Novel hybrid extraction systems for fetal heart rate variability monitoring based on non-invasive fetal electrocardiogram

Received June 16, 2019, accepted July 26, 2019, date of publication August 7, 2019, date of current version September 25, 2019. Digital Object Identifier 10.1109/ACCESS.2019.2933717 Novel Hybrid Extraction Systems for Fetal Heart Rate Variability Monitoring Based on Non-Invasive Fetal Electrocardiogram RENE JAROS , RADEK MARTINEK, RADANA KAHANKOVA, AND JIRI KOZIOREK Department of Cybernetics and Biomedical Engineering, Faculty of Electrical Engineering and Computer Science, VSB—Technical University of Ostrava, 708 00 Ostrava, Czech Republic Corresponding author: Rene Jaros ([email protected]) This work was supported in part by the Ministry of Education of the Czech Republic under Project SP2019/85 and Project SP2019/118, and in part by the European Regional Development Fund of the Research Centre through the Advanced Mechatronic Systems Project within the Operational Programme Research, Development and Education under Project CZ.02.1.01/0.0/0.0/16_019/0000867. ABSTRACT This study focuses on the design, implementation and subsequent verification of a new type of hybrid extraction system for noninvasive fetal electrocardiogram (NI-fECG) processing. The system designed combines the advantages of individual adaptive and non-adaptive algorithms. The pilot study reviews two innovative hybrid systems called ICA-ANFIS-WT and ICA-RLS-WT. This is a combination of independent component analysis (ICA), adaptive neuro-fuzzy inference system (ANFIS) algorithm or recursive least squares (RLS) algorithm and wavelet transform (WT) algorithm. The study was conducted on clinical practice data (extended ADFECGDB database and Physionet Challenge 2013 database) from the perspective of non-invasive fetal heart rate variability monitoring based on the determination of the overall probability of correct detection (ACC), sensitivity (SE), positive predictive value (PPV) and harmonic mean between SE and PPV (F1). System functionality was verified against a relevant reference obtained by an invasive way using a scalp electrode (ADFECGDB database), or relevant reference obtained by annotations (Physionet Challenge 2013 database). The study showed that ICA-RLS-WT hybrid system achieve better results than ICA-ANFIS-WT. During experiment on ADFECGDB database, the ICA-RLS-WT hybrid system reached ACC > 80 % on 9 recordings out of 12 and the ICA-ANFIS-WT hybrid system reached ACC > 80 % only on 6 recordings out of 12. During experiment on Physionet Challenge 2013 database the ICA-RLS-WT hybrid system reached ACC > 80 % on 13 recordings out of 25 and the ICA-ANFIS-WT hybrid system reached ACC > 80 % only on 7 recordings out of 25. Both hybrid systems achieve provably better results than the individual algorithms tested in previous studies. INDEX TERMS Noninvasive fetal electrocardiography, independent component analysis (ICA), adaptive neuro fuzzy inference system (ANFIS), recursive least squares (RLS), wavelet transform (WT), ICA-ANFISWT, ICA-RLS-WT, hybrid methods, fetal heart rate variability monitoring, extraction systems. I. INTRODUCTION Before the advent of electronics in obstetrics and gynaecology, doctors had to rely on their senses and experience. One of the first methods to detect non-invasive fetal cardiac activity was to listen to (auscultation) heart sounds using a stethoscope [1]. In this way, only basic information such as indicative fetal heart rate (fHR), significant arrhythmias, or cardiac arrest could be obtained about fetal health [2]. With The associate editor coordinating the review of this manuscript and approving it for publication was Sunil Karamchandani. the development of electrical engineering, it was possible to switch to more advanced fetal health monitoring - electronic fetal monitoring (EFM) [3]. In the 1960s, cardiotocography (CTG) was introduced; this is a new method using the ultrasound principle that allows EFM while watching uterine contractions [4]. EFM breeds the possibility of continuous monitoring and, thus, early detection of symptoms of fetal hypoxia [5] and other life-threatening conditions. This has led to a significant reduction in neonatal mortality, as shown, for example, by the study presented by Chen et al. in [6]. CTG is currently 131758 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/ VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG technically sophisticated and forms an essential part of modern obstetrics for its simplicity, speed, non-invasiveness and painlessness [7]. Despite these innumerable advantages, however, it faces problematic reliability and accuracy. In addition, its great drawback is the fact that it does not provide any information about Beat to beat (BTB) variabilities [8]. This is mainly due to the methods that are used to reduce the number of signal loss episodes. These methods include, in particular, correlation methods based on signal periodicity analysis. Their use leads to significant averaging of instantaneous fHR values [8]. It is therefore not possible to detect rapid changes in fHR - so-called short-term variability (STV) [8], [9], which is important for the assessment of fetal health. Cases, when fetal hypoxia is diagnosed falsely positively, pose a big problem. Many studies [10]–[16] blame CTG for the large increase in the number of Caesarean sections, which has provably increased since this method was put into practice. The acute caesarean section, as a very invasive surgical procedure, is a great burden for the woman’s, as well as the child’s, body compared to the natural mode of delivery. Many researches [17]–[19] show that it is not only a physical, but also mental, burden that can lead to post-traumatic stress. In addition, this type of delivery is associated with considerable economic losses for hospital facilities [20], [21]. Fetal electrocardiography is a monitoring technique based on electrical potential monitoring, manifesting cardiac activity, specifically in the form of a fetal electrocardiogram (fECG). Fetal ECG has been studied for over 100 years - it was first mentioned in the scientific literature as early as 1906, when Cremer presented the first record of an abdominal fECG [22]. However, significant progress in this area took place about 50 years later, mainly due to the work of three independent groups of scientists in London, Stockholm and Paris [1]. This led in particular to the first direct fECG recordings that were presented at the Mount Sinai Hospital meeting in 1956 [23]. This accelerated further research in this field, in particular, the investigation of fECG morphology, and the heart rate derived therefrom, and their changes in the case of pathological conditions [1]. These findings have led to a better understanding of the pathophysiology of the fetal cardiovascular system during childbirth and, thus, to recommendations for obstetricians used also for CTG monitoring [1]. The first attempts to capture the non-invasive fECG at the beginning of the 1960s were limited mainly by the technical possibilities at that time. Many authors [24]–[26] have been able to capture abdominal ECGs (aECGs) and describe them in relation to invasive recordings, but the major limitation was the large amount of interference that was sensed along with the useful signal, especially the maternal ECG (mECG). Although there were efforts to perform automatic reduction of the maternal component from the aECG, its complete elimination was not achieved [1]. The biggest problem with the reduction of the maternal component is the fact that the fECG signal overlaps in the time and frequency domain and, moreover, its amplitude is several times higher. Hence, classical linear filtration is ineffective in this respect and more advanced methods are required. With advances in computer technology and signal processing techniques, significant progress has been made in the extraction of fECG in recent decades. A large number of authors have achieved successful extraction using various methods and approaches, including: 1) Methods using only abdominal electrodes (AES methods) [27]–[65], 2) Methods using a combination of abdominal and thoracic electrodes (CS methods) [52], [66]–[89]. Healthcare professionals prefer the former, as it is clinically better feasible and more comfortable for the patient. Thus, most commercially available fECG-based devices use only abdominal measuring electrodes. At present, the main parameter in obstetrics is fHR. In this respect, a large number of AES algorithms have proven to be sufficient [32], [90], [91], leading to the implementation of the first commercially available NI-fECG-based systems in clinical practice in recent years [92], [93]. On the other hand, it should be noted that many authors have achieved very good results with adaptive methods using the CS approach [52], [66], [67], [71], [89]. As already mentioned, AES methods are useful for determining fHR obtained by detecting R-peaks of the ECG waveform. Nevertheless, processing by these methods generally distorts the fECG waveform, thus changing its morphology [94]–[96] (ST segment, QT interval, T/QRS ratio), which is important from a diagnostic perspective. From this perspective, adaptive CS methods are more effective. Therefore, combining both approaches is one way to improve the quality of fECG extraction. Examples of such research have been presented in [97], where hybrid methods based on combination of CS and AES algorithms appear to be the most effective methods, see [59], [98]–[102]. By developing new or improving existing methods of fECG monitoring, it could lead to both improved diagnosis of fetal hypoxia and minimized unnecessary Caesarean sections due to putative hypoxia. At the same time, the ST analysis of the ST section of the fECG waveform (STAN) is an available option, see [94]–[96]. This sophisticated method enables observation of the ST segment of the ECG waveform of the fetus, which is sensed transvaginally using a fetal scalp electrode (FSE). It is an invasive method that can only be used during delivery. An example of a device capable of performing ST analysis is, for example, STAN S31 made by the Swedish company Neoventa Medical AB [103]. The main challenge of current research is developing a system based on NI-fECG able to provide morphological analysis similar to the invasive ST analysis (i.e. NI-STAN). This paper proposed a method that, when optimized, would be able to achieve this task. To achieve this, it is necessary to find suitable algorithms in terms of both the extraction quality and the clinical feasibility, which is associated with both the performance and the computational cost. Some of the authors tested the methods separately, some combined other methods. The results of studies [98]–[102] VOLUME 7, 2019 131759 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG show that hybrid methods outperform the other methods by means of the quality of the fECG extraction. The aim of the study proposed therein is combining the methods in order to combine their advantages and to eliminate their drawbacks. The approach of evaluating the results is also novel since we investigated the overall trend of the fHR curve in comparison with the reference FSE along with the standard evaluation by means of quality parameters such as accuracy (ACC), sensitivity (SE), positive predictive value (PPV), the harmonic mean between SE and PPV (F1), and so on. The original contribution of this article is to test new hybrid methods combining independent component analysis (ICA), adaptive neuro-fuzzy inference system (ANFIS) algorithm, recursive least squares (RLS) algorithm and wavelet transform (WT) algorithm. Separately, these methods are relatively effective in fECG extraction [52], [66], [67], [71], [89]. Thus, their incorporation into the extraction and post-processing phases in a complex fully automated hybrid system seems very advantageous. Many authors have only tested their algorithms using synthetic data. However, the algorithm presented in this article has been verified on real clinical data - using the gold standard in the form of a signal continuously scanned with a FSE [104], [105]. The team of authors, as well as other research teams around the world in the past, tested both ICA and principal component analysis (PCA) [29], [51], [58], [98], [106] and adaptive systems using algorithms such as least mean squares (LMS) and RLS [69], [71], [107]–[109]. Based on the results supported by the statistical analysis, it can be stated that the robust hybrid system presented here achieves far better results in the determination of the fHR than the two aforementioned approaches alone. At the same time, this system is potentially able to extract the fECG waveform in such a way that a more detailed morphological analysis can be performed. This could lead to creating a new non-invasive alternative to the current STAN method. Based on the in-depth research [97] and the initial experiments [58], [110] conducted, the ICA algorithm for the first part of the hybrid methods was selected. It has been found that the ICA algorithm is able to extract an mECG component from the aECG signals containing only the mECG signal and an aECG component containing the mECG signal and the fECG signal that is highlighted and is located at the same amplitude level as the mECG signal (marked as aECG*). Pre-processing of input aECG signals for adaptive algorithms thus forms an ICA algorithm. The mECG and aECG signals obtained after pre-processing are used as inputs to two different adaptive algorithms. These are ANFIS and RLS adaptive algorithms. The output fECG signal from both adaptive algorithms needs to be smoothed using WT (after fECG signal processing). The combination of these algorithms resulted in 2 hybrid methods, ICA-ANFIS-WT and ICA-RLS-WT. This chapter will deal with the 4 algorithms (ICA algorithm, ANFIS algorithm, RLS algorithm and WT algorithm) and their mathematical description. A. INDEPENDENT COMPONENT ANALYSIS This is an algorithm that tries to find a linear representation of non-Gaussian data that contains statistically independent components. When processing the fECG signal, the ICA algorithm principle can be easily explained. There are 2 electrodes located in the abdominal area of a pregnant woman providing 2 time signals: x1(t) and x2(t). These signals then include the sum of the signals induced by the maternal and fetal cardiac activity designated s1(t) and s2(t) (in reality, moreover, the noise-induced signals). In general, the composition of the signals x1(t) and xn(t) can be described by equations (1) and (2), where Amix denotes a mixing matrix, a denote parameters depending on the distance of individual cardiac activities from the electrodes and ndenotes the number of statistically independent components. The problem is that the parameters aare not known. The only possibility is to assume that the signals s1(t) and s2(t) are statistically independent, which is confirmed when processing the fECG signal. Equation (3) is then used to estimate independent components from mixed aECG signals, where Wis the inverse matrix from the Amix matrix [29], [58], [110]–[112]. xj=aj1s1+aj2s2+ · · · + ajnsn.(1) −→ x=Amix−→ s= n X i=1 aisi.(2) −→ s=W−→ x.(3) The most widely used type of ICA algorithm is the FastICA algorithm. It is based on a fixed iteration scheme looking for maxima of data not derived from −→ wT−→ xnormal data distribution. To devise the FastICA algorithm, weight vector −→ wand derivative gof the non-quadratic Gfunction are needed. The FastICA algorithm is based on 4 steps. Before performing the following steps, pre-processing using centring must be applied to create data of zero mean value and whitening to create data vectors whose components are subsequently uncorrelated with unit scattering. Furthermore, the convergence criterion δ(δ=0.00001 is often used), the maximum number of iterations of the kICA cycle (often kICA =100) and the number of ICA,noutput components must be selected. The smallest recommended number of output components for fECG signal processing is ICA,n=3. Convergence seeks to achieve practically zero scalar product between old and new vector values. First, random standardized initial weights of vector −→ wT−→ xare created. Then the current vector −→ w+is stored in vector −→ w, and the equation (4) is used to calculate the kurtosis, or the negentropy can be calculated in this step. Subsequently, the standardization is performed using equation (5). The last step checks whether the scalar product of the new vector −→ w+and vector −→ wis smaller than the selected convergence criterion δ, and whether the cycle has been run more times than the maximum number of kICA iterations selected. If the condition is not met, the second and third steps of the FastICA algorithm are repeated[110], [112]. −→ w+=E{−→ x g(−→ wT−→ x)} − E{g0(−→ wT−→ x)}−→ w.(4) 131760 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG −→ w+= −→ w+ k−→ w+k.(5) B. RECURSIVE LEAST SQUARES FILTER Adaptive algorithms calculate error e(n) between the desired and real output of the adaptive algorithm using equation (6), where d(n) is the desired output of the adaptive algorithm and y(n) is the actual output of the adaptive algorithm. The adaptive algorithm attempts to adjust the filter coefficients w(n) to achieve an output as much correlated as the desired one [71], [113]. e(n)=d(n)−y(n).(6) The RLS algorithm is based on recursive determination of weight coefficients, Kalman filter theory, time averaging and also on the LMS algorithm. The advantage of this algorithm is that it uses the values of previous error estimates and has very high performance in time-varying environments. The disadvantage then is that it has higher computing complexity and has stability problems. Its aim is to minimize the result of the objective function ξ, see equation (7), where nrepresents the external time index indicating the number of last values and kdenotes the internal time index. The required output pcan be expressed by equation (8), where λdenotes the forgetting factor being in the interval <0,1>. Most often, the forgetting factor ranges from 0.95 to 0.99. The forgetting factor λserves for forgetting the previous values. It is desirable to make a compromise between trying to achieve parameter convergence using λ=1 and between the ability to monitor the algorithm’s sensitivity to changing parameters using λ < 1. The ideal solution to the problem is to use a variable forgetting factor [71], [113], [114]. ξ(n)= n X k=1 pn(k)e2 n(k).(7) pn(k)=λn−k.(8) From equation (7), it is clear that at a certain time n, all the values obtained since the start of the RLS algorithm must be available. This means that the number of values processed rises with increasing time and, therefore, the RLS algorithm is very memory intensive. To reduce the computational complexity, the NRLS filter order is used, which indicates the final number of previous values processed [71], [113], [114]. This algorithm updates the current variables in each iteration cycle based on the state in the previous iteration. When implementing the RLS algorithm, it is possible to reduce the computational complexity (except for decreasing the order of the filter) by omitting the inverse matrix step, whose calculation is not necessary in practice. In the first step, the filter output is calculated using the input vector of the current iteration and using the filter weights obtained during the previous iteration, see equation (9), where the vector −→ x(n) denotes the input signal. In the second step, the mean gain vector is calculated using equation (10) and (11). The third step solves equation (12), which determines the value of the estimation error. In the following step, the balance vector −→ w(n) is updated using the estimation error value and the gain vectors, see equation (13). In the last step, the inverse matrix is calculated, see equation (14), and the pattern selection from the training set is terminated. In Figure 6, the RLS algorithm application block can be seen, where the λparameter shows the forgetting coefficient and NRLS indicates the filter order [71], [113], [114]. −→ yn−1(n)=−→ wT(n−1)−→ x(n).(9) −→ u(n)=˜ 9−1 λ(n−1)−→ x(n).(10) −→ k(n)=1 λ+−→ xT(n)−→ u(n) −→ u(n).(11) −→ en−1(n)=d(n)−−→ yn−1(n).(12) −→ w(n)=−→ wT(n−1) +−→ k(n)−→ en−1(n).(13) ˜ 9−1 λ(n)=λ−1˜ 9−1 λ(n−1) −−→ k(n)[−→ xT(n)˜ 9−1 λ(n−1)].(14) C. ADAPTIVE NEURO FUZZY INFERENCE SYSTEM This is the most commonly used adaptive algorithm to extract the fECG signal, developed by Jang in his work in 1993 [115]. This adaptive algorithm belongs to softcomputing methods, which are methods based on analytical methods, boolean logic, sharp classification, and deterministic search. The ANFIS algorithm can often be referred to as a hybrid adaptive algorithm because it combines a fuzzy inference system such as Takagi-Sugeno (fuzzy logic) [116], [117] and a feed-forward neural network learning algorithm [118], [119]. It is, therefore, a very powerful algorithm utilizing the advantage of fuzzy expert systems (the ability to work with inaccurate data) and neural networks (learning from the environment). In addition, the ANFIS algorithm can work with a learning algorithm consisting only of a backpropagation algorithm (BP) or a combination of BP and LMS algorithm [87], [120]. Thus, the ANFIS algorithm manifests itself as a fuzzy expert system implemented by a multilayer feed-forward neural network, and it is important to include Sugeno zero or first order model. Next, the system needs only one output, has no shared rules, the output membership functions are of the same type (linear or constant), and the number of rules correlates with the number of membership functions. Figure 5shows the block diagram of the ANFIS algorithm. As the inputs of the noise cancellation system using ANFIS algorithm, we use mECG a mECG*, which is mECG with a 1 sample delay that is necessary for a correct estimation of the maternal component from the aECG* signal. It can be seen from the figure that the ANFIS algorithm architecture consists of 5 feed-forward layers. The ANFIS algorithm rule base can be described by two IF-THEN rules, see equations (15) and (16) [80], [87], [120]. R1: IF xis A1and yis B1, THEN z1=p1X+q1Y+r1.(15) VOLUME 7, 2019 131761 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG R2: IF xis A2and yis B2, THEN z2=p2X+q2Y+r2.(16) In the 1st layer, which is adaptive and referred to as an input layer, fuzzification of input language variables representing individual nodes is performed. The functions for individual nodes can be written using equation (17). The neural network adapts, by its learning in the first layer, parameters of membership functions representing the antecedent. In the 2nd layer, which is called a rule layer, a multiplication of the first layer output signals is performed to determine the rules of weights w. The antecedent of rules is composed of language values of language variables. Thus, this layer consists of only non-adaptive nodes representing individual Takagi-Sugeno fuzzy rules. The weight of each rule is at the output of the 2nd layer and can be calculated using equation (18). The next layer represents the standardization layer composed only of non-adaptive nodes. The output of the 3rd layer is the ratio of the weights of individual rules to the sum of all rules. The standardized force of each rule is calculated using equation (19). The defuzzification layer (4th layer) uses adaptive nodes that have a linear or constant transmission function determined by the consequent. The 4th layer is connected to the standardization nodes as well as to the language variables xand y. The calculation of the adaptive nodes of the 4th layer is performed by equation (20), where p,qand rrepresent the parameters of the consequent. The last, 5th layer, consists of one non-adaptive node called summation, which uses equation (21) to determine ANFIS [80], [87], [120]. o1,i=µAi(x). o1,i=µBi(y).(17) o2,i=wi=µAi(x)·µBi(y).(18) o3,i=wi=wi w1+w2.(19) o4,i=wifi=wi·zi=wi(pix+qiy+ri).(20) o5,i=X i wifi=Piwifi Piwi.(21) Correct functionality of the ANFIS algorithm is very often performed by a combination of suitably set nonlinear parameters of the 1st layer using the BP and appropriately set linear parameters of the 4th layer using the LMS algorithm. This hybrid combination of BP and LMS algorithm was also used in our study. At the output of the ANFIS algorithm, the total error between the ANFIS output and the desired output is calculated. The program ends when an optimally small total error is reached or when the number of selected epochs (iterations) is exceeded. This combination of the ANFIS algorithm using BP and the LMS algorithm includes forward and reverse run [80], [87], [120]. D. WAVELET TRANSFORM It is a very similar algorithm as the Fourier transform, but its great advantage is that it is very effective in processing non-stationary signals and signals containing multiple components (such as fECG signal processing). The essence of the WT algorithm is to select the appropriate shape and, then, the width of the wavelet. Often, wavelet types such as Daubechies, Symlets, and Coiflet are used in fECG signal processing. The Daubechies wavelet type seems to be the most suitable [39], [121], so in this work, this wavelet is selected when performing R-peak detection and smoothing the resulting fECG signal. In case a wrong type of mother wavelet is used, the result of WT could be unsufficient. Indeed, it could cause distortion of the output signal leading to inaccurate R-peaks detection. Basically, this algorithm performs signal decomposition using the selected type and width of the maternal wavelet 9[46], [122]. When applying discrete WT (DWT), it is necessary to choose the type of maternal wavelet 9and the decomposition level n. In the first decomposition stage, the signal is decomposed by the Low-pass filter to one cA approximation component (containing the lower half of the frequencies and providing the overall signal trend) and by the high-pass filter to one detailed cD component (containing the upper half of the frequencies and providing additional fineness information). In the second decomposition stage, the approximation component of the signal is divided into another one approximation component and one detailed component. This is happening until the final selected degree of decomposition [46], [122]. Equation (22) is used to describe DWT, where 9∗ j,kis a complex conjugate function to the daughter wavelet 9j,k. The daughter wave, which is dilated and shifted, can be described by equation (23), where jis the number of wavelets needed to cover the maternal wavelet and kis the wavelet position over time. The reverse DWT contains, at its input, signal c composed of the last approximation component and of all the detailed components from the highest degree of decomposition to the lowest one, see equation (24), where ndenotes the degree of decomposition chosen. This composite signal is adjusted by thresholding before the DTW is completed. With reverse DWT, convolution with reconstruction filters is applied. Equation (25) is used to calculate the inverse DWT [46], [122]. DWT (f)=F(j,k)=ˆ f(j,k) =Z∞ −∞ f(t)·9∗ j,k(t)dt.(22) 9j,k(t)=2j 29(2jt−k).(23) c=cAn+cDn+cAn−1+ · · · + cA1.(24) DWT −1(F)=f(t)=X j,k cj,k·9j,k(t).(25) Thresholding removes part of the noise by setting the coefficients to zero. The coefficients representing the useful portion of the signal are left undamaged. Adaptive thresholding uses the calculation of the noise standard deviation σin a floating window with the selected length land the selected empirical constant Kto calculate the empirical threshold λ, see equation (26). Subsequently, it is advisable to select soft thresholding. First, the coefficients having a value 131762 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG smaller than the threshold value are set to zero. Thereafter, the remaining coefficients having a value greater than the threshold value are shifted to zero by the threshold size. Equation (27) shows soft thresholding where Udenotes the data processed [46], [122]–[124]. λj,k=σj,k·K.(26) D(U, λ)=sgn(U)max(0,|U| − λ).(27) II. EVALUATION PARAMETERS Correct evaluation is vital to verify the accuracy of the proposed method. Various evaluation parameters are used among different studies, so it makes an objective comparison challenging. This chapter describes the evaluation parameters that were selected for the evaluation of the results in order to ensure objectivity and repeatability of the experiments. A. DETERMINATION OF EXTRACTION ACCURACY To determine the accuracy of the extraction of the fECG signal relative to the reference fECG signal, the use of parameters determining the true positive values (TP), false positive values (FP) and false negative values (FN) was chosen. The number of TPs indicates how many times the R-peak was determined by the detector in the extracted signal at certain signal locations where R-peaks are to be determined by reference. The TP determination interval was 50 ms left and 50 ms right of the reference annotations. This interval was determined based on a study by Billeci and Varanini in 2017 [125]. This interval then serves to determine the number of FNs. If, according to the reference annotation, an Rpeak was to be a frequency in the given point of the signal, but it was not determined in the extracted fECG signal; this means omitting the R-peak while increasing the number of FNs. The number FP indicates the determination of R-peaks outside the intervals in which the R-peaks are located by reference. When displaying the variable of fHR over time. An example of TP, FP and FN determination can be seen in Chapter IV, Figure 11. By means of the TP, FP and FN values obtained, ACC can be calculated using equation (28), which is an estimate of the probability of correct detection of the R-peak (TP) by the hybrid method used relative to the reference. Parameter SE can be calculated using equation (29), PPV can be calculated using equation (30) and F1 can be calculated using equation (31) [58], [125]. ACC =TP n·100 =TP TP +FP +FN ·100 (%).(28) SE =TP TP +FN ·100 (%).(29) PPV =TP TP +FP ·100 (%).(30) F1=2·SE ·PPV SE +PPV =2·TP 2·TP +FP +FN ·100 (%).(31) B. BLAND-ALTMAN GRAPH If, in statistics, the goal is to evaluate pairs of random variables (X1,Y1), (X2,Y2) to (Xn,Yn), which form pairs of dependent observations, it is necessary to use pair tests when verifying the position. Such tests include the Bland-Altman graph. In the first step of the pair test, the difference −→ Dand the averages −→ Mbetween the pairs of random variables is calculated, see equations (32) and (33), where −→ Xis a vector of random variables obtained from the fHR reference variable over time, −→ Yis a vector of random variables obtained from the estimated fHR variable over time using the test method, nis the number of elements of the individual vectors and i is the i-th element of the individual vectors. It can then be assumed that the quantities (D1,D2,...,Dn) are statistically independent and have the same distribution with the mean value µ=µ1−µ2[58], [126]–[128]. −→ D=(D1,D2,...,Dn),where Di=Xi−Yi.(32) −→ M=(M1,M2,...,Mn),where Mi=Xi+Yi 2.(33) The standardized normal distribution is based on point µ±1.96σindicating the 97.5 % quantile of the normal distribution, where µindicates the mean value and σthe standard deviation. Then, from equation (34) and the modification of this equation, see equations (35), (36) and (37), where P denotes the probability and zdenotes the normal distribution quantile point, it follows that 95 % of the area of normal distribution lies in the interval µ±1.96σ. The mean value of µand 1.96σfrom the difference vector −→ Dis calculated on the basis of equations (38) and (39), where nis the total number of elements of the difference vector −→ Dand iis the i-th element of the vector −→ D[58], [126]–[128]. P(z0.025 < −→ D−µ σ<z0.975)=0.95.(34) P(−z0.975 < −→ D−µ σ<z0.975)=0.95.(35) P(−1.96 < −→ D−µ σ<1.96) =0.95.(36) P(µ−1.96σ < −→ D< µ +1.96σ)=0.95.(37) µ= n P i=1 Di n.(38) 1.96σ=1.96 · v u u u t n P i=1 (Di−µ)2 n−1. (39) To construct a Bland-Altman graph, first, the values of the vectors −→ Dand −→ Mare plotted, then the middle line marking the mean value µand then the upper and lower limits of compliance (LoA) are plotted using two lines indicating µ±1.96σ[58], [126]–[128]. VOLUME 7, 2019 131763 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 1. Block diagram of the hybrid system. III. MATERIAL AND METHODS In this chapter, the basic methods used in this article will be presented. This is a description of the dataset that was used to test the methods presented. Furthermore, both the hybrid extraction systems tested and the validation process are discussed in detail. A. DATASET In this work, a database of abdominal and direct fetal electrocardiogram (ADFECGDB) was used, which is a freely available database from Physionet sites [104], [105], [129]–[131]. This database contains 5 five-channel recordings from 5 different pregnant women in the 38th to 41stweek of pregnancy which were taken at birth. All 5 recordings were taken at the pulmonary department of the Medical University of Zabrze, Poland. The individual signals are recorded with a bandwidth of 1 to 150 Hz (with 50 Hz network interference removed), a sampling rate of 1 kHz, a resolution of 16 bits and a length of 5 minutes. Individual recordings contain 4 aECG signals measured on the maternal abdomen and one fECG signal measured directly from the fetal head surface. 4 silver chloride electrodes placed around the navel (electrode surface was ground to reduce skin impedance), a reference electrode placed above the pubic symphysis, and an active electrode located on the left lower limb were used to measure aECG signals. The direct fECG signal was measured transvaginally using a typical spiral electrode. From the direct fECG signal measured, the positions of the R-peaks were automatically marked with the sensing system used and were subsequently verified by a group of cardiologists to compile accurate reference markers (annotations). In addition to the 5 recordings listed on the Physionet site (r01, r04, r07, r08, and r10), other 7 recordings (r02, r03, r05, r06, r09, r11, and r12) with sampling rate of 500 Hz were included in this study. Another database used was set A of Physionet Challenge 2013 [131]. This database was created to improve the development of accurate fHR, fetal RR-interval, or fetal QT-interval estimation algorithms. Total of 25 records (a01 to a25) contain 4 aECG signals. The length of the individual signals is 1 minute with a sampling frequency of 1 kHz and a resolution of 12 bits. set A is consists of reference annotations that have been created and indicate the positions of the individual R-peaks. B. HYBRID SYSTEM DESCRIPTION This part describes in detail the concept of novel hybrid extraction systems for fHR variability monitoring based on NI-fECG. Subsequently, the modification of the fHR variable curves detected over time based on decision making and moving averaging used in all 12 recordings employed is described. Finally, a statistical comparison of the fHR variables estimated over time follows using two new hybrid methods of ICA-ANFIS-WT and ICA-RLS-WT against the fHR reference variable over time. In Fig. 1, a block diagram of the hybrid system implemented can be seen. The methods employed consist of several steps which are performed successively to complete the successful extraction of the fECG signal. An example of extraction on the block diagram is shown on recording r01. First, at least 2 measured aECG signals are fed to the preprocessing block where the initial signal adjustment is performed, see Fig. 2. The electrode layout in this figure is based on the ADFECGDB database [104], [105], [129]–[131]. The input signals are labelled as AE1to AE4, the reference signal is denoted as AE0and the active ground is denoted as N. It is possible to select any filter of a certain band, but, in our case, the FIR filter was chosen. When using a bandpass filter, it is possible to set lower limit frequency fFIR,L, upper limit frequency fFIR,U, filter 131764 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 2. Input signal preprocessing. order NFIR and, in addition, it is necessary to set sampling frequency fs of the input aECG signals used. To adjust aECG signals, fFIR,L=3Hz,fFIR,U=150 Hz,NFIR =500 and fs =500;1000 Hz were set (5 recordingss from the ADFECGDB database had a sampling frequency of 1 kHz and 7 recordings had a sampling frequency of 500 Hz). The band selected cannot, in any way, damage the fECG signal because the fetal QRS complex is, according to the study by Sameni et al. in 2010 [132], in the range of 10 to 15 Hz. The FIR filter used with the bandpass set is designed to eliminate the variation of isolines, but also to remove noise, for example in the form of maternal and fetal movements. Signals after preprocessing are then brought to the ICA algorithm block, see Fig. 3. In this block, the number of components ICA,n, the convergence criterion δand the maximum number of iterations of the cycle kICA can be selected. ICA,n=3, δ=0.00001, and kICA =100 were set for our experiments. The output of the ICA algorithm is then formed by 3 components. The scheme described shows that, in most cases, one component practically corresponds to the mECG signal, the second component corresponds to the aECG signal with the enhanced fECG signal at the same amplitude level as the mECG signal (marked as aECG*), and the third component corresponds to noise. The individual components are in a different order each time the ICA algorithm is started, they have an altered amplitude due to standardization within the cycle, the components are rotated, and moreover, the components are time-shifted by several samples. For this reason, the output components are fed to an auto-centring block where the components are first differentiated and, then, rotated in the right direction, as shown in Fig. 4. Subsequently, the component corresponding to the mECG signal is amplitude and time aligned, based on the mQRS complexes, with the component corresponding to the aECG* signal with enhanced fECG signal. Finally, re-standardization is performed because amplitude adjustments occur during centring. If the automatic centring result is unsatisfactory, it is possible to use manual centring, where the user subjectively chooses the mECG signal and the aECG* signal with the fECG signal enhanced (or the VOLUME 7, 2019 131765 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 3. Implementation of the ICA algorithm to obtain components for adaptive systems. FIGURE 4. Progress of automatically selected algorithm and centring of output components from the ICA algorithm. aECG signal can be selected from the input signals), and then the person aligns the signals by amplitude and time. After finishing the automatic or manual centring, the mECG signal and the aECG* signal are standardized. In the remaining course of the program, the output signals then have a dimensionless unit, and, thus, output fECG signals too. Thus, the signals are ready for entering the adaptive algorithm and for the subsequent extraction of the fECG signal. It is now possible to select either the extraction of the fECG signal using the ANFIS algorithm, see Fig. 5, or the RLS algorithm, see Fig. 6. Using the ANFIS algorithm, it is possible to select the shape of membership functions µANF, the number of membership functions ANF,nand the number 131766 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG TABLE 4. Determination of TP, FP and FN based on the detection of R-peak locations in the fECG signals extracted from ADFECGDB database using the hybrid ICA-RLS-WT method relative to R-peak locations in the reference signal, and calculation of the ACC, SE, PPV and F1 parameters. FIGURE 12. Graphical comparison of estimated and reference fHR variables over time from ADFECGDB database. a) comparison of estimated variables using the ICA-ANFIS-WT hybrid method; and b) comparison of estimated variables using the hybrid ICA-RLS-WT method. surgical termination of pregnancy. Fig. 12 shows that the ICA-RLS-WT hybrid method provides more similar fHR variable curves over time than the ICA-ANFIS-WT hybrid method. Thus, it can be stated that, based on visual assessment, the ICA-RLS-WT hybrid method achieves better fECG signal extraction accuracy. C. BLAND-ALTMAN GRAPHS In this part of the test, Bland-Altman graphs were plotted between the individual estimated and reference fHR variables were over time after adjustments from chapter IV-B. Fig. 13 and 14 show 2 examples of accurate estimated variables of fHR over time and Fig. 15 and 16 show 2 examples of inaccurately estimated variables of fHR over time. Initially, before the Bland-Altman graphs were plotted, two vectors were calculated betweenthe individual estimated fHR variables over time and between the fHR reference variables over time using both hybrid methods. The first difference vectors −→ Dwere made for the Y-axes, which show the difVOLUME 7, 2019 131773 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 13. Comparison of estimated and reference variables of fHR over time from recording r01. a) comparison of the reference and estimated variables of fHR over time using ICA-ANFIS-WT based on the Bland-Altman graph; and b) comparison of the reference and estimated variables of fHR over time using ICA-RLS-WT based on the Bland-Altman graph. FIGURE 14. Comparison of estimated and reference variables of fHR over time from recording r08. a) comparison of the reference and estimated variables of fHR over time using ICA-ANFIS-WT based on the Bland-Altman graph; and b) comparison of the reference and estimated variables of fHR over time using ICA-RLS-WT based on the Bland-Altman graph. FIGURE 15. Comparison of estimated and reference variables of fHR over time from recording r11. a) comparison of the reference and estimated variables of fHR over time using ICA-ANFIS-WT based on the Bland-Altman graph; and b) comparison of the reference and estimated variables of fHR over time using ICA-RLS-WT based on the Bland-Altman graph. ferences between the values of the fHR variables at the time of the reference and estimated signals, see equation (32). The second vectors −→ Mwere made for the X-axes, which show the averages between the values of the fHR variables at the time of the reference and estimated signals, see equation (33). The elements of both of the vectors −→ Dand −→ M 131774 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 16. Comparison of estimated and reference variables of fHR over time from recording r12. a) comparison of the reference and estimated variables of fHR over time using ICA-ANFIS-WT based on the Bland-Altman graph; and b) comparison of the reference and estimated variables of fHR over time using ICA-RLS-WT based on the Bland-Altman graph. were then plotted into individual Bland-Altman graphs. The main part of the Bland-Altman graph construction is the calculation of mean value µ, which is shown in the graph by the middle thick horizontal line. Subsequently, on the basis of this mean value µand the calculated standard deviation σ, two LoAs are determined using µ±1.96σwhich are represented by the remaining two thinner lines in the graph. Equations (37), (38) and (39) are used for the calculation. The individual mean values µand values 1.96σdetermined were recorded in Table 5. In the tables, it is necessary to monitor whether the mean value µapproaches 0. At a high positive or negative value, the mean value µwill show that there were large values in the difference vector −→ Dand, therefore, there is a large difference between the estimated and the reference variables of fHR over time. A similar phenomenoncan bethen observedinvalues 1.96σ. Table5shows that a low mean value µwas achieved for r01, r02, r03, r05, r08, and r09 recordings, indicating that the ICA-ANFIS-WT hybrid method is effective in these recordings. Nevertheless, r03 recording shows a fairly high value of 1.96σ, so, for this recording, we have to disprove the statement. This means that good fECG signal extraction efficiency using the hybrid ICA-ANFIS-WT method can be monitored, based on this testing, only in r01, r02, r05 r08 and r09 recordings. Table 5also shows that the ICA-RLS-WT hybrid method has low mean values µfor r01, r02, r03, r05, r06, r08, r09 and r10 recordings, indicating good efficiency of the ICA-RLS-WT hybrid method in these recordings. From the recordings marked, large value of 1.96σcan be seen in r03, r06 and 10 recordings, which means that good fECG signal extraction efficiency using the hybrid ICA-RLS-WT method can be monitored in the remaining 5 recordings. D. ECG PHYSIONET CHALLENGE 2013 This chapter deals with evaluation of both hybrid methods on dataset from set A, Physionet Challenge 2013. Previous experiments proved that ICA-RLS-WT hybrid method TABLE 5. The recorded mean values and values after plotting Bland-Altman graph for individual adjusted and estimated fHR variables over time using the ICA-ANFIS-WT and ICA-RLS-WT hybrid methods relative to fHR reference variables over time. performs better than ICA-ANFIS-WT hybrid method. Experiments carried out using ADFECGDB database were very extensive, so experiment carried out using Physionet Challenge 2013 database is performed only on evaluation of R-Peak detection accuracy (similar to experiment in chapter IV-A). After performing optimization and associated extractions on all 25 recordings using both hybrid methods, the detection of R-peaks was performer using a CWT-based detector. Then determination of of TP, FP and FN and calculation of selected parameters, ACC, SE, PPV and F1, was performed on all recording from Physionet Challenge 2013 database. Table 6shows the results for the ICA-ANFIS-WT hybrid method. It can be seen that the method worked, based on ACC, above 95 % with 6 recordings (a04, a05, a08, a15, a17 and a22), and over 80 % for a01 recording. On the basis of PPV, it worked, moreover, above 95 % with a01 recording, which, according to the ACC, SE and F1 parameters, did not reach 95 % accuracy. This fact indicates that, in the signal extracted, on a01 recording, there was a low number VOLUME 7, 2019 131775 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG TABLE 6. Determination of TP, FP and FN based on the detection of R-peak locations in the fECG signals extracted from Physionet Challenge 2013 database using the hybrid ICA-ANFIS-WT method relative to R-peak locations in the reference signal, and calculation of the ACC, SE, PPV and F1 parameters. TABLE 7. Determination of TP, FP and FN based on the detection of R-peak locations in the fECG signals extracted from Physionet Challenge 2013 database using the hybrid ICA-RLS-WT method relative to R-peak locations in the reference signal, and calculation of the ACC, SE, PPV and F1 parameters. of FP values, but a high number of FN values. With the other recordings, this method did not achieve good results, and, thus, it was not, as an extraction method, effective in these recordings. The same assessment was performed for the hybrid ICARLS-WT method and the results can be seen in Table 7. Based on the ACC parameter, this method achieved values above 95 % with 7 recordings (a03, a04, a05, a08, a15, a17 and 131776 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 17. Estimated fHR variable curve over time using the ICA-RLS-WT hybrid method on r10 recording. Graphs a), b) and c) show 3 selected sections where filtration did not work, and graphs d), e) and f) show 3 selected sections where filtration worked properly. a22), values over 90 % for a12 and a24 recordings, and over 80 % for a01, a14, a19 and a25 recordings. With the signal extracted from a12 recording, the method, based on the PPV and F1 parameters, reached a values above 95 %. This fact indicates that, in the signal extracted, on a12 recording, there was a low number of FP values, but a high number of FN values. With the signal extracted from a24 recording, the method, based on the SE, PPV and F1 parameters, reached a values above 95 %. With the signal extracted from a19 recording, the method, based on PPV, reached a value above 95 %. This fact indicates that, in the signal extracted, on a19 recording, there was a low number of FP values, but a high number of FN values. With the other 12 recordings out of 25, the ICA-RLS-WT hybrid method was not effective. During this experiment, the ICA-RLS-WT hybrid method achieved significantly better results than the ICA-ANFISWT hybrid method. It can be seen that, in 9 recordings out of 25, the ICA-RLS-WT hybrid method was able to extract fECG signals that reached at least 90 % based on the ACC parameter, and, as a result, with 13 recordings out of 25, they reached at least 80 % based on the ACC parameter. The ICA-ANFIS-WT hybrid method was able to extract usable fECG signals in only 7 recordings out of 25. V. CONCLUSION AND DISCUSSION The results summarized in the previous subchapters show that the algorithms presented are very effective for most of the recordings tested. However, for some recordings or their sections, their effectiveness is limited. As an example, recording r10 will be discussed in detail in this chapter in connection with the ICA-RLS-WT method. Fig. 17 shows the estimated fHR variable curve using the given method and in comparison with the reference fHR curve obtained from FSE recording. Further, this figure contains six selected 5 s sections, three of which relate to cases where the resulting fHR curve deviated significantly from the reference and, in case of longer duration, this could lead to a false positive diagnosis of fetal distress. The remaining three sections then relate to cases where the specified fHR matches the reference. The analysis of these six sections includes an example of three signals from a given time interval: the reference signal from the FSE, abdominal electrode 1, and, also, the signal estimated using the ICA-RLS-WT method. It is clear from these examples that the quality of the extraction depends on the quality of the input abdominal signals. In the case of sections where the recording quality was inadequate, i.e. a) to c), this algorithm was not able to suppress this interference, which resulted in an increase in the number of false positive peaks (FP), thus reducing the parameters evaluating extraction quality (ACC, PPV, F1). With sections d) - f) having good-quality abdominal recordings, the useless signal was suppressed to such an extent that successful detection of fetal R waves was possible; therefore, the fHR curve determined is consistent with the reference. Hence, VOLUME 7, 2019 131777 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 18. Estimated fHR variable curve over time using the ICA-ANFIS-WT and ICA-RLS-WT hybrid methods on r03 recording. Graphs a) to f) show selected sections for a comparison of the efficiency of both methods. in the case of clinical use of this method, it is essential to maintain a high quality of the abdominal recording. It is negatively influenced in particular by insufficient adhesion of the measuring electrodes and their placement, fetal position and maternal movement. Another factor that has a tremendous impact on the quality of extraction using the hybrid system presented here is the optimal adjustment of adaptive algorithms. As examined in [71], this setting varies depending on the placement of the input electrodes (reference thoracic and abdominal ones). In this case, however, the reference maternal signal is estimated using AES methods. Therefore, in further research, a more detailed study of optimizing the adaptive algorithm settings is required, particularly depending on various factors, such as the selection of the quantity and location of the hybrid system input electrodes, the gestational age of the fetus or its position. At the same time, it is necessary to test and compare combinations of different algorithms in each block of the hybrid extraction system depending on the factors mentioned above. Fig. 18 shows a comparison of fHR variable curves determined using the ICA-ANFIS-WT and ICA-RLS-WT methods tested in r03 recording compared to the reference fHR curve obtained from the FSE recording. It can be stated that the ICA-RLS-WT method is more capable of copying the trend within the resulting fHR curve. The method using ANFISwasnot ableto suppress the maternal component from the abdominal recording in some sections, and the resulting fHR thus correlated with the maternal heart rate that is lower than fetal one. On the other hand, the method using the RLS algorithm, with some exceptions as section e), was able to successfully suppress the maternal component to such an extent that it did not affect the subsequent detection of R-R intervals, hence the calculation of the resulting fHR. Fig. 19 shows the example of the reference output signal and the fHR waveforms along with those estimated by means of hybrid methods ICA-ANFIS-WT and ICA-RLS-WT on a08 recording, where the filtrations achieved 100 % for all the evaluation parameters. The fHR determined using the estimated fECG signal is almost identical with the reference fHR signal. Fig. 20 shows the example of the reference output signal and the fHR waveforms along with those estimated by means of hybrid methods ICA-ANFIS-WT and ICA-RLS-WT on a18 recording, i.e. the recording with the worst results. It can be noticed that in the input aECG signal, the magnitude of the fetal component is negligible in comparison with the maternal one. For such signal, the extraction is challenging. It should be noted that the Physionet Challenge 2013 dataset is composed of variety of data. Part of the data was taken from various fECG databases using various electrode deployment, while some were synthetic. It can 131778 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG FIGURE 19. Estimated fHR variable curve over time using the ICA-ANFIS-WT and ICA-RLS-WT hybrid methods on a08 recording. Graphs a)–f) show selected sections in recording where filtration worked properly. FIGURE 20. Estimated fHR variable curve over time using the ICA-ANFIS-WT and ICA-RLS-WT hybrid methods on a18 recording. Graphs a)–f) show selected sections in recording where filtration did not work. VOLUME 7, 2019 131779 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG be noticed that the signals significantly differ from the real recordings that were used in the rest of the experiments in the time domain and also in terms of the determined fHR. Hybrid methods combine the benefits of both adaptive and non-adaptive approaches, such as accuracy, stability, and minimization of the number of measuring electrodes. This study shows that this approach is very promising for the extraction of fECG and very accurate for the needs of determining fHR as the main parameter observed in current clinical practice. The subject of further research should be deeper morphological analysis, such as ST segment, QT interval or T/QRS ratio analysis. This would lead to the creation of a new diagnostic method - a non-invasive variant of STAN, i.e. NI-STAN, thus improving diagnosis of fetal hypoxia. In this work, wavelet transform was used in the last step. This method is very effective for the purpose of detecting R-peaks, which is important for determining the heart rate. On the other hand, this method has negative effects on the ECG curve morphology. Therefore, for the purposes of NI-STAN, a replacement for the final post-processing step would have to be found in this hybrid system in the future. Nevertheless, it can be stated that the construction of hybrid methods achieves promising efficiency in fECG signal extraction, hence, it can be assumed that these methods could be used in clinical practice as a replacement for classical CTG. However, before introducing this method into clinical practice, the negative factors affecting the efficiency of the extraction system need to be addressed. They include, in particular, the quality of abdominal recordings, which is primarily influenced by the location and adhesion of the measuring electrodes, and the maternal or fetal movement. In this initial study, particular attention was paid to the precise determination of fHR. However, in clinical practice, both fHR and uterine contractions are monitored by CTG during electronic fetal monitoring. In the case of an ECG-based system, the electrical signal produced by the contracting uterus could be recorded by a method known as electrohysterography (EHG), whose efficacy has been verified in e.g. [137]–[140]. For future research, it is necessary to incorporate measurement and filtration of uterine contractions into a complex NI-fECG-based fetal monitoring system. The proposed hybrid system could be improved by modifying its individual parts. The adaptive algorithms could be upgraded or replaced by different adaptive methods. For example, in the future research, we aim to test LMS, normalized LMS, QR-RLS, or adaptive linear neuron (ADALINE). Most of these algorithms are part of MATLAB digital signal processing toolbox. The advantage of using an adaptive algorithm as a major part of a hybrid system is that it compensates delays and the inference. Another part that can be modified is the initial block of the extraction system for estimation of the mECG input to the adaptive algorithm. The currently used ICA algorithm could be replaced by another method able to estimate the maternal component, such as the PCA algorithm. The advantage of using the ICA algorithm (or PCA algorithm) is that it only requires abdominal electrodes as an input (the chest electrodes are not needed). In terms of the WT algorithm, it is necessary to mention that its utilization influences the morphology of the ECG waveform and thus disables the morphological analysis. So, in our future research we will only use WT algorithm to find correct R-peak positions (for fHR monitoring) while the morphological analysis will be performed only on the signals that were not processed by the WT algorithm. The approach introduced in this paper is only an example of possible hybrid fECG extraction systems. There is a large variety of the hybrid systems among the literature [97]. Most of the currently available hybrid methods utilize the ICA algorithm. Very interesting hybrid methods are for example combination of ICA, ensemble empirical mode decomposition (EEMD) and wavelet shrinkage (WS) [98], combination of ICA and adaptive fECG enhancer (AFE) [99], combination of ICA and projective filtering (PF) [100], combination of ICA and PCA [101], combination of singular value decomposition (SVD) and ICA [102], etc. In our future research, we aim to test some of the most promising methods and modify our system according to the results. ETHICS STATEMENT The study protocol was approved by the Ethical Committee of the Silesian Medical University, Katowice, Poland (NN-013-345/02). Subjects read the approved consent form and gave written informed consent to participate in the study. REFERENCES [1] C. Sureau, ‘‘Historical perspectives: Forgotten past, unpredictable future,’’ Baillière’s Clin. Obstetrics Gynaecol., vol. 10, no. 2, pp. 167–184, 1996. doi: 10.1016/S0950-3552(96)80032-0. [2] H. D. Banta and S. B. Thacker, ‘‘Historical controversy in health technology assessment: The case of electronic fetal monitoring,’’ Obstetrical Gynecol. Survey, vol. 56, no. 11, pp. 707–719, 2001. doi: 10.1097/00006254-200111000-00023. [3] C. B. Martin, Jr., ‘‘Electronic fetal monitoring: A brief summary of its development, problems and prospects,’’ Eur. J. Obstetrics Gynecol. Reproductive Biol., vol. 78, no. 2, pp. 133–140, 1998. doi: 10.1016/S0301-2115(98)00059-1. [4] C. A. Menihan and E. Kopel, Electronic Fetal Monitoring: Concepts and Applications. Philadelphia, PA, USA: Lippincott Williams & Wilkins, 2007. [5] V. A. Browne, C. G. Julian, L. Toledo-Jaldin, D. Cioffi-Ragan, E. Vargas, and L. G. Moore, ‘‘Uterine artery blood flow, fetal hypoxia and fetal growth,’’ Philos. Trans. Roy. Soc., Biol. Sci., vol. 370, no. 1663, 2015, Art. no. 20140068. doi: 10.1098/rstb.2014.0068. [6] H. Y. Chen, S. P. Chauhan, C. V. Ananth, A. M. Vintzileos, and A. Z. Abuhamad, ‘‘Electronic fetal heart rate monitoring and its relationship to neonatal and infant mortality in the United States,’’ Amer. J. Obstetrics Gynecol., vol. 204, no. 6, pp. 491.e1–491.e10, 2011. doi: 10.1016/j.ajog.2011.04.024. [7] Z. Alfirevic, D. Devane, G. M. Gyte, and A. Cuthbert, ‘‘Continuous cardiotocography (CTG) as a form of electronic fetal monitoring (EFM) for fetal assessment during labour,’’ Cochrane Database Syst. Rev., no. 2, pp. 1–137, Feb. 2017, Art. no. CD006066. doi: 10.1002/14651858.CD006066.pub3. [8] J. Jezewski, J. Wrobel, and K. Horoba, ‘‘Comparison of Doppler ultrasound and direct electrocardiography acquisition techniques for quantification of fetal heart rate variability,’’ IEEE Trans. Biomed. Eng., vol. 53, no. 5, pp. 855–864, May 2006. doi: 10.1109/TBME.2005.863945. [9] M. F. MacDorman and E. C. W. Gregory, ‘‘Fetal and perinatal mortality: United States, 2013,’’ Nat. Vital Statist. Rep., vol. 64, no. 8, pp. 1–24, 2015. 131780 VOLUME 7, 2019 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG [10] H. D. Banta and S. B. Thacker, ‘‘Costs and benefits of electronic fetal monitoring: A review of the literature,’’ US Dept. Health, Educ., Welfare, Public Health Service, Office Health Res., Statist., Technol., Nat. Center Health Services Res., Hyattsville, MD, USA, ERIC no. ED171429, 1979. [11] P. J. Placek, ‘‘Type of delivery associated with social, demographic, maternal health, infant health and health insurance factors,’’ Nat. Center Health Statist., Hyattsville, MD, USA, Tech. Rep., 1972, pp. 619–624. [12] R. H. Paul, J. R. Huey, Jr., and C. F. Yaeger, ‘‘Clinical fetal monitoring: Its effect on cesarean section rate and perinatal mortality: Fiveyear trends,’’ Postgraduate Med., vol. 61, no. 4, pp. 160–166, 1977. doi: 10.1080/00325481.1977.11714552. [13] D. Petitti, R. O. Olson, and R. L. Williams, ‘‘Cesarean section in California—1960 through 1975,’’ Amer. J. Obstetrics Gynecol., vol. 133, no. 4, pp. 391–397, 1979. doi: 10.1016/0002-9378(79)90058-9. [14] R. L. Williams and W. E. Hawes, ‘‘Cesarean section, fetal monitoring, and perinatal mortality in California,’’ Amer. J. Public Health, vol. 69, no. 9, pp. 864–870, 1979. doi: 10.2105/AJPH.69.9.864. [15] S. M. Taffel, P. J. Placek, M. Moien, and C. L. Kosary, ‘‘1989 U.S. Cesarean section rate steadies—VBAC rate rises to nearly one in five,’’ Birth, vol. 18, no. 2, pp. 73–77, 1991. doi: 10.1111/j.1523536X.1991.tb00063.x. [16] S. B. Thacker, D. F. Stroup, and H. B. Peterson, ‘‘Efficacy and safety of intrapartum electronic fetal monitoring: An update,’’ Obstetrics Gynecol., vol. 86, no. 4, pp. 613–620, 1995. doi: 10.1016/S0029-7844(95)80027-1. [17] E. L. Ryding, K. Wijma, and B. Wijma, ‘‘Psychological impact of emergency cesarean section in comparison with elective cesarean section, instrumental and normal vaginal delivery,’’ J. Psychosomatic Obstetrics Gynecol., vol. 19, no. 3, pp. 135–144, 1998. doi: 10.3109/01674829809025691. [18] E. L. Ryding, B. Wijma, and K. Wijma, ‘‘Posttraumatic stress reactions after emergency cesarean section,’’ Acta Obstetricia Gynecologica Scandinavica, vol. 76, no. 9, pp. 856–861, 1997. doi: 10.3109/00016349709024365. [19] J. E. Swain, E. Tasgin, L. C. Mayes, R. Feldman, R. Todd, and J. F. Leckman, ‘‘Maternal brain response to own baby-cry is affected by cesarean section delivery,’’ J. Child Psychol. Psychiatry, vol. 49, no. 10, pp. 1042–1052, 2008. doi: 10.1111/j.1469-7610.2008.01963.x. [20] S. Petrou, J. Henderson, and C. Glazener, ‘‘Economic aspects of caesarean section and alternative modes of delivery,’’ Best Pract. Res. Clin. Obstetrics Gynaecology, vol. 15, no. 1, pp. 145–163, 2001. doi: 10.1053/beog.2000.0154. [21] J. Henderson, R. McCandlish, L. Kumiega, and S. Petrou, ‘‘Systematic review of economic aspects of alternative modes of delivery,’’ Int. J. Obstetrics Gynaecol., vol. 108, no. 2, pp. 149–157, 2001. [22] M. Cremer, ‘‘Ueber die direkte ableitung der aktionsstrome des menschlichen herzens vom oesophagus und uber das elektrokardiogramm des fotus,’’ Miinchener Medizinische Wochenschrift, vol. 53, no. 17, pp. 811–813, 1906. [23] C. Sureau and R. Foulon, ‘‘Etude de l’activite electrique de l’uterus au cours de la gestation et du travail. . .,’’ Tech. Rep., 1955. [24] S. D. Larks, ‘‘Fetal electrocardiography. The electrical activity of the fetal heart,’’ Amer. J. Med. Sci., vol. 242, no. 3, p. 395, 1961. [25] C. Sureau and R. Trocellier, ‘‘Un problème technique d’électrocardiographie foetale. Note sur l’élimination de l’électrocardiogramme maternel,’’ Gynecology Obstetrics, vol. 60, no. 1, pp. 43–54, 1961. [26] J. B. Roche and E. H. Hon, ‘‘The fetal electrocardiogram: V. Comparison of lead systems,’’ Amer. J. Obstetrics Gynecol., vol. 92, no. 8, pp. 1149–1159, 1965. doi: 10.1016/S0002-9378(15)33099-4. [27] S. Zaunseder, F. Andreotti, M. Cruz, H. Stepan, C. Schmieder, H. Malberg, and N. Wessel, ‘‘Fetal QRS detection by means of Kalman filtering and using the event synchronous canceller,’’ in Proc. 7th Int. Workshop Biosig. Interpretation, Como, Italy, Jul. 2012, pp. 2–4. [28] L. D. Lathauwer, B. D. Moor, and J. Vandewalle, ‘‘Fetal electrocardiogram extraction by blind source subspace separation,’’ IEEE Trans. Biomed. Eng., vol. 47, no. 5, pp. 567–572, May 2000. doi: 10.1109/10.841326. [29] R. Sameni, C. Jutten, and M. B. Shamsollahi, ‘‘What ICA provides for ECG processing: Application to noninvasive fetal ECG extraction,’’ in Proc. IEEE Int. Symp. Signal Process. Inf. Technol., Vancouver, BC, Canada, Aug. 2006, pp. 656–661. doi: 10.1109/ISSPIT.2006.270882. [30] M. Richter, T. Schreiber, and D. T. Kaplan, ‘‘Fetal ECG extraction with nonlinear state-space projections,’’ IEEE Trans. Biomed. Eng., vol. 45, no. 1, pp. 133–137, Jan. 1998. doi: 10.1109/10.650369. [31] M. Kotas, J. Jezewski, T. Kupka, and K. Horoba, ‘‘Detection of low amplitude fetal QRS complexes,’’ in Proc. 30th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Vancouver, BC, Canada, Aug. 2008, pp. 4764–4767. doi: 10.1109/IEMBS.2008.4650278. [32] J. Behar, J. Oster, and G. D. Clifford, ‘‘Combining and benchmarking methods of foetal ECG extraction without maternal or scalp electrode data,’’ Physiol. Meas., vol. 35, no. 8, p. 1569, 2014. doi: 10.1088/09673334/35/8/1569. [33] S. M. M. Martens, C. Rabotti, M. Mischi, and R. J. Sluijter, ‘‘A robust fetal ECG detection method for abdominal recordings,’’ Physiol. Meas., vol. 28, no. 4, pp. 373–388, 2007. doi: 10.1088/0967-3334/28/4/004. [34] P. P. Kanjilal, S. Palit, and G. Saha, ‘‘Fetal ECG extraction from single-channel maternal ECG using singular value decomposition,’’ IEEE Trans. Biomed. Eng., vol. 44, no. 1, pp. 51–59, Jan. 1997. doi: 10.1109/10.553712. [35] J. Behar, ‘‘Extraction of clinical information from the non-invasive fetal electrocardiogram,’’ (2016), arXiv:1606.01093. [Online]. Available: https://arxiv.org/abs/1606.01093 [36] R. Vullings, B. de Vries, and J. W. M. Bergmans, ‘‘An adaptive Kalman filter for ECG signal enhancement,’’ IEEE Trans. Biomed Eng., vol. 58, no. 4, pp. 1094–1103, Apr. 2011. doi: 10.1109/TBME.2010. 2099229. [37] G. Bachmann, L. Narici, and E. Beckenstein, Fourier and Wavelet Analysis. New York, NY, USA: Springer-Verlag, 2012. [38] S. H. Jothi and K. H. Prabha, ‘‘Fetal electrocardiogram extraction using adaptive neuro-fuzzy inference systems and undecimated wavelet transform,’’ IETE J. Res., vol. 58, no. 6, pp. 469–475, 2012. doi: 10.4103/03772063.106753. [39] H. Hassanpour and A. Parsaei, ‘‘Fetal ECG extraction using wavelet transform,’’ in Proc. Int. Conf. Comput. Intell. Modelling Control Automat. Intell. Agents Web Technol. Int. Commerce (CIMCA), Sydney, NSW, Australia, Nov./Dec. 2006, p. 179. doi: 10.1109/CIMCA. 2006.98. [40] P. Kumar, S. K. Sharma, and S. Prasad, ‘‘CAD for detection of fetal electrocardiogram by using wavelets and neuro-fuzzy systems,’’ Int. J. Appl. Eng. Res., vol. 11, no. 4, pp. 2321–2326, 2016. [41] K. Assaleh and H. Al-Nashash, ‘‘A novel technique for the extraction of fetal ECG using polynomial networks,’’ IEEE Trans. Biomed. Eng., vol. 52, no. 6, pp. 1148–1152, Jun. 2005. doi: 10.1109/TBME.2005.844046. [42] S. Kadambe, R. Murray, and G. F. Boudreaux-Bartels, ‘‘Wavelet transform-based QRS complex detector,’’ IEEE Trans. Biomed. Eng., vol. 46, no. 7, pp. 838–848, Jul. 1999. doi: 10.1109/10.771194. [43] R. Almeida, H. Gonçalves, J. Bernardes, and A. P. Rocha, ‘‘Fetal QRS detection and heart rate estimation: A wavelet-based approach,’’ Physiol. Meas., vol. 35, no. 8, pp. 1723–1735, 2014. doi: 10.1088/09673334/35/8/1723. [44] M. G. Jafari and J. A. Chambers, ‘‘Fetal electrocardiogram extraction by sequential source separation in the wavelet domain,’’ IEEE Trans. Biomed. Eng., vol. 52, no. 3, pp. 390–400, Mar. 2005. doi: 10.1109/TBME.2004.842958. [45] E. Castillo, D. P. Morales, G. Botella, A. García, L. Parrilla, and A. J. Palma, ‘‘Efficient wavelet-based ECG processing for single-lead FHR extraction,’’ Digit. Signal Process., vol. 23, no. 6, pp. 1897–1909, 2013. doi: 10.1016/j.dsp.2013.07.010. [46] S. A. Chouakri, F. Bereksi-Reguig, S. Ahmaldi, and O. Fokapu, ‘‘Wavelet denoising of the electrocardiogram signal based on the corrupted noise estimation,’’ in Proc. Comput. Cardiol., Lyon, France, Sep. 2005, pp. 1021–1024. doi: 10.1109/CIC.2005.1588284. [47] C. H. L. Peters, R. Vullings, M. J. Rooijakkers, J. W. M. Bergmans, S. G. Oei, and P. F. F. Wijn, ‘‘A continuous wavelet transform-based method for time-frequency analysis of artefact-corrected heart rate variability data,’’ Physiol. Meas., vol. 32, no. 10, pp. 1517–1527, 2011. doi: 10.1088/0967-3334/32/10/001. [48] A. Khamene and S. Negahdaripour, ‘‘A new method for the extraction of fetal ECG from the composite abdominal signal,’’ IEEE Trans. Biomed. Eng., vol. 47, no. 4, pp. 507–516, Apr. 2000. doi: 10.1109/10.828150. [49] F. Mochimaru, Y. Fujimoto, and Y. Ishikawa, ‘‘Detecting the fetal electrocardiogram by wavelet theory-based methods,’’ Prog. Biomed. Res., vol. 7, pp. 185–193, Sep. 2002. [50] E. C. Karvounis, C. Papaloukas, D. I. Fotiadis, and L. K. Michalis, ‘‘Fetal heart rate extraction from composite maternal ECG using complex continuous wavelet transform,’’ in Proc. Comput. Cardiol., Chicago, IL, USA, Sep. 2004, pp. 737–740. doi: 10.1109/CIC.2004.1443044. VOLUME 7, 2019 131781 R. Jaros et al.: Novel Hybrid Extraction Systems for fHR Variability Monitoring Based on NI-fECG [51] B. Azzerboni, F. La Foresta, N. Mammone, and F. C. Morabito, ‘‘A new approach based on wavelet-ICA algorithms for fetal electrocardiogram extraction,’’ in Proc. ESANN, Bruges, Belgium, Apr. 2005, pp. 193–198. [52] S. Wu, Y. Shen, Z. Zhou, L. Lin, Y. Zeng, and X. Gao, ‘‘Research of fetal ECG extraction using wavelet analysis and adaptive filtering,’’ Comput. Biol. Med., vol. 43, no. 10, pp. 1622–1627, 2013. [53] A. Daamouche, L. Hamami, N. Alajlan, and F. Melgani, ‘‘A wavelet optimization approach for ECG signal classification,’’ Biomed. Signal Process. Control, vol. 7, no. 4, pp. 342–349, Jul. 2012. doi: 10.1016/j.bspc.2011.07.001. [54] V. S. Chourasia and A. K. Mittra, ‘‘Selection of mother wavelet and denoising algorithm for analysis of foetal phonocardiographic signals,’’ J. Med. Eng. Technol., vol. 33, no. 6, pp. 442–448, 2009. doi: 10.1080/03091900902952618. [55] R. Martinek, R. Kahankova, J. Nedoma, M. Fajkus, and K. Cholevova, ‘‘Fetal ECG preprocessing using wavelet transform,’’ in Proc. 10th Int. Conf. Comput. Modeling Simulation, 2018, pp. 39–43. [56] S. Ravindrakumar and K. B. Raja, ‘‘Fetal ECG extraction and enhancement in prenatal monitoring—Review and implementation issues,’’ in Proc. Trendz Inf. Sci. Comput. (TISC), Chennai, India, Dec. 2010, pp. 16–20. doi: 10.1109/TISC.2010.5714599. [57] V. Zarzoso and A. K. Nandi, ‘‘Noninvasive fetal electrocardiogram extraction: Blind separation versus adaptive noise cancellation,’’ IEEE Trans. Biomed. Eng., vol. 48, no. 1, pp. 12–18, Jan. 2001. doi: 10.1109/10.900244. [58] R. Martinek, R. Kahankova, J. Jezewski, R. Jaros, J. Mohylova, M. Fajkus, J. Nedoma, P. Janku, and H. Nazeran, ‘‘Comparative effectiveness of ICA and PCA in extraction of fetal ECG from abdominal signals: Toward non-invasive fetal monitoring,’’ Frontiers Physiol., vol. 9, p. 648, May 2018. doi: 10.3389/fphys.2018.00648. [59] R. Sameni, C. Jutten, and M. B. Shamsollahi, ‘‘Multichannel electrocardiogram decomposition using periodic component analysis,’’ IEEE Trans. Biomed. Eng., vol. 55, no. 8, pp. 1935–1940, Aug. 2008. doi: 10.1109/TBME.2008.919714. [60] S. Kharabian, M. B. Shamsollahi, and R. Sameni, ‘‘Fetal R-wave detection from multichannel abdominal ECG recordings in low SNR,’’ in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Minneapolis, MN, USA, Sep. 2009, pp. 344–347. doi: 10.1109/IEMBS.2009.5333578. [61] J. L. Camargo-Olivares, R. Martin-Clemente, S. Hornillo-Mellado, M. M. Elena, and I. Roman, ‘‘The maternal abdominal ECG as input to MICA in the fetal ECG extraction problem,’’ IEEE Signal Process. Lett., vol. 18, no. 3, pp. 161–164, Mar. 2011. doi: 10.1109/LSP.2011.2104415. [62] Y. Sevim and A. Atasoy, ‘‘Performance evaluation of nonparametric ICA algorithm for fetal ECG extraction,’’ Turkish J. Elect. Eng. Comput. Sci., vol. 19, no. 4, pp. 657–666, 2011. doi: 10.3906/elk-0912-311. [63] D. E. Marossero, D. Erdogmus, N. Euliano, J. C. Principe, and K. E. Hild, ‘‘Independent components analysis for fetal electrocardiogram extraction: A case for the data efficient mermaid algorithm,’’ in Proc. IEEE 13th Workshop Neural Netw. Signal Process., Toulouse, France, Sep. 2003, pp. 399–408. doi: 10.1109/NNSP.2003.1318039. [64] Y. Ye, Z.-L. Zhang, J. Zeng, and L. Peng, ‘‘A fast and adaptive ICA algorithm with its application to fetal electrocardiogram extraction,’’ Appl. Math. Comput., vol. 205, no. 2, pp. 799–806, 2008. doi: 10.1016/j.amc.2008.05.117. [65] V. Zarzoso, A. K. Nandi, and E. Bacharakis, ‘‘Maternal and foetal ECG separation using blind source separation methods,’’ Math. Med. Biol., J. IMA, vol. 14, no. 3, pp. 207–225, Sep. 1997. doi: 10.1093/imammb/14.3.207. [66] R. Swarnalatha and D. V. Prasad, ‘‘A novel technique for extraction of FECG using multi stage adaptive filtering,’’ J. Appl. Sci., vol. 10, no. 4, pp. 319–324, 2010. [67] R. Kahankova, J. Jezewski, J. Nedoma, M. Jezewski, M. Fajkus, A. Kawala-Janik, H. Wen, and R. Martinek, ‘‘Influence of gestation age on the performance of adaptive systems for fetal ECG extraction,’’ Adv. Electr. Electron. Eng., vol. 15, no. 3, pp. 491–501, 2017. doi: 10.15598/aeee.v15i3.2207. [68] G. Camps, M. Martinez, and E. Soria, ‘‘Fetal ECG extraction using an FIR neural network,’’ in Proc. Comput. Cardiol., Rotterdam, The Netherlands, vol. 28, Sep. 2001, pp. 249–252. doi: 10.1109/CIC.2001.977639. [69] R. Kahankova, R. Martinek, and P. Bilik, ‘‘Non-invasive fetal ECG extraction from maternal abdominal ECG using LMS and RLS adaptive algorithms,’’ in Proc. Int. Afro-Eur. Conf. Ind. Adv., Marrakesh, Morrocco, Nov. 2016, pp. 258–271. doi: 10.1007/978-3-319-60834-1_27. [70] R. Martinek and J. Žídek, ‘‘Refining the diagnostic quality of the abdominal fetal electrocardiogram using the techniques of artificial intelligence,’’ VSB–Tech. Univ. Ostrava, Ostrava, Czech Republic, Tech. Rep., 2012. [71] R. Martinek, R. Kahankova, H. Nazeran, J. Konecny, J. Jezewski, P. Janku, P. Bilik, J. Zidek, J. Nedoma, and 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, vol. 17, no. 5, p. 1154, 2017. doi: 10.3390/s17051154. [72] M.B.I.ReazandL.S.Wei, ‘‘Adaptive linear neural network filter for fetal ECG extraction,’’ in Proc. Int. Conf. Intell. Sens. Inf. Process., Chennai, India, Jan. 2004, pp. 321–324. doi: 10.1109/ICISIP.2004.1287675. [73] W. Jia, C. Yang, G. Zhong, M. Zhou, and S. Wu, ‘‘Fetal ECG extraction based on adaptive linear neural network,’’ in Proc. 3rd Int. Conf. Biomed. Eng. Informat., Yantai, China, vol. 2, Oct. 2010, pp. 899–902. doi: 10.1109/BMEI.2010.5639886. [74] M. A. Hasan, M. I. Ibrahimy, and M. B. I. Reaz, ‘‘Fetal ECG extraction from maternal abdominal ECG using neural network,’’ J. Softw. Eng. Appl., vol. 2, no. 5, p. 330, 2009. [75] R. Kahankova, R. Martinek, M. Mikolášová, and R. Jaroš, ‘‘Adaptive linear neuron for fetal electrocardiogram extraction,’’ in Proc. IEEE 20th Int. Conf. e-Health Netw., Appl. Services (Healthcom), Ostrava, Czech Republic, Sep. 2018, pp. 1–5. doi: 10.1109/HealthCom.2018.8531135. [76] W. Zheng, H. Liu, A. He, X. Ning, and J. Cheng, ‘‘Single-lead fetal electrocardiogram estimation by means of combining R-peak detection, resampling and comb filter,’’ Med. Eng. Phys., vol. 32, no. 7, pp. 708–719, 2010. doi: 10.1016/j.medengphy.2010.04.012. [77] Z. Wei, W. Xueyun, Z. J. jian, and L. Hongxing, ‘‘Noninvasive fetal ECG estimation using adaptive comb filter,’’ Comput. Methods Programs Biomed., vol. 112, no. 1, pp. 125–134, 2013. doi: 10.1016/j.cmpb.2013.07.015. [78] M. Shadaydeh, Y. Xiao, and R. K. Ward, ‘‘Extraction of fetal ECG using adaptive volterra filters,’’ in Proc. 16th Eur. Signal Process. Conf., Lausanne, Switzerland, Aug. 2008, pp. 1–5. [79] P. Wang and S. Tan, ‘‘Soft computing and fuzzy logic,’’ Soft Comput., vol. 1, no. 1, pp. 35–41, 1997. doi: 10.1007/s005000050004. [80] M. Nasiri, K. Faez, and A. M. Nasrabadi, ‘‘A new method for extraction of fetal electrocardiogram signal based on adaptive nero-fuzzy inference system,’’ in Proc. IEEE Int. Conf. Signal Image Process. Appl. (ICSIPA), Kuala Lumpur, Malaysia, Nov. 2011, pp. 456–461. doi: 10.1109/ICSIPA.2011.6144151. [81] A. Sargolzaei, K. Faez, and S. Sargolzaei, ‘‘A new method for foetal electrocardiogram extraction using adaptive nero-fuzzy interference system trained with PSO algorithm,’’ in Proc. IEEE Int. Conf. Electro/Inf. Technol., Mankato, MN, USA, May 2011, pp. 1–5. doi: 10.1109/EIT.2011.5978624. [82] S. S. Priyadharsini, S. E. Rajan, and S. Saranya, ‘‘An efficient softcomputing technique for extracting fetal ECG from maternal ECG signal,’’ Program. Device Circuits Syst., vol. 3, no. 1, pp. 1–7, 2011. [83] S. R. Rathod and V. R. Bansod, ‘‘Separation of FECG from complex ECG in fetal monitoring,’’ in Proc. Online Int. Conf. Green Eng. Technol. (IC-GET), Coimbatore, India, Nov. 2016, pp. 1–5. doi: 10.1109/GET.2016.7916788. [84] R. Swarnalatha and D. V. Prasad, ‘‘Maternal ECG cancellation in abdominal signal using ANFIS and wavelets,’’ Appl Sci., vol. 10, pp. 868–877, Nov. 2010. [85] H. Ahmadieh and B. M. Asl, ‘‘Fetal ECG extraction via type-2 adaptive neuro-fuzzy inference systems,’’ Comput. Methods Programs Biomed., vol. 142, pp. 101–108, Apr. 2017. doi: 10.1016/j.cmpb.2017.02.009. [86] P. Kumar, S. K. Sharma, and S. Prasad, ‘‘Detection of fetal electrocardiogram through OFDM, neuro-fuzzy logic and wavelets systems for telemetry,’’ in Proc. 10th Int. Conf. Intell. Syst. Control (ISCO), Coimbatore, India, Jan. 2016, pp. 1–4. doi: 10.1109/ISCO.2016.7726970. [87] K. Assaleh, ‘‘Extraction of fetal electrocardiogram using adaptive neurofuzzy inference systems,’’ IEEE Trans. Biomed. Eng., vol. 54, no. 1, pp. 59–68, Jan. 2007. doi: 10.1109/TBME.2006.883728. [88] R. Martinek, H. Skutova, R. Kahankova, P. Koudelka, P. Bilik, and J. Koziorek, ‘‘Fetal ECG extraction based on adaptive neuro-fuzzy interference system,’’ in Proc. 10th Int. Symp. Commun. Syst., Netw. Digit. Signal Process. (CSNDSP), Prague, Czech Republic, Jul. 2016, pp. 1–6. doi: 10.1109/CSNDSP.2016.7573973. [89] J. Behar, A. Johnson, G. D. Clifford, and J. Oster, ‘‘A comparison of single channel fetal ECG extraction methods,’’ Ann. Biomed. Eng., vol. 42, no. 6, pp. 1340–1353, 2014. doi: 10.1007/s10439-014-0993-9. 131782 VOLUME 7, 2019