recordings. The computational complexity of individual ICA and FTF methods was classified as low and the CEEMDAN method as a medium (see, Table 2), and thus the implementation within real-time devices appears to be very promising. However, the disadvantage of the algorithm was its dependence of high performance on optimal setting of parameters. Parameter optimization leads to an increase in the computational time of the algorithm. Therefore, if necessary to reduce the computational time, it is possible to reduce the time when optimizing the parameters as follows: (1) The algorithm setting could be optimized for a shorter signal section (such with 15, 30 or 60 s) and extraction would be performed with this setting for the rest of the signal; (2) The optimization could be also performed at shorter but more frequent intervals (for example, the first 10 s of the signal would be used for optimization, the next 60 s of the signal would be extracted with this setting followed by another 10 s for optimization, etc.). The algorithm could also find application in other areas where it seems promising to use adaptive algorithms. These include the processing of other human biological signals, such as adult ECG [109,110], speech signals [111,112] or signals used in telecommunications [113]. Future research will focus on verifying the effectiveness of the method in the above-mentioned real prototype device and its effectiveness will be tested on a larger number of records, including pathological records. Using the prototype, a dataset will be created with real fECG records, both physiological and pathological, including different gestational gestation ages and fetal positions. In addition to non-invasive aECG recordings, a simultaneously recorded CTG reference or direct fECG signal using a scalp electrode so that we can create annotations. The dataset can then be used by researchers to evaluate the accuracy of filtration techniques. In the future, it is also possible to perform other types of morphological analyzes (such as QT interval analysis), which provide physicians with valuable information about the health of the fetus. Further research should also focus on the classification of fECG signals into categories, such as physiological and pathological, and on the detection of arrhythmias, as in the case of adult ECG [114,115]. Here, it would be appropriate to 16 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 test algorithms based on artificial intelligence and machine learning (e.g. ANNs, k-means, k-medoids, support vector machines or fuzzy logic) within the classification tasks. 7. Conclusion The results of the study showed that the ICA-FTF-CEEMDAN hybrid method was able to efficiently extract fECG from abdominal recordings. Also, it achieved more accurate statistical results in the detection of fQRS complexes than the previously presented approaches. Compared to the RLS algorithm, the detection of fQRS complexes by FTF did not improve significantly. However, with increasing filter order, the RLS filter required linearly more computation time, while the FTF filter was able to perform filtering quickly regardless of the filter order (the computation time was approximately 2 s for filter order values ranging from 1 to 100). When comparing the EEMD and CEEMDAN methods, the CEEMDAN method achieved slightly better suppression of the maternal component. Further, the shape of the fECG waveform was not as deformed as when using EEMD. Preservation of the signal morphology has a fundamental influence on the subsequent morphological analysis of the fECG signal. As for both methods, there was a linear increase in computational time for both methods depending on the increasing number of ensemble trials in both methods, but the total computational time for the CEEMDAN method was approximately 40% lower. The proposed algorithm thus proved to be effective both in terms of accuracy of fHR determination and in terms of computational speed, i.e. enabling its implementation within real-time operating devices. Moreover, no additional lead is needed to record the reference mECG from the mother’s chest, leading to greater maternal comfort and mobility. The main benefit of this approach is that it can obtain the signal of a sufficient quality to perform an accurate ST analysis. This confirms the assumption that non-invasive fECG monitoring can achieve as accurate results as when monitored invasively. Thus, non-invasive fECG appears to be a promising alternative to CTG and invasive fECG, which are used in the today’s clinical practice. To verify the effectiveness of this combined algorithm, it would be necessary to have a database with a larger number of signals. Unfortunately, there are still relatively a few databases with real recordings and annotations. Therefore, future research will focus on the design of a monitoring device for fECG, where this algorithm will be tested on our own data. Furthermore, the algorithm will also be tested on pathological recordings, and the research will also focus on the classification of fECG signals into categories, such as physiological and pathological and other types of morphological analysis (such as QT interval analysis). CRediT authorship contribution statement Katerina Barnova: Methodology, Investigation, Validation, Writing – original draft. Radek Martinek: Conceptualization, Supervision, Funding acquisition. Rene Jaros: Software, Visualization, Data curation. Radana Kahankova: Investigation, Validation, Formal analysis, Editing. Khosrow Behbehani: Writing – review & editing. Vaclav Snasel: Project administration, Supervision, Validation. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgment This article was supported by the Ministry of Education of the Czech Republic (Projects No. SP2021/32, SP2021/24). This article was also supported by the European Regional Development Fund in the Research Centre of Advanced Mechatronic Systems project, project number CZ.02.1.01/0.0/0.0/16_019/0000867 within the Operational Programme Research, Development and Education. References [1] D. Hutter, J. Kingdom, E. Jaeggi, Causes and mechanisms of intrauterine hypoxia and its impact on the fetal cardiovascular system: a review, Int. J. Pediatrics 2010 (2010) 1–9, http://dx.doi.org/10.1155/2010/401323. [2] L.J. Millar, L. Shi, A. Hoerder-Suabedissen, Z. Molnár, Neonatal hypoxia ischaemia: mechanisms, models, and therapeutic challenges, Front. Cell. Neurosci. 11 (2017) 78, http://dx.doi.org/10.3389/fncel.2017.00078. [3] J. Samuel, C. Franklin, Hypoxemia and hypoxia, in: J.A. Myers, K.W. Millikan, T.J. Saclarides (Eds.), Common Surgical Diseases, Springer New York, New York, NY, 2008, pp. 391–394, http://dx.doi.org/10.1007/978-0387-75246-4_97. [4] S. Vannuccini, C. Bocchi, F.M. Severi, F. Petraglia, Diagnosis of fetal distress, in: G. Buonocore, R. Bracci, M. Weindling (Eds.), Neonatology, Springer International Publishing, Cham, 2016, pp. 1–23, http://dx.doi.org/ 10.1007/978-3-319-18159-2_156-1. [5] M. Herrera-Marschitz, T. Neira-Pena, E. Rojas-Mancilla, P. EspinaMarchant, D. Esmar, R. Perez, V. Muãoz, M. Gutierrez-Hernandez, B. Rivera, N. Simola, D. Bustamante, P. Morales, P.J. Gebicke-Haerter, Perinatal asphyxia: CNS development and deficits with delayed onset, Front. Neurosci. 8 (2014) http://dx.doi.org/10.3389/fnins.2014.00047. [6] W. Künzel (Ed.), Fetal Heart Rate Monitoring: Clinical Practice and Pathophysiology, Springer Berlin Heidelberg, Berlin, Heidelberg, 1985, http://dx.doi.org/10.1007/978-3-642-70358-4. [7] C. Smyth, J. Farrow, Present place in obstetrics for foetal phonocardiography and electrocardiography, BMJ (Clin. Res. Ed.) 2 (5103) (1958) 1005–1009, http://dx.doi.org/10.1136/bmj.2.5103.1005. [8] C.B. Martin, Electronic fetal monitoring: A brief summary of its development, problems and prospects, Eur. J. Obstet. Gynecol. Reprod. Biol. 78 (2) (1998) 133–140, http://dx.doi.org/10.1016/S0301-2115(98)00059-1. [9] R. Kahankova, R. Martinek, R. Jaros, K. Behbehani, A. Matonia, M. Jezewski, J.A. Behar, A review of signal processing techniques for non-invasive fetal electrocardiography, IEEE Rev. Biomed. Eng. 13 (2020) 51–73, http: //dx.doi.org/10.1109/RBME.2019.2938061. [10] A. Pinas, E. Chandraharan, Continuous cardiotocography during labour: analysis, classification and management, Best Pract. Res. Clin. Obstet. Gynaecol. 30 (2016) 33–47, http://dx.doi.org/10.1016/j.bpobgyn.2015.03. 022. [11] T. Khangura, E. Chandraharan, Electronic fetal heart rate monitoring: the future, Curr. Women’s Health Rev. 9 (3) (2014) 169–174, http://dx.doi. org/10.2174/157340480903140131111807. [12] S. Vannuccini, C. Bocchi, F.M. Severi, F. Petraglia, Diagnosis of fetal distress, in: G. Buonocore, R. Bracci, M. Weindling (Eds.), Neonatology, Springer International Publishing, Cham, 2018, pp. 105–127, http://dx. doi.org/10.1007/978-3-319-29489-6_156. [13] R.L. Williams, W.E. Hawes, Cesarean section, fetal monitoring, and perinatal mortality in california., Am J Public Health 69 (9) (1979) 864–870, http://dx.doi.org/10.2105/AJPH.69.9.864. [14] H.-Y. Chen, S.P. Chauhan, C.V. Ananth, A.M. Vintzileos, A.Z. Abuhamad, Electronic fetal heart rate monitoring and its relationship to neonatal and infant mortality in the United States, Am. J. Obstet. Gynecol. 204 (6) (2011) 491.e1–491.e10, http://dx.doi.org/10.1016/j.ajog.2011.04.024. [15] A.P. Betrán, J. Ye, A.-B. Moller, J. Zhang, A.M. Gülmezoglu, M.R. Torloni, The increasing trend in caesarean section rates: global, regional and national estimates: 1990–2014, PLoS One 11 (2) (2016) e0148343, http: //dx.doi.org/10.1371/journal.pone.0148343. [16] E. Blix, K.G. Brurberg, E. Reierth, L.M. Reinar, P. Øian, ST waveform analysis versus cardiotocography alone for intrapartum fetal monitoring: A systematic review and meta-analysis of randomized trials, Acta Obstet. Gynecol. Scand. 95 (1) (2016) 16–27, http://dx.doi.org/10.1111/aogs. 12828. [17] T. Kazmi, F. Radfer, S. Khan, ST Analysis of the fetal ECG, as an adjunct to fetal heart rate monitoring in labour: a review, Oman Med. J. 26 (6) (2011) 459–460, http://dx.doi.org/10.5001/omj.2011.118. [18] E. Karvounis, M. Tsipouras, C. Papaloukas, D. Tsalikakis, K. Naka, D. Fotiadis, A non-invasive methodology for fetal monitoring during pregnancy, Methods Inf. Med. 49 (03) (2010) 238–253, http://dx.doi.org/10.3414/ ME09-01-0041. 17 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 [19] D. Jagannath, A.I. Selvakumar, Issues and research on foetal electrocardiogram signal elicitation, Biomed. Signal Process. Control 10 (2014) 224–244, http://dx.doi.org/10.1016/j.bspc.2013.11.001. [20] Sameni, A review of fetal ECG signal processing issues and promising directions, the open pacing, Electrophysiol. Ther. J. (2010) http://dx.doi. org/10.2174/1876536X01003010004. [21] G.D. Clifford, I. Silva, J. Behar, G.B. Moody, Non-invasive fetal ECG analysis, Physiol. Meas. 35 (8) (2014) 1521–1536, http://dx.doi.org/10.1088/09673334/35/8/1521. [22] R. Jaros, R. Martinek, R. Kahankova, J. Koziorek, Novel hybrid extraction systems for fetal heart rate variability monitoring based on non-invasive fetal electrocardiogram, IEEE Access 7 (2019) 131758–131784, http://dx. doi.org/10.1109/ACCESS.2019.2933717. [23] D.A. Ramli, Y.H. Shiong, N. Hassan, Blind source separation (BSS) of mixed maternal and fetal electrocardiogram (ECG) signal: A comparative study, Procedia Comput. Sci. 176 (2020) 582–591, http://dx.doi.org/10.1016/j. procs.2020.08.060. [24] A. Ghazdali, A. Hakim, A. Laghrib, N. Mamouni, S. Raghay, A new method for the extraction of fetal ECG from the dependent abdominal signals using blind source separation and adaptive noise cancellation techniques, Theor. Biol. Med. Model. 12 (1) (2015) 25, http://dx.doi.org/10.1186/ s12976-015-0021-2. [25] P.J. He, X.M. Chen, Y. Liang, H.Z. Zeng, Extraction for fetal ECG using single channel blind source separation algorithm based on multi-algorithm fusion, in: W.-P. Sung, J. Kao (Eds.), MATEC Web Conf. 44 (2016) 01026, http://dx.doi.org/10.1051/matecconf/20164401026. [26] R. Martinek, R. Kahankova, J. Jezewski, R. Jaros, J. Mohylova, M. Fajkus, J. Nedoma, P. Janku, H. Nazeran, Comparative effectiveness of ICA and PCA in extraction of fetal ECG from abdominal signals: toward non-invasive fetal monitoring, Front. Physiol. 9 (2018) 648, http://dx.doi.org/10.3389/ fphys.2018.00648. [27] P. Manorost, N. Theera-Umpon, S. Auephanwiriyakul, Fetal electrocardiogram extraction by independent component analysis, in: 2017 7th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), IEEE, Penang, 2017, pp. 220–225, http://dx.doi.org/10.1109/ ICCSCE.2017.8284408. [28] M. Kotas, J. Giraldo, S.H. Contreras-Ortiz, G.I.B. Lasprilla, Fetal ECG extraction using independent component analysis by jade approach, in: J. Brieva, J.D. García, N. Lepore, E. Romero (Eds.), 13th International Conference on Medical Information Processing and Analysis, SPIE, San Andres Island, Colombia, 2017, p. 55, http://dx.doi.org/10.1117/12.2285962. [29] R. Petrolis, V. Gintautas, A. Krisciukaitis, Multistage principal component analysis based method for abdominal ECG decomposition, Physiol. Meas. 36 (2) (2015) 329–340, http://dx.doi.org/10.1088/0967-3334/36/2/329. [30] H. Hassanpour, A. Parsaei, Fetal ECG extraction using wavelet transform, in: 2006 International Conference on Computational Inteligence for Modelling Control and Automation and International Conference on Intelligent Agents Web Technologies and International Commerce (CIMCA’06), IEEE, Sydney, NSW, 2006, p. 179, http://dx.doi.org/10.1109/CIMCA.2006.98. [31] Y. Alshebly, M. Nafea, Isolation of fetal ECG signals from abdominal ECG using wavelet analysis, IRBM 41 (5) (2020) 252–260, http://dx.doi.org/10. 1016/j.irbm.2019.12.002. [32] S. Wu, Y. Shen, Z. Zhou, L. Lin, Y. Zeng, X. Gao, Research of fetal ECG extraction using wavelet analysis and adaptive filtering, Comput. Biol. Med. 43 (10) (2013) 1622–1627, http://dx.doi.org/10.1016/j.compbiomed. 2013.07.028. [33] A. Matonia, J. Jezewski, K. Horoba, A. Gacek, P. Labaj, The maternal ECG suppression algorithm for efficient extraction of the fetal ECG from abdominal signal, in: 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, IEEE, New York, NY, 2006, pp. 3106–3109, http://dx.doi.org/10.1109/IEMBS.2006.260221. [34] H. Liu, D. Chen, G. Sun, Detection of fetal ECG r wave from singlelead abdominal ecg using a combination of RR time-series smoothing and template-matching approach, IEEE Access 7 (2019) 66633–66643, http://dx.doi.org/10.1109/ACCESS.2019.2917826. [35] G. Camps, M. Martinez, E. Soria, Fetal ECG extraction using an FIR neural network, in: Computers in Cardiology 2001. Vol.28 (Cat. No.01CH37287), IEEE, Rotterdam, Netherlands, 2001, pp. 249–252, http://dx.doi.org/10. 1109/CIC.2001.977639. [36] G. Camps-Valls, M. Martínez-Sober, E. Soria-Olivas, R. MagdalenaBenedito, J. Calpe-Maravilla, J. Guerrero-Martínez, Foetal ECG recovery using dynamic neural networks, Artif. Intell. Med. 31 (3) (2004) 197–209, http://dx.doi.org/10.1016/j.artmed.2004.03.005. [37] M. Bin Ibne Reaz, L.S. Wei, Adaptive linear neural network filter for fetal ECG extraction, in: International Conference on Intelligent Sensing and Information Processing, 2004. Proceedings of, IEEE, Chennai, India, 2004, pp. 321–324, http://dx.doi.org/10.1109/ICISIP.2004.1287675. [38] R. Kahankova, R. Martinek, M. Mikolasova, R. Jaros, Adaptive linear neuron for fetal electrocardiogram extraction, in: 2018 IEEE 20th International Conference on E-Health Networking, Applications and Services (Healthcom), IEEE, Ostrava, 2018, pp. 1–5, http://dx.doi.org/10.1109/HealthCom. 2018.8531135. [39] L. Billeci, M. Varanini, A combined independent source separation and quality index optimization method for fetal ECG extraction from abdominal maternal leads, Sensors 17 (5) (2017) 1135, http://dx.doi.org/10.3390/ s17051135. [40] G. Liu, Y. Luan, An adaptive integrated algorithm for noninvasive fetal ECG separation and noise reduction based on ICA-EEMD-WS, Med. Biol. Eng. Comput. 53 (11) (2015) 1113–1127, http://dx.doi.org/10.1007/s11517015-1389-1. [41] A. Gupta, M. Srivastava, V. Khandelwal, A. Gupta, A novel approach to fetal ECG extraction and enhancement using blind source separation (BSSICA) and adaptive fetal ECG enhancer (AFE), in: 2007 6th International Conference on Information, Communications & Signal Processing, IEEE, Singapore, 2007, pp. 1–4, http://dx.doi.org/10.1109/ICICS.2007.4449716. [42] R. Jaros, R. Martinek, K. Barnova, M. Ladrova, Use of a hybrid method ICAPCA-ICA for fetal electrocardiography extraction, in: 2019 International Symposium on Advanced Electrical and Communication Technologies (ISAECT), IEEE, Rome, Italy, 2019, pp. 1–6, http://dx.doi.org/10.1109/ ISAECT47714.2019.9069682. [43] C. Li, B. Fang, H. Li, P. Wang, A novel method of FECG extraction combined self-correlation analysis with ICA, in: 2016 8th IEEE International Conference on Communication Software and Networks (ICCSN), IEEE, Beijing, China, 2016, pp. 107–111, http://dx.doi.org/10.1109/ICCSN.2016.7586629. [44] V. Ionescu, Fetal ECG extraction from multichannel abdominal ECG recordings for health monitoring during labor, Proc. Technol. 22 (2016) 682–689, http://dx.doi.org/10.1016/j.protcy.2016.01.143. [45] M. Hasan, M. Reaz, M. Ibrahimy, Fetal electrocardiogram extraction and r-peak detection for fetal heart rate monitoring using artificial neural network and correlation, in: The 2011 International Joint Conference on Neural Networks, IEEE, San Jose, CA, USA, 2011, pp. 15–20, http: //dx.doi.org/10.1109/IJCNN.2011.6033193. [46] K. Barnova, R. Martinek, R. Jaros, R. Kahankova, Hybrid methods based on empirical mode decomposition for non-invasive fetal heart rate monitoring, IEEE Access 8 (2020) 51200–51218, http://dx.doi.org/10.1109/ ACCESS.2020.2980254. [47] R. Martinek, R. Kahankova, H. Nazeran, J. Konecny, J. Jezewski, P. Janku, P. Bilik, J. Zidek, J. Nedoma, M. Fajkus, Non-invasive fetal monitoring: a maternal surface ECG electrode placement-based novel approach for optimization of adaptive filter control parameters using the LMS and RLS algorithms, Sensors 17 (5) (2017) 1154, http://dx.doi.org/10.3390/ s17051154. [48] R. Martinek, R. Kahankova, H. Skutova, P. Koudelka, J. Zidek, J. Koziorek, Adaptive signal processing techniques for extracting abdominal fetal electrocardiogram, in: 2016 10th International Symposium on Communication Systems, Networks and Digital Signal Processing (CSNDSP), IEEE, Prague, Czech Republic, 2016, pp. 1–6, http://dx.doi.org/10.1109/CSNDSP. 2016.7573974. [49] K.-A. Lee, W.-S. Gan, S.M. Kuo, Subband Adaptive Filtering: Theory and Implementation, J. Wiley, Chichester, U.K, 2009. [50] S. Douglas, R. Losada, Adaptive filters in matlab: from novice to expert, in: Proceedings of 2002 IEEE 10th Digital Signal Processing Workshop, 2002 and the 2nd Signal Processing Education Workshop., IEEE, Pine Mountain, GA, USA, 2002, pp. 168–173, http://dx.doi.org/10.1109/DSPWS. 2002.1231097. [51] H. Kaur, R. Talwar, Performance comparison of adaptive filter algorithms for noise cancellation, in: 2013 International Conference on Emerging Trends in Communication, Control, Signal Processing and Computing Applications (C2SPCA), IEEE, Bangalore, India, 2013, pp. 1–5, http://dx. doi.org/10.1109/C2SPCA.2013.6749421. [52] Z. Shengkui, M. Zhihong, K. Suiyang, A fast variable step-size LMS algorithm with system identification, in: 2007 2nd IEEE Conference on Industrial Electronics and Applications, IEEE, Harbin, China, 2007, pp. 2340–2345, http://dx.doi.org/10.1109/ICIEA.2007.4318828. [53] P.S. Diniz, Conventional rls adaptive filter, in: Adaptive Filtering, Springer US, Boston, MA, 2008, pp. 1–36, http://dx.doi.org/10.1007/978-0-38768606-6_5. [54] J. Cioffi, T. Kailath, Fast, recursive-least-squares transversal filters for adaptive filtering, IEEE Trans. Acoust. Speech Signal Process. 32 (2) (1984) 304–337, http://dx.doi.org/10.1109/TASSP.1984.1164334. [55] M. Setareh, M. Parniani, F. Aminifar, Non-stationary stabilized fast transversal RLS filter for online power system modal estimation, IEEE Trans. Power Syst. 34 (4) (2019) 2744–2754, http://dx.doi.org/10.1109/ TPWRS.2019.2898168. [56] A. Bessekri, M. Djendi, A. Guessoum, A new simplified fast transversal filter algorithm based on subband approach (SSFTF) for acoustic echo cancellation, Appl. Acoust. 161 (2020) 107178, http://dx.doi.org/10.1016/ j.apacoust.2019.107178. [57] S.-j. Liu, D.-l. Liu, J.-q. Zhang, Y.-j. Zeng, Extraction of fetal electrocardiogram using recursive least squares and normalized least mean squares algorithms, in: 2011 3rd International Conference on Advanced Computer Control, IEEE, Harbin, China, 2011, pp. 333–336, http://dx.doi.org/10.1109/ ICACC.2011.6016426. 18 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 [58] 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, http://dx.doi.org/10.1007/s10439-014-0993-9. [59] S.L. Lima-Herrera, C. Alvarado-Serrano, P.R. Hernandez-Rodriguez, Fetal ECG extraction based on adaptive filters and wavelet transform: validation and application in fetal heart rate variability analysis, in: 2016 13th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE), IEEE, Mexico City, Mexico, 2016, pp. 1–6, http://dx.doi.org/10.1109/ICEEE.2016.7751243. [60] S. Ziani, Y. El Hassouani, Fetal electrocardiogram analysis based on LMS adaptive filtering and complex continuous wavelet 1-d, in: Y. Farhaoui (Ed.), Big Data and Networks Technologies, Vol. 81, Springer International Publishing, Cham, 2020, pp. 360–366, http://dx.doi.org/10.1007/978-3030-23672-4_26. [61] A.M. Kaleem, R.D. Kokate, An efficient adaptive filter for fetal ECG extraction using neural network, J. Intell. Syst. 28 (4) (2019) 589–600, http://dx.doi.org/10.1515/jisys-2017-0031. [62] R. Swarnalatha, D. Prasad, A novel technique for extraction of FECG using multi stage adaptive filtering, J. Appl. Sci. 10 (4) (2010) 319–324, http://dx.doi.org/10.3923/jas.2010.319.324. [63] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.-C. Yen, C.C. Tung, H.H. Liu, The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis, Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. 454 (1971) (1998) 903–995, http://dx.doi.org/10.1098/rspa.1998.0193. [64] Y. Ren, P. Suganthan, N. Srikanth, A comparative study of empirical mode decomposition-based short-term wind speed forecasting methods, IEEE Trans. Sustain. Energy 6 (1) (2015) 236–244, http://dx.doi.org/10.1109/ TSTE.2014.2365580. [65] S. Jha, O. Singh, R.K. Sunkaria, Modified approach for ECG signal denoising based on empirical mode decomposition and moving average filter, Int. J. Med. Eng. Inform. 6 (2) (2014) 165, http://dx.doi.org/10.1504/IJMEI.2014. 060250. [66] Y. Gao, G. Ge, Z. Sheng, E. Sang, Analysis and solution to the mode mixing phenomenon in EMD, in: 2008 Congress on Image and Signal Processing, IEEE, Sanya, China, 2008, pp. 223–227, http://dx.doi.org/10. 1109/CISP.2008.193. [67] Z. Wu, N.E. Huang, Ensemble empirical mode decomposition: a noiseassisted data analysis method, Adv. Adapt. Data Anal. 01 (01) (2009) 1–41, http://dx.doi.org/10.1142/S1793536909000047. [68] K.-M. Chang, Ensemble empirical mode decomposition based ECG noise filtering method, in: 2010 International Conference on Machine Learning and Cybernetics, IEEE, Qingdao, China, 2010, pp. 210–213, http://dx.doi. org/10.1109/ICMLC.2010.5581064. [69] J.-R. Yeh, J.-S. Shieh, N.E. Huang, Complementary ensemble empirical mode decomposition: a novel noise enhanced data analysis method, Adv. Adapt. Data Anal. 02 (02) (2010) 135–156, http://dx.doi.org/10.1142/ S1793536910000422. [70] M.E. Torres, M.A. Colominas, G. Schlotthauer, P. Flandrin, A complete ensemble empirical mode decomposition with adaptive noise, in: 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, Prague, Czech Republic, 2011, pp. 4144–4147, http://dx. doi.org/10.1109/ICASSP.2011.5947265. [71] D.D. Taralunga, I. Gussi, R. Strungaru, A new method for fetal electrocardiogram denoising using blind source separation and empirical mode decomposition, Rev. Roum. Sci. Tech. Ser. Électrotech. Énerg. 61 (1) (2016) 94–98. [72] X. Wei, W. Zheng, An integrated approach for fetal heart rate estimation from abdominal electrocardiogram signal, Chin. J. Electron. 28 (6) (2019) 1198–1203, http://dx.doi.org/10.1049/cje.2019.08.002. [73] P. Ghobadi Azbari, M. Abdolghaffar, S. Mohaqeqi, M. Pooyan, A. Ahmadian, N. Ghanbarzadeh Gashti, A novel approach to the extraction of fetal electrocardiogram based on empirical mode decomposition and correlation analysis, Australas. Phys. Eng. Sci. Med. 40 (3) (2017) 565–574, http://dx.doi.org/10.1007/s13246-017-0560-4. [74] P.R. Muduli, A. Mukherjee, Noise-assisted trend-filtering of fetalelectrocardiogram signals, in: 2016 IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES), IEEE, Malaysia, 2016, pp. 465–469, http://dx.doi.org/10.1109/IECBES.2016.7843494. [75] A. Bin Queyam, S. Kumar Pahuja, D. Singh, Quantification of feto-maternal heart rate from abdominal ecg signal using empirical mode decomposition for heart rate variability analysis, Technologies 5 (4) (2017) 68, http://dx.doi.org/10.3390/technologies5040068. [76] T. Liu, Z. Luo, J. Huang, S. Yan, A comparative study of four kinds of adaptive decomposition algorithms and their applications, Sensors 18 (7) (2018) 2120, http://dx.doi.org/10.3390/s18072120. [77] M.A. Colominas, G. Schlotthauer, M.E. Torres, Improved complete ensemble EMD: A suitable tool for biomedical signal processing, Biomed. Signal Process. Control 14 (2014) 19–29, http://dx.doi.org/10.1016/j.bspc.2014. 06.009. [78] Y.-x. Li, L. Wang, A novel noise reduction technique for underwater acoustic signals based on complete ensemble empirical mode decomposition with adaptive noise, minimum mean square variance criterion and least mean square adaptive filter, Def. Technol. 16 (3) (2020) 543–554, http://dx.doi.org/10.1016/j.dt.2019.07.020. [79] C.J. James, C.W. Hesse, Independent component analysis for biomedical signals, Physiol. Meas. 26 (1) (2005) R15–R39, http://dx.doi.org/10.1088/ 0967-3334/26/1/R02. [80] V. Gupta, M. Mittal, V. Mittal, R-peak detection based chaos analysis of ECG signal, Analog Integr. Circuits Signal Process. 102 (3) (2020) 479–490, http://dx.doi.org/10.1007/s10470-019-01556-1. [81] V. Gupta, M. Mittal, A comparison of ECG signal pre-processing using frft, frwt and IPCA for improved analysis, IRBM 40 (3) (2019) 145–156, http://dx.doi.org/10.1016/j.irbm.2019.04.003. [82] G.R. Naik, D.K. Kumar, An overview of independent component analysis and its applications, Informatica 35 (1) (2011). [83] J.J.R. Immanuel, V. Prabhu, V.J. Christopheraj, D. Sugumar, P. Vanathi, Separation of maternal and fetal ECG signals from the mixed source signal using FASTICA, Procedia Eng. 30 (2012) 356–363, http://dx.doi.org/ 10.1016/j.proeng.2012.01.872. [84] D. Taralunga, M. Ungureanu, R. Strungaru, W. Wolf, Performance comparison of four ICA algorithms applied for fECG extraction from transabdominal recordings, in: ISSCS 2011 - International Symposium on Signals, Circuits and Systems, IEEE, Iasi, Romania, 2011, pp. 1–4, http://dx.doi.org/10.1109/ISSCS.2011.5978768. [85] D. Slock, T. Kailath, Numerically stable fast transversal filters for recursive least squares adaptive filtering, IEEE Trans. Signal Process. 39 (1) (Jan./1991) 92–114, http://dx.doi.org/10.1109/78.80769. [86] D.J. Dechene, Fast transversal recursive least-squares (FT-RLS) algorithm, in: IEEE Trans Signal Proc, Citeseer, 2007. [87] Z. Ren, H. Schütze, A stabilized fast transversal filters algorithm for recursive least squares adaptive filtering, Signal Process. 39 (3) (1994) 235–246, http://dx.doi.org/10.1016/0165-1684(94)90087-6. [88] K. Bensafia, A. Mansour, A.-O. Boudraa, S. Haddab, P. Ariès, B. Clement, Blind separation of ECG signals from noisy signals affected by electrosurgical artifacts, Analog Integr. Circuits Signal Process. 104 (2) (2020) 191–204, http://dx.doi.org/10.1007/s10470-020-01674-1. [89] A. Hyvärinen, E. Oja, Independent component analysis: Algorithms and applications, Neural Netw. 13 (4–5) (2000) 411–430, http://dx.doi.org/10. 1016/S0893-6080(00)00026-5. [90] A. Matonia, J. Jezewski, T. Kupka, M. Jezewski, K. Horoba, J. Wrobel, R. Czabanski, R. Kahankova, Fetal electrocardiograms, direct and abdominal with reference heart beats annotations, 2020, http://dx.doi.org/10.6084/ M9.FIGSHARE.C.4740794.V1. [91] A.L. Goldberger, L.A. Amaral, L. Glass, J.M. Hausdorff, P.C. 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) (2000) http://dx. doi.org/10.1161/01.CIR.101.23.e215. [92] V. Gupta, M. Mittal, V. Mittal, A.K. Sharma, N.K. Saxena, A novel feature extraction-based ECG signal analysis, J. Inst. Eng. (India): Ser. B (2021) http://dx.doi.org/10.1007/s40031-021-00591-9. [93] V. Gupta, M. Mittal, V. Mittal, Performance evaluation of various preprocessing techniques for r-peak detection in ECG signal, IETE J. Res. (2020) 1–16, http://dx.doi.org/10.1080/03772063.2020.1756473. [94] J. Martin Bland, D. Altman, Statistical methods for assessing agreement between two methods of clinical measurement, Lancet 327 (8476) (1986) 307–310, http://dx.doi.org/10.1016/S0140-6736(86)90837-8. [95] K. Lee, B. Lee, Sequential total variation denoising for the extraction of fetal ECG from single-channel maternal abdominal ECG, Sensors 16 (7) (2016) 1020, http://dx.doi.org/10.3390/s16071020. [96] F. Andreotti, J. Behar, S. Zaunseder, J. Oster, G.D. Clifford, An opensource framework for stress-testing non-invasive foetal ECG extraction algorithms, Physiol. Meas. 37 (5) (2016) 627–648, http://dx.doi.org/10. 1088/0967-3334/37/5/627. [97] D. Gurve, S. Krishnan, Separation of fetal-ECG from single-channel abdominal ecg using activation scaled non-negative matrix factorization, IEEE J. Biomed. Health Inf. 24 (3) (2020) 669–680, http://dx.doi.org/10. 1109/JBHI.2019.2920356. [98] R. Sameni, C. Jutten, M. Shamsollahi, Multichannel electrocardiogram decomposition using periodic component analysis, IEEE Trans. Biomed. Eng. 55 (8) (2008) 1935–1940, http://dx.doi.org/10.1109/TBME.2008.919714. [99] B. De Moor, P. De Gersem, B. De Schutter, W. Favoreel, DAISY: A database for identification of systems, J. A 38 (1997) 4–5. [100] E. Castillo, D.P. Morales, A. García, L. Parrilla, V.U. Ruiz, J.A. ÁlvarezBermejo, A clustering-based method for single-channel fetal heart rate monitoring, PLoS One 13 (6) (2018) e0199308, http://dx.doi.org/10.1371/ journal.pone.0199308. 19 K. Barnova, R. Martinek, R. Jaros et al. Applied Soft Computing 113 (2021) 107940 [101] M. Ayat, K. Assaleh, H. Al-Nashash, Extracting fetal ECG from a single maternal abdominal record, in: 2015 IEEE 8th GCC Conference & Exhibition, IEEE, Muscat, Oman, 2015, pp. 1–4, http://dx.doi.org/10.1109/ IEEEGCC.2015.7060027. [102] Y. Zhang, S. Yu, Single-lead noninvasive fetal ECG extraction by means of combining clustering and principal components analysis, Med. Biol. Eng. Comput. 58 (2) (2020) 419–432, http://dx.doi.org/10.1007/s11517-01902087-7. [103] G. Da Poian, R. Bernardini, R. Rinaldo, Separation and analysis of fetal-ECG signals from compressed sensed abdominal ECG recordings, IEEE Trans. Biomed. Eng. 63 (6) (2016) 1269–1279, http://dx.doi.org/10.1109/TBME. 2015.2493726. [104] D. Panigrahy, P. Sahu, Extraction of fetal ECG signal by an improved method using extended Kalman smoother framework from single channel abdominal ECG signal, Australas. Phys. Eng. Sci. Med. 40 (1) (2017) 191–207, http://dx.doi.org/10.1007/s13246-017-0527-5. [105] J. Sandall, R.M. Tribe, L. Avery, G. Mola, G.H. Visser, C.S. Homer, D. Gibbons, N.M. Kelly, H.P. Kennedy, H. Kidanto, P. Taylor, M. Temmerman, Short-term and long-term effects of caesarean section on the health of women and children, Lancet 392 (10155) (2018) 1349–1357, http: //dx.doi.org/10.1016/S0140-6736(18)31930-5. [106] E. Orovou, M. Dagla, G. Iatrakis, A. Lykeridou, C. Tzavara, E. Antoniou, Correlation between kind of cesarean section and posttraumatic stress disorder in greek women, Int. J. Environ. Res. Public Health 17 (5) (2020) 1592, http://dx.doi.org/10.3390/ijerph17051592. [107] S. Petrou, J. Henderson, C. Glazener, Economic aspects of caesarean section and alternative modes of delivery, Best Pract. Res. Clin. Obstet. Gynaecol. 15 (1) (2001) 145–163, http://dx.doi.org/10.1053/beog.2000. 0154. [108] D.M. Okada, A.W. Chow, V.T. Bruce, Neonatal scalp abscess and fetal monitoring: factors associated with infection, Am. J. Obstet. Gynecol. 129 (2) (1977) 185–189, http://dx.doi.org/10.1016/0002-9378(77)90742-6. [109] A.C. Mugdha, F.S. Rawnaque, M.U. Ahmed, A study of recursive least squares (RLS) adaptive filter algorithm in noise removal from ECG signals, in: 2015 International Conference on Informatics, Electronics & Vision (ICIEV), IEEE, Fukuoka, Japan, 2015, pp. 1–6, http://dx.doi.org/10.1109/ ICIEV.2015.7333998. [110] B. Sharma, R.J. Suji, ECG denoising using weiner filter and adaptive least mean square algorithm, in: 2016 IEEE International Conference on Recent Trends in Electronics, Information & Communication Technology (RTEICT), IEEE, Bangalore, India, 2016, pp. 53–57, http://dx.doi.org/10.1109/RTEICT. 2016.7807781. [111] D. Graupe, D. Veselinovic, Blind adaptive filtering of speech from noise of unknown spectrum using a virtual feedback configuration, IEEE Trans. Speech Audio Process. 8 (2) (2000) 146–158, http://dx.doi.org/10.1109/ 89.824699. [112] K.-C. Yen, Y. Zhao, Adaptive co-channel speech separation and recognition, IEEE Trans. Speech Audio Process. 7 (2) (1999) 138–151, http: //dx.doi.org/10.1109/89.748119. [113] B. Widrow, J. Glover, J. McCool, J. Kaunitz, C. Williams, R. Hearn, J. Zeidler, J. Eugene Dong, R. Goodlin, Adaptive noise cancelling: principles and applications, Proc. IEEE 63 (12) (1975) 1692–1716, http://dx.doi.org/10. 1109/PROC.1975.10036. [114] V. Gupta, M. Mittal, A novel method of cardiac arrhythmia detection in electrocardiogram signal, Int. J. Med. Eng. Inform. 12 (5) (2020) 489, http://dx.doi.org/10.1504/IJMEI.2020.109943. [115] A. Giorgio, C. Guaragnella, D.A. Giliberti, Improving ECG signal denoising using wavelet transform for the prediction of malignant arrhythmias, Int. J. Med. Eng. Inform. 12 (2) (2020) 135, http://dx.doi.org/10.1504/IJMEI. 2020.106898. 20