Fetal ECG signal processing and identification of hypoxic pregnancy conditions in-utero

1.   Introduction

Fetal heart activity tracking and wellbeing investigation is an essential part of prenatal care. Clinical fetal monitoring based on both antepartum fHR and fECG shows that most of the cardiac defects, as well as oxygen deprivation, have some appearance in the fECG morphology. In this sense, electrocardiography could be seen as one of the most important reference sources about fetal status during pregnancy and at labour with a high predictive value. This information can be exploited to diagnose fetal hypoxia and indicate changes in fetal heart patterns. Fetal heart monitoring serves the purpose of detecting signs of life in utero, including abnormal fHR or patterns, which may indicate oxygen deficiency or other pathological processes at their early stage to perform the appropriate intervention and prevent intrauterine fetal neurological damage and mortality. Besides other metabolic changes, fetal response to oxygen deprivation includes decreased heart rate, cessation of gross body movements and a reduction in oxygen consumption ^[1,2]. These changes enable to use of fHR monitoring as a potentially valuable approach for assessing hypoxaemic fetus and their oxygenation status in real-time. In addition, accurate identification of fetuses with sufficient heart rate and oxygenation can prevent unnecessary intervention; reduce the fetal morbidity and number of operative deliveries ^[3].

Although the fHR changes markedly during uninterrupted hypoxia, fetal deprivation or non-reassuring fetal status can be observed only indirectly. To date, the majority of fECG studies on long‐term electronic fetal monitoring are based on data acquired on a clinical basis which requires costly equipment, extensive training, a high level of technical skill, and is subject to high variability in data interpretation. To overcome these challenges, significant efforts are being made for both remote solutions for home fHR monitoring and signal processing technique. While the current standards of fetal monitoring do not support remote or at-home monitoring and require a high skilled physician to interpret results, research community and medical systems developers tend to capacity building and development of non-invasive, portable devices feasible to telemedical home care and long-term monitoring of fHR and fECG ^[4,5,6]. This is particularly relevant for developing countries and rural areas where access to prenatal care can be challenging due to different reasons, including physician scarcity, unavailability of equipment for accurate record keeping and peculiarities of rural geography. Moreover, remote, outpatient fHR monitoring provides the means to collect and analyze large ECG data, enabling pregnant women to receive timely prenatal and antenatal care, which is crucial in high-risk pregnancies.

A new generation of perinatal care instruments offers the capability to deliver an undistorted fECG output from electrodes placed on the maternal abdomen, perform advanced morphological fECG signal extraction and analysis, and send raw and proceeded data directly to the physician. The idea is to develop a tool being intrinsically ambulatory, could also to perform signal recording and processing in the home environment on a long‐term basis ^[7] and offering the capability to carry out advanced morphological fECG signal analysis ^[8]. Improvement of technology for fHR monitoring is promising; however, it has not yet reached a sufficient level of accuracy to be approved for commercial devices ^[9]. A limitation of this approach is that it utilizes a just noticeable, very low signal to noise ratio (SNR). Most notably, this is because the fECG signal is produced by the fetal heart that is a relatively small source compared to the mother's body. In particular, the size of the maternal electrocardiogram (mECG) found in the abdominal cavity is approximately 2–10 times the size of the fECG ^[10], which also makes it difficult to extract the fECG. Moreover, received abdominal signal includes not only mECG and weak signals of fECG but also other signals as maternal respiration, the electrohysterogram (EHG), the power line interference, and other noise coming mainly from maternal muscle activity. Thus, the suppression of mECG while preserving the fetal QRS complex is the most important stage in the fECG extraction from the maternal abdominal signal ^[11]. Further development and evaluation of these technologies remain a global priority.

The present study is a continuation of the previous work ^[12] aimed at discovering interesting associations in gestation course data and finding risk factors in pregnancy-related to fetal hypoxia. Our current goal is to automate the in-utero fetal hypoxia risk factors detection to achieve early warning on fetus' safety and provide the physician with additional information about the potential threat, if any. We investigate the possibilities of mECG elimination and fECG extraction from signal recorded non-invasively on the mother's abdomen and assess the feasibility of the fECG morphological evaluation to be included in the automatic screening tool for identifying fetuses at high risk of hypoxia. To be adopted with this process, we propose a novel methodology for automatic fECG signal processing and tune it into a fetal hypoxia risk-scoring tool. To our knowledge, this is the first trial to introduce auto-generated risk scoring in fECG. It is assumed that risk evaluation allows early warning on fetus' non-reassuring fetal status and provides the physician with additional information about the possible fetal complications.

2.   State of the art

Several surveys and systematic reviews of mECG elimination and fECG detection were found in the literature. In ^[13], the most common techniques developed over the last four decades with discussion both their advantages and shortcomings are given. Another review ^[14] is devoted to fECG signal processing and mECG cancelling methods based on the power line interference component (PLI). In ^[15], the system for indirect fECG was proposed and evaluated with direct fECG, which currently is the gold standard of fECG.

A comparative analysis of single-channel fECG extraction methods is given in ^[16]. In the vast majority of cases, fECG is obtained from multichannel devices by evaluating independent sources of fetal heart bioelectrical activity ^[17]. While extracting unobservable signals, the sources are assumed to be statistically uncorrelated to each other and the known mixture of these signals ^[18]. In many practical applications, the multichannel fECG extraction is based on Blind Source Separation (BSS) and its variations as BSS with Reference (BSSR) ^[19] and others. The study on this topic based on a solid theoretical approach known as Independent Component Analysis (ICA) ^[20] which requires the use of Higher-Order Statistics (HOS). However, for some cases as Gaussian sources, HOS methods do not allow to perform the separation. An alternative approach for source separation is second-order statistics (SOS). SOS-based methods include ​​Singular Value Decomposition (SVD) ^[21], Principal Component Analysis (PCA) ^[22,23], Periodic Component Analysis (πCA) ^[24], etc. Besides, many other methods have been applied for the same purpose, such as Extended Barros's extraction algorithm and Zhang algorithm (ZA) ^[25], polynomial eigenvalue decomposition (PEVD) ^[26], generalized eigenvalue decomposition (GEVD) ^[27], fuzzy clustering c-means (FCM) ^[28], Compressed Perception (CS), parental suppression method (MCSM), Hilbert Huang Transform (HHT) ^[29], and others. Figure 1 shows a comparison of the most popular multichannel methods for fECG signal separation. Following the idea discussed in ^[30], ten approaches were compared according to their overall performance, SNR improvement level, computational cost, and implementation complexity. We used blue lines for real-time approaches and red lines for non-real-time approaches.

The ICA can be seen as a baseline technique since this is the most popular non-adaptive method for fECG extraction and has the average characteristics across the board. There are many modifications of the ICA technique discussed in the literature, including the Fast ICA algorithm (FastICA) ^[31], Joint Approximate Diagonalization of Eigen matrices algorithm (JADE) ^[32]. In ^[33], three-lead abdominal ECG (aECG) signals are collected and sent to the smartphone application, where the signal is processed using the FastICA algorithm to extract the signal components. The sample's entropy is used to determine the fECG signal and calculate the fHR. In ^[34], the Space-Time ICA (ST-ICA) was implemented by combining the delays with FastICA. The preliminary results indicate that the temporal dynamics and the spatial information used by ST-FastICA enable a robust separation of ECGs in low-dimensional recordings where the traditional implementation of FastICA is more likely to fail. The SVD is a spatial filtering approach that seeks a linear decomposition of a data matrix, which acts as a spatial filter and efficiently removes electrocardiographic artefacts ^[17]. PCA is a method that is often used to reduce the dimensionality of large data sets. PCA might also be applicable to find a linear combination of vectors and separating mixed signals from statistically independent sources. Transforming the data into a new orthogonal coordinate system enables us to find the most significant vectors, and this operation increase interpretability while minimizing the loss of information ^[35]. The polynomial matrix eigenvalue decomposition (PMED) method is an extension of EVD in the case of the para-Hermitian matrix. It represents polynomially (or convolutively) mixed signals in the diagonal matrix form. This transformation provides both improved signal and noise identification and strong decorrelation.

The subspace decomposition enables signal separation and estimation of the fECG signal. PMED is a blind approach that does not require a priori knowledge of the source signals or the mixing matrix; instead, it utilizes the space-time covariance matrix's information carried by the observed sensor signals. Assuming that source signals to be statistically independent, the mECG amplitude can be compared to fECG to separate from aECG ^[26]. PMED does not require mECG as a reference and shows relatively low sensitivity to sensor placement.

An analysis of mECG suppression techniques shows that more accurate results are obtained mainly with hybrid non-adaptive methods ^[36,37], ICA & PF ^[38]. The methods' performance highly depends on the tasks for which they are applied (filtering, correlation, transformation, separation, etc.), which also explains the advantage of hybrid methods. In many cases, non-adaptive methods are used for preprocessing, and further analysis is performed with adaptive algorithms. Applying non-adaptive techniques enables partially separate source components and facilitate retrieval for an adaptive system. By this reasoning, we use a hybrid approach for fECG extraction and fetal R-peaks detection, which incorporates the ECG signal's periodic nature and shows the accuracy of fECG extraction sufficient for its analysis. At the stage of the fECG extraction, the following techniques have been applied: PCA to determine the correlation between layering, JADE for blind signal separation, GEVD and πCA for separation of the mECG and fECG.

3.   Materials and methods

Proposed fECG signal processing and hypoxia risk estimation pipeline comprises the following stages: 1) fECG extraction and separation, 2) fHR and fHR variability (fHRV) calculation, 3) hypoxia index (HI) evaluation and risk estimation. We detail the basic mathematical concepts and procedures step by step below.

3.1.   Fetal ECG extraction and separation

The data model of signals received from the maternal abdominal surface can be represented as follows ^[39]

where $x\left(t\right)\in {\mathbb{R}}^{N}$ denotes the N-channel measurements from the surface of the maternal abdomen and thoracic, ${s}_{m}\left(t\right)\in {\mathbb{R}}^{M}$ describes the mECG entry, ${s}_{f}\left(t\right)\in {\mathbb{R}}^{L}$ are the fECG components, $v\left(t\right)\in {\mathbb{R}}^{K}$ and $n\left(t\right)\in {\mathbb{R}}^{N}$ denotes the correlated (low-rank) noise and uncorrelated (full-rank) noise respectively; ${H}_{m}\in {\mathbb{R}}^{N\times M}$ and ${H}_{f}\in {\mathbb{R}}^{N\times L}$ matrices describe the signals received from a set of electrodes placed on the surface of the maternal body. Eq (1) in a compact form:

here matrix $H\in {\mathbb{R}}^{N\times (M\times L\times K)}$ is a general mixing matrix through the sources and $s\left(t\right)\in {\mathbb{R}}^{M\times L\times K}$ represents all sources and structural noise jointly. When the general mixing matrix H is non-singular and the number of monitored channels N is equal to or greater than the effective number of M, L, K sources ($N\ge M+L+K)$, the observed mixture can be treated as defined or redefined. Thus, ${s}_{m}\left(t\right), {s}_{f}\left(t\right)$ and $v\left(t\right)$ are considered as groups of mutually independent sources with some internal correlations.

The fECG extraction procedure is summarized in Figure 2, with key elements described in detail below. In a generic way, the procedure comprises of six stages: 1) signal preprocessing, 2) mECG R-peak detection, 3) sources separation, 4) mECG cancellation, 5) fECG R-peak detection, 6) fHR analysis for detection of fetal hypoxia. At the first stage, six maternal ECG signals (five abdominal and one thoracic) are preprocessed. The key operations include removing the baseline wander and cancelling higher frequency content with a second-order zero-phase low pass filter. On stage two, PCA decorrelation is applied. The mECG R-peak detection occurs on each pre-filtered channel, including PCA-transformed one.

As a result, one of the mECG R-peak channels is selected as being the reference component. The third stage runs the source separation algorithm on the aECG data with the purpose of extract the fECG. On stage four, the mECG channels are cancelled using generalized eigenvalue decomposition and periodic component analysis. Stage five assumes fECG R-peak detection is performed based on the residuals containing the fECG through the fECG enhancing. Finally, a channel for fECG R-peak detection is selected. Finally, extracted fECG R-peak and corresponding fHR are scored for detection of fetal hypoxia. Below all the steps are presented and explained in more detail.

3.1.1.   Signal preprocessing

As it was mentioned above, readings obtained from the maternal surface include three types of data, 1) mix of low-amplitude fECG signal together with mECG signal, 2) high-amplitude baseline wander and 3) noise. The preprocessing stage involves the techniques for eliminating low frequency (below 0.05 Hz) baseline wander and removing high-frequency (above 50 Hz) noise from each channel. For this purpose, the second-order zero-phase low pass filter is used.

3.1.2.   The mECG R-peak detection

The mECG R-peak is detectable from an independent maternal thoracic lead. The principal component analysis helps to obtain the maternal component for R-peak detection. To increase the mECG R-peak contribution on the principal components, the mECG R-peak detection is executed per both pre-filtered channels and channel after PCA transformation. As a result, one of the mECG R-peak channels is selected as being the reference channel.

3.1.3.   Source separation

The source separation is carried out on the aECG to extract the fECG. With respect to the data model (1), this step assumes a specific separation of fECG from the structured noise v(t). From our previous analysis, BSS can be used to take advantage of the inter-channel spatial correlation of fECG sources, slightly improving the quality of fECG before detecting R-peaks. Namely, the JADE algorithm ^[40] is used. The number of extracted components is the same as the number of maternal channels recorded from the surface of the abdomen and chest of a pregnant woman.

It should also be mentioned that due to the peculiarities of the implementation of a blind source separation algorithm, the order of the putative sources (in our case fECG) could not be guaranteed. This requires additional visual inspection of each fECG channel after the JADE has been applied.

3.1.4.   The mECG cancellation

Whereas the mECG is considered as the main source of interference in fECG, iterative subspace decomposition is applied to eliminate maternal components from the recordings. As a basis for our method, we used the approach proposed in ^[41] and modified it as follows.

First, we run GEVD to address the challenge with source separation. The GEVD of a pair of matrices (A, B) generalizes the EVD of a single matrix and produces the new matrices W and D in such way that

where A, B, D, W, I $\in {\mathbb{R}}^{N\times N}$ and A, B are initial symmetric matrices to be diagonalized, D denotes the generalized diagonal matrix of eigenvalues ​​correspond to the generalized eigenvalue W with a set of real eigenvalues ​​sorted in ascending order along its diagonal, and I is the identity matrix.

Using GEVD, the pair of matrices are computed over the entire data set. The efficiency of linear decomposition with GEVD vary with the utilization of preliminary signal and noise subspace information collected in matrix pairs A and B.

Next, a linear source separation on πCA ^[42] was applied to the new matrix pairs. The periodic component analysis utilizes the concept of cardiac phase to sequentially change the ECG rhythm on a controlled rhythm. The algorithm calculates the timespan τt for which the ECG signal x(t) maximally correlate with x(t + τt). The πCA is purely compatible with the pseudo-periodic structure underlying the maternal and fetal signals and can be evaluated using the cost function as follows

here τ_t denotes the interested timespan for the corresponding ECG, C_τ refers to the lagged covariance matrix calculated over the timespan τ_t and W is a matrix resulting from GEVD of diagonalized matrices C_x and C_τ.

The method reapplies a linear decomposition sequence to autonomously separate the mECG and fECG subsets, followed by noise removal for suppressing the mECG components, and finally, the background projection of the disabled elements into the input data space. This is an iterative procedure repeatedly carrying out a process until most of the maternal components removed from the fECG. In this regard, the preprocessing stage where simple R-peak detection is performed enables to detect the location of the maternal R-peaks. This information is passed to the mECG cancellation module, where fine-grained separation and sorting of the mECG and fECG components formed with respect to the degree of similarity across their periodic properties and the frequency of the mECG signal.

3.1.5.   The fECG R-peak detection

At this step, fECG is enhanced using the FastICA algorithm as an additional filter and a channel for fECG R-peak detection is selected. The fECG R-peaks detection carried out via a dedicated fECG channel using the following procedure. The selected fECG component is run through a matched filter with a narrow fECG R-peak amplitude for maximizing its impulse response. The amplitude is specified further to visual inspection of the output. Then the results are squared and time-averaged through a moving average filter. The resulting local peaks can be thought-out as the fetal R-peaks.

3.2.   Fetal ECG parameters detection

The fHR and fHRV are the primary sources of information about fetal well-being. Instantaneous fHR in beats per minute (bpm) is calculated for each cardiac cycle T (in milliseconds) as stated by Eq (5) ^[43].

Following recommendations of the heart variability standard ^[44], time and frequency analysis of the fHR is calculated on 5-minute segment blocks.

For the fHR values s[n] (n = 1, ..., N), also denoted in a vector form as $\overline{s}$, the fHR baseline variability (fHRV) is considered as the sign of severe fetal brain damage ^[45]. The fHRV is represented in the time domain as follows

3.3.   Hypoxia index (HI) evaluation and risk estimation

Considering the fact that intrapartum fECG could be used as a screening method for detecting hypoxia ^[46], it would be safe to assume that hypoxia risk estimation is achieved through the hypoxia index (HI) proposed in ^[47], at least for the labor stage. The threshold HI for a set of fHR values s[n] is the sum of all deceleration periods (min) divided by the lowest fHR (bpm) in the decelerations, multiplied by 100. The resulting value indicates the intensity of hypoxia in the intrapartum fetal heart rate monitoring.

Denoting one period of deceleration of fHR Tdec, and the lowest value of fHR as fHRmin_dec, the threshold hypoxia index HI can be expressed as follows

The fetal hypoxia risk assessment is performed in accordance with ^[48], where the value of the Hi = 25 is a threshold level, for values greater than 25, the risk of fetal hypoxia shall be considered as high. With a calculated hypoxia index of Hi ≤ 24, the risk of fetal hypoxia is considered low.

Additional information for detecting fetal distress can be obtained by a more profound analysis of the morphology of the fECG signal, e.g., the ratio between the amplitude of the fetal QRS complex and that of the T wave, ST-segment changes, etc., to be included in the risk scoring tool.

4.   Experiment and results

We tested the implementation of the first stage of the fECG extraction for automatic processing and analysis of ECG and risk assessment of hypoxia. The experiment was performed using the Open-source electrophysiological instrument set (OSET) tools ^[49]. As a result, the fECG was isolated, and R-peaks were determined.

4.1.   Data description

For this experiment, the dataset ^[50] from PhysioNet ^[51] were used. Non-invasive fetal ECG signals were collected using routine abdominal electrocardiography. The fECG data came from the automated Cardiolab Babycard fetal monitor. The data consist of records in which fetal arrhythmias are diagnosed by echocardiography and equal control records in which normal rhythm is diagnosed. Records are annotated as follows: ARR - fetal arrhythmia, NR - normal fetal rhythm. The experiment used 10-second fragments of records NR_02 and ARR_01.

Each record is with a record of four or five abdominal channels-abdominal ECG (aECG): Abdomens_1–5 and one thoracic lead, which corresponds to the selected ECG of the mother-maternal ECG (mECG). The experiment performed 10-second fragments of records ARR_01 and NR_02. The sampling frequency of signals is 1000 Hz. The fragments of records ARR_01 and NR_02 are as follows (see Figures 3 and 4). In Figures 3 and 4, the upper signal corresponds to the thoracic lead, i.e., the mECG, the five lower signals correspond to the aECG.

Raw signals from abdominal channels contain not only the fECG but also primarily the mECG and various types of noise (e.g., maternal muscles activity or artifacts resulting from fetal movements).

From Figures 3 and 4, the initial raw data under identical conditions, however, are visually different. This may be due to even the slightest difference in the location of the electrodes, the position of the fetus, the different ADC noises and many other factors that affect the ECG properties of the fetus discussed above.

4.2.   fECG extraction

Firstly, the preprocessing step was conducted by means of a second-order zero-phase lowpass filter with the parameters described above. The filter removes data from the baseline wander and cuts-off high-frequency values equal to 50 and 60 Hz. The data ARR_01 and NR_02 after the preprocessing are presented in Figures 5 and 6. It can be observed that, unlike the raw data, there is no baseline wander in the processed.

The preprocessing parameters are defined in such a way as to remove the most of noise while preserving the elements of the fetal and maternal QRS complexes. Signal processing with a low pass filter enables filtering out the dc potentials that trigger the baseline wander.

At the next step, the principal component analysis (PCA) was applied for maternal R-peak detection. In general, PCA enables to reduce the data dimensionality while preserving their informative value. With fECG extraction, isolating the main components of the fetal ECG signal is complicated by the fact that even after filtering, the data contains noises that, along with the fetal and maternal ECG signals, will affect the dimension of the extracted components. Data of ARR_01 and NR_02 after the principal component analysis are presented as follow (see Figures 7 and 8).

In this matter, PCA is, in a way, signal preprocessing before extracting independent components using JADE. Following the overall procedure, the mECG was used for blind source separation using the JADE algorithm. The results of source separation of ARR_01 and NR_02 are presented in Figures 9 and 10.

For ARR_01, 2 out of 6 extracted components clearly correspond to the maternal ECG (1 and 5), two components correspond to the fetal heart (4, 6), and two components comprise the noise (2, 3). Based on the results of source separation, channel 4 was selected for fECG extraction for ARR_01, and channel 1 was selected for more accurate mECG extraction.

For NR_02, 2 out of 6 extracted components clearly correspond to the maternal ECG (1 and 6), two components correspond to the fetal heart (2, 3), and two components comprise to the noise (4, 5). Based on source separation results, channel 2 was selected for fECG extraction and channel 1 for more accurate mECG extraction.

After mECG cancellation, signals be subject to the GEVD and periodic component analysis for fECGs extraction. In ARR_01, the R-peaks of the selected independent fetal component (4) are used to calculate the fetal ECG period. The results of the periodic component analysis for ARR_01 are shown in Figure 11. As seen, the first component (corresponding to the largest eigenvalue) bears the most significant similarity to the fetal ECG, while as the eigenvalues decrease (from the first to the last component), the signals become less similar to the fECG.

Same as previous, the R-peaks of the independent fetal component (2) are used to calculate the fetal ECG period for NR_02. The first component in resulting periodic components for NR_02 has the most remarkable similarity to the fetal ECG.

Finally, R-peaks were detected in the fECGs; the result for ARR_01 and NR_02 is shown in Figures 13 and 14. The top row of each figure depicts fECG with marked R-peaks, and the bottom row shows the time between two sequential R-peaks (RR intervals).

The fECG looks still noisy though R peaks are clearly demarcated, and P wave can be traced. Compared to NR_02, the R peaks in ARR_01 (bottom line) appear at irregular time intervals, and visually the P wave is unseen. However, the signs of irregularity occur between 6 and 7.5 seconds in NR_02. It can be considered as model perturbation and needs further investigation. Prolonged skewness and asymmetry of the RR-intervals can be used as a sign of arrhythmia or fetal heart failure and can potentially be used as a statistical indicator of the autonomic response to fetal hypoxia.

In practice, only up to 30% of abdominal ECG recordings allow the R-wave of the fetus to be clearly seen. When the fetus is in a cephalic presentation, the R waves on the ECG of the mother and the fetus are in different planes (have opposite signs), so the fetal signals can be identified relatively easily. In the case of co-directional R-waves in the mother and fetus, the fetal signals can be identified by averaging them after the maternal signal has been removed. In some cases, the averaging every 4 beats enable to obtain a smooth shape of the beats ^[52].

5.   Discussion

For evaluation resulting fECG and its morphological elements, such as QRS-complex, P-wave, T-wave two statistical estimators, the Mean Squared Error (MSE) between raw and extracted ECG channels, and signal-to-noise ratio (SNR) between an original signal and the extracted component were used. The sensitivity of the selected components has been assessed.

MSE is the most widely used and also the simplest complete reference metric, which is calculated by squaring the differences between raw and extracted fECG signal components and averaging them with the SNR of the related quantity. We used this statistic to assess the accuracy of the algorithm in terms of evaluating the fHR from their R-peaks. MSE is defined as ^[53].

MSE metric was used to estimate the differences between matched values and computed fHR at twelve samples per one-minute segment, with, where denotes the assessed test fHR value on i-th segment and that denotes the referent fHR value for the same segment.

Finally, the SNR variation in decibels (dB) was calculated to find the relationship between the desired fHR signal and extracted component ^[54].

here r denotes a reference signal, and f is an extracted signal with K components.

The normalized MSE for ARR_01 and NR_02 between raw and extracted ECG channels are shown in Figure 15.

The MSE in dB for the overall assessment of the isolated 5-lead fECG signal for ARR_01 and NR_02 records is presented in Table 1. As can be seen, the MSE is different for the two records with the same preprocessing and analyzing procedure. All MSE for NR_02 is higher than MSE for ARR_01. This could be explained by the higher noise level of the original recording NR_02. MSE enables to track the clean-up quality of the original recording for the highlighted fetal ECG. The result obtained indicates a significant improvement in the recording after the preprocessing and analysis of the ECG recordings.

The SNR estimates for the extracted fetal ECG, QRS-complex, P-wave, T-wave are shown in Figure 16, where the x-axis describes the noise level, and the y-axis describes the ECG channels.

The visualization of the results indicates that the SNR level for extracted fECG varies for different morphological elements and channels. For instance, SNR for both records is the lowest for P-wave. This suggests that P-wave is the noisiest fECG element and, therefore, be the biggest challenge for extraction. The QRS-complex is less noisy and, therefore, the most easily extracted component. As a general matter, human eyes unable to recognize fetal heart sounds when the SNR is lower than -15 dB. Our experiment shows a moderate SNR for the extracted fECG that makes it possible recognition procedure. We compared our results with a set of similar studies conducted using both synthetic and real data from PhysioNet (see Table 2). In most baseline research, two real ECG datasets (adfecgdb and DAISY) and synthesized (FECGSYNDB) ECG datasets were used.

Results obtained in this study are comparable to the benchmarks, but as shown above, the mixing technique allows the detection of arrhythmias and ectopic contractions when sudden, and short-term changes in the morphology of the fECG signal occur. It can adapt quickly to abrupt changes in fECG, enabling the identification of new artefacts. Moreover, since the target application of this technology is improving the quality of the fetal ECG to detect hypoxia, it is possible to propose the inclusion of the results in the risk assessment scheme for further investigation on a long-term basis.

In summary, non-invasive fECG has the potential to monitor fetal heart rate and fECG morphological parameters. In terms of fetal hypoxia, this information could be included in the auto-generated screening tool to achieve early warning on fetus' safety, identifying fetuses at risk of hypoxia, and enabling early intervention to improve perinatal outcomes. In this respect, the following should be taken into account: 1) separate abnormalities detected in fetal heart rate in the absence of disturbances of beat-to-beat variability or decelerations may not flag fetal hypoxia; 2) the likelihood of fetal hypoxia increases when there are signs of baseline bradycardia coupled with the loss of beat-to-beat variability and decelerations ^[52].

Following up on our previous research, the updated profile of pregnancy complications associated with fetal hypoxia ^[12] includes a new section based mostly on fECG data received on non-invasive techniques. As is the case of the fetal biophysical profile that combines data obtained from different sources and calculates the score points, the profile assigns different scores for abnormal and normal points to achieve the risk level. Table 3 shows the profile of pregnant mother on both the fetal and maternal status related to the presence or absence of pregnancy complications associated with intrauterine hypoxia.

6.   Conclusions

In this study, the approach for automatic fECG processing and updating profile for fetal hypoxia risk estimation was proposed. The ultimate goal was to investigate the possibilities of extraction a fECG from signal recorded non-invasively on the mother's abdomen and assess the feasibility of the fECG morphological evaluation to be included in the automatic screening tool for identifying fetuses at high risk of hypoxia. The set of new artefacts were defined and included in the profile of the pregnant mother for detecting hypoxic fetus. This is the first step towards a comprehensive risk score for fetal hypoxia that can be used not only in a clinical setting but also for remote outpatient monitoring enabling pregnant women to receive timely prenatal and antenatal care, which is crucial in high-risk pregnancies. The main advantage of including fECG parameters into the risk profile is its simplicity of use and programming. As compared to others, this approach can be used for measuring the risk of fetal hypoxia in-utero. Particular attention was paid to fECG signal extraction from maternal abdominal ECG signals using a hybrid filtering technique. The blind source separation, the GEVD and periodic component analysis were applied at the stage of fetal ECG extraction. Detected R-peaks were used to determine RR-intervals and average fHR per minute. The study showed that the quality of fECG extraction is highly dependent on extraneous noise; under equal processing and signal extraction conditions, the MSE and SNR results are slightly different for ARR_01 and NR_02 records. For all that, the result of MSE and SNR shows that our approach is suitable for fetal heartbeat extraction and further fHR analysis from multiple abdominal recording. Moving forward, the appropriate choice of filters could significantly improve the quality and adequacy of extracted fECG signals and raise the prospect of their clinical approval. Therefore, the next work will be aimed at fine-tuning and studying different filtering methods for more accurate fECG extraction and fetal R-peaks detection. In this aspect, further work will focus on performing automatic extraction of morphological features such as PR intervals, QT intervals and ST segments. Non-linear control stability of heartbeat based on modified Zeeman's heartbeat models ^[67] will also be discovered. Elaboration and validation of technology for fHRV and fetal risk assessment through the hypoxia index are also the subjects of future research, as well as an integrated model for fetal hypoxia risk assessment. We also expect that this approach will enable the provision of new medical services with improved accuracy and the advantages of low cost and easy implementation in the near future.

7.   Limitations

We pay attention to the following limitations of this study. First, given the small number of published datasets, there was not enough data for analysis. Second, signal filters reduce noise and improve the signal, but at the same time, they can cause phase delays that affect morphology and temporal alignment between different leads. Next, there is a piece of evidence ^[68] that computerized interpretation of CTG in women with continuous fetal monitoring in labor did not improve clinical outcomes. Finally, this technique has not been tested for real data with the pregnant mother, which raises some doubts about the quick win of non-invasive fECG as an additional method for diagnosing fetal hypoxia. Multifetal pregnancy has not been taken into consideration during signal processing but is included in the profile. Clinical validation and more deep analysis are needed to completely adapt and include these parameters into the risk assessment tool.

Acknowledgment

We sincerely thank anonymous reviewers for critically reading the paper, for their insightful comments and suggesting substantial improvements.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.