Abstract
Purpose: Continuous monitoring of fetal heart rate (FHR) is essential to diagnose heart abnormalities. Therefore, FHR measurement is considered as the most important parameter to evaluate heart function. One method of FHR extraction is done by using fetal phonocardiogram (fPCG) signal, which is obtained directly from the mother abdominal surface with a medical stethoscope. A variety of high-amplitude interference such as maternal heart sound and environmental noise cause a low SNR fPCG signal. In addition, the signal is nonstationary because of changes in features that are highly dependent on pregnancy age, fetal position, maternal obesity, bandwidth of the recording system and nonlinear transmission environment. Methods: In this paper, a sources separation process from the recorded fPCG signal is proposed. Independent component analysis (ICA) has always been one of the most efficient methods for extracting background noise from multichannel data. In order to extract the source signals from the single-channel fPCG data using ICA algorithm, it is necessary to first decompose the signal into multivariate data using a proper decomposition technique. In this paper, we implemented three combined methods of SSA-ICA, Wavelet-ICA and EEMD-ICA. Results: In order to validate the performance of the methods, we used simulated and real fPCG signals. The results indicated that SSA-ICA recovers sources of single-channel signals with different SNRs. Conclusion: The performance criteria such as power spectral density (PSD) peak and cross correlation value show that the SSA-ICA method has been more successful in extracting independent sources.
Keywords: Fetal phonocardiogram, Heart sound, Independent component analysis, Singular spectrum analysis
Introduction
Since 1970, fetal heart rate (FHR) measurement has been the most important noninvasive diagnostic parameter for fetal heart monitoring [1]. Normal fetal heart rate is between 120 and 160 bpm (beat per minute), which decreases with pregnancy age [2]. Higher or lower FHR values indicate a problem with the structure and condition of the fetal heart.
The mechanical function of the heart results in the production of sounds whose analysis can provide useful information on the structural status of the heart. So far, many improvements have been made in development of sensors, data acquisition systems, and heart sound signal processing techniques. Monitoring of fetal heart function begins at the 24th week of pregnancy and can be continued daily or weekly depending on the condition of the pregnant woman. In high-risk pregnancies, continuous monitoring of FHR changes is necessary for early diagnosis of heart malformations to prevent intrauterine death or permanent fetal injury [3].
One of the methods of FHR extraction is the use of fetal phonocardiogram (fPCG) signal, which is obtained directly from the mother abdominal surface using medical stethoscope. Main advantage of this method is that the signal recording is completely non-invasive and low cost. However, the recorded signal is highly sensitive to extra sounds of internal organs and outdoor environment. Noises affecting the recorded signal include:
Fetal movements
Mother's gastrointestinal sound
Mother's heart sound
Mother's respiratory sound
Sensor movement during recording
Environmental noise [4]
Our goal is to extract the source components of the single-channel fPCG signal recorded from the mother abdomen. The signal processing methods used so far to extract fPCG components are direct analysis, adaptive filters, linear analysis including short-time fourier transform, wigner- ville distribution, wavelet transform (WT) and blind source separation (BSS), nonlinear analysis, and polynomial analysis [3].
By reviewing previous researches and existing methods, it seems that the works done in the field of BSS and in particular the Independent Component Analysis (ICA) algorithm have yielded acceptable results in extracting fetal heart sound from background noise. The ICA method is a data-dependent algorithm that decomposes the signal to its underlying sources [5]. Since recorded abdominal signal is single-channel, an alternative approach named single-channel ICA (SCICA) is proposed [6]. Previous studies in this field demonstrate capability of the method in separating the signal into its constituent components. Most of the processing methods that have examined this signal have focused on extraction of the fetal component of this signal. Gonzalez et al. presented a method in which single-channel abdominal phonocardiograms recorded from three individuals were analyzed using ICA algorithm to extract fetal heart sound from noisy signal [7]. They used the power spectrum density (PSD) of the two extracted sources to identify independent components. In addition, the abdominal ECG signal was used as an auxiliary signal with respect to its synchronization with the heart sound. In another similar work, Gonzales et al. extracted three sources of fetal heart sound, maternal respiration along with maternal pulse wave, and line-noise from the single-channel abdominal phonocardiogram [8]. This group continued their work by utilizing the SCICA to extract independent sources. A post processing step with the K-means method was used to group the independent components (ICs) with similar characteristics together for better separation of the desired sources. In order to identify the ICs corresponding to an independent source, they adopted a criterion based on periodic analysis [9]. Subsequently, Gonzales et al. proposed a two-stage noise removal algorithm including SCICA and experimental filtering to extract fetal heart sound and FHR. They compared the results with the reference abdominal ECG signal for validation [10].
So far, most previous works have focused on recovering only fetal heart sound while discarding other informative sources. The only work that has considered the retrieving of physiological sources, requires post-processing operations on extracted ICs which is time-consuming task [9]. Our focus in this paper is to recover the independent physiological sources underlying the single-channel signal using ICA algorithm. An appropriate method could break down the recorded single-channel abdominal signal to ICs including fetal heart sound, maternal heart sound, respiratory and low-frequency noise, and environmental noise.
Since ICA requires multi-channel input, we initially need to map the signal into multivariate data using appropriate decomposition method. Singular spectrum analysis (SSA) is a technique which decomposes mixed signal based on desired sources subspaces. SSA has been widely applied in EEG signal processing to separate EEG artifacts [11, 12]. Recently, combination of SSA and ICA has been used to efficiently separate available sources in the single-channel EEG signals [13]. We have investigated three different methods of singular spectrum analysis, wavelet transform and ensemble empirical mode decomposition (EEMD) in combination with ICA, denoted by SSA-ICA, Wavelet-ICA and EEMD-ICA, respectively. Simulation results showed that SSA-ICA with a new criterion based on PSD of each eigenvector performs well in this regard.
Materials and methods
Independent component analysis (ICA)
Independent component analysis is a method that is capable of extracting appropriate information buried in noisy signals by decomposing the recorded signal into independent components. ICA method has been widely used in blind source separation, feature extraction and pattern detection of biological signals. Suppose observed random variables which are combination of random variables by matrix as follows:
| 1 |
The main issue in ICA is estimation of unknown mixing matrix and sources using observations with assumptions about the statistical properties of the sources distribution. In other words, the goal of ICA is to find the appropriate weight matrix as un-mixing matrix. Then, the source components , are revealed by using the equation . In ICA theory, we consider the following assumptions and limitations:
Source components are linearly mixed and statistically independent.
Source components must have a non-Gaussian distribution.
The number of observed signals is greater than or equal to the number of source components [14].
Mapping a single-channel signal into a multidimensional dataset
ICA method assumes that a multichannel observations measured by a set of sensors can be modeled as a linear mixture of underlying sources [9]. In general, the number of observations or measured signals should be greater than or equal to the number of sources. To apply ICA to a single-channel signal, data must be mapped to a multivariate data. If a multichannel recording can be created from a single-channel signal, ICA might be used to extract the signal components, which are independent from each other. Such mapping can be done by decomposing the single-channel signal into different spectral modes.
In this section, we review some existing methods and study their source extraction capabilities in single-channel fPCG signal processing. All the methods are based on decomposing the signal into components of disjoint spectra.
Singular spectrum analysis-ICA (SSA-ICA)
The SSA is a subspace based method that involves two stages of decomposition and reconstruction. The analysis consists of two steps of embedding and singular value decomposition (SVD). In the embedding step, sample single-channel signal is mapped into trajectory matrix of size as follows:
| 2 |
where is the window length and . Then, the SVD of the matrix is calculated to decompose it into number of trajectory matrices, Suppose the eigenvalues and eigenvectors of the covariance matrix are and , respectively. Note that eigenvalues and corresponding eigenvectors are arranged in descending order. If , the th matrix can be represented as:
| 3 |
After substitution of in (3), matrix is given by
| 4 |
where the term constitutes the subspace for the th component of . By projecting the matrix onto the subspace, trajectory matrix for th signal present in is obtained.
The reconstruction step consists of two stages of grouping and diagonal averaging. After decomposition of the matrix into matrices, they are separated into groups. Suppose are indices related to eigenvalues. Then, the matrix corresponding to the th group can be expressed as follows:
| 5 |
After splitting the indices into subsets, say , the trajectory matrix will be expressed as follows:
| 6 |
Finally, to reconstruct a specific component, corresponding trajectory matrix is mapped into a single-channel signal. The th sample of the desired signal is obtained as follows:
| 7 |
Such that is element of the trajectory matrix in the th row and the th column [11].
The critical step in SSA is identifying the subspaces of the desired signal. Unlike classical SSA, where magnitude of the eigenvalues of is used to identify signal subspaces, Maddirala et al. proposed a novel grouping criteria based on frequency of the eigenvectors to separate the sources from single-channel EEG signal [13]. They first applied SSA on the measured single-channel signal to decompose it into higher and lower frequency components, say and , respectively. This decomposition was done based on the dominant frequency of PSD of each eigenvector. With the dominant frequency information of each eigenvector and a proper threshold value, the eigenvectors were decomposed into two groups and the corresponding signal was computed using (7). Again, SSA was applied on the low frequency signal and decomposed it into two components. This process continued for levels to create a data matrix of size .
Wavelet-ICA
Wavelet transform is a powerful tool in biological signal processing that breaks down a given single-channel signal into spectrally non-overlapping components. In first level, the signal is passed through a low and high pass filters, simultaneously. Then, WT is applied on the low frequency component and this process will be continued for levels. Finally, a multivariate data matrix of size is constructed by each level’s coefficients. Choosing the appropriate mother wavelet and also the order of WT plays an important role on the overall performance [15].
Ensemble empirical mode decomposition-ICA (EEMD-ICA)
EMD is used to decompose a single-channel signal into a finite set of modes called intrinsic modes functions (IMFs). IMFs are spectrally independent, zero mean oscillatory functions. Thus, the given signal can be expressed as the sum of IMFs and the residual signal as . Here, is the th IMF and is the residual signal after extraction of IMFs.
One of the main problems of the EMD algorithm is its high sensitivity to noise. A more robust version of the EMD, called EEMD, has been introduced. In this method, first, identically distributed and independent white noise is added to the given signal. Then, each IMF is obtained using EMD. This process is repeated for ensemble of trials () to drive the averaged IMF. The ratio of noise standard deviation to standard deviation of the signal is referred to as the noise parameter () [15].
After extracting the mean IMFs, the matrix is constructed to apply the ICA algorithm. Unlike Wavelet-ICA, EEMD-ICA method does not require prior information from extracted sources. However, the performance of the EEMD-ICA method largely depends on the number of trials and the noise parameter.
Datasets
The methods were evaluated on two different fPCG datasets available at Physionet [16]. The databases are Simulated Fetal Phonocardiograms [2] and Fetal PCG Database [4].
First dataset includes a series of synthetic fPCGs relative to different fetal states and recording conditions which are generated using simulation software [2]. The simulation software can simulate different physiological and pathological fetal conditions and recording situations by changing some parameters of the system. Simulated signals are sequence of frames, each of them including simulated first heart sound (S1) and second heart sound (S2) wavelets, corrupted by noise. The added noise includes mother heart sound, sound of mother internal organs, fetal movements, and environmental noise. These signals are constructed with a range of SNR values from −4.4 dB to −26.8 dB that are typically present in actual recordings. The sampling frequency is 1 kHz and 16 bits resolution.
The second dataset contains phonocardiographic signals recorded from 26 pregnant women during 31–40 weeks of gestation. This dataset was recorded by a portable phonocardiograph during 20-min sessions with a sampling frequency of 333 Hz and 8 bits resolution. The sources of noise include vibrations caused by maternal heart sound, maternal body organ sounds, fetal movements, and surrounding environment [4].
Results
To evaluate the performances of the techniques, two signals with high SNR of −4.4 dB and low SNR of −23.8 dB were selected from the first dataset. After applying each algorithm, the PSD of the extracted sources were estimated using burg method to relate them to four signals mentioned above.
Application of the SSA-ICA on simulated signals
First, SSA-ICA was applied to 10 s of higher SNR signal. To reduce the computational demands, down sampling to 500 Hz was performed prior to any processing block. Figure 1 shows the time domain noisy signal with a SNR of −4.4 dB and its PSD. As expected, a dominant frequency around 60 Hz associated with the fetus component and a smaller peak around 20 Hz associated with the mother activity can be observed. Also, low and high frequency noises are practically very weak.
Fig. 1.
The abdominal signal with SNR of −4.4 dB and its PSD
In the simulations, the window length is set to 33, the number of decomposition levels is set to five and the number of ICs to be extracted is set to three. Figure 2 shows the extracted sources and their corresponding PSDs. Efficient separation of fetal heart sound, maternal heart sound and low frequency (LF) noise can be seen according to single-peak of retrieved sources spectrums. The method not only separates fPCG, but also clearly distinguishes the two main sounds including S1 and S2. The downward arrows in first extracted source represent fetal S1 and the upward arrows represent fetal S2. In second source, which relates to the mother heart sound, the red circles indicate maternal S1 and the green circles indicate maternal S2. The third estimated source associates with the noise due to vibrations created by maternal body organs and fetal movements.
Fig. 2.
The extracted sources by SSA-ICA and their corresponding PSDs for SNR of −4.4 dB
Next, a signal with very low SNR of −23.8 dB was selected to evaluate the method. To simulate low SNR condition, the mother and fetal heart sounds are combined with approximately equal amplitude ratios. Also, the components related to low and high frequency noises are stronger.
Figure 3 shows the time domain signal and its PSD. It is obvious from Fig. 3 that the low SNR signal has more frequency spectrum components from 0 to 250 Hz. To obtain the input matrix to the ICA algorithm, the number of decomposition levels is set to five and the number of sources is set to four. Figure 4 shows the extracted sources as well as corresponding PSDs. Considering the PSDs, the extracted sources can be related to low frequency noise, maternal heart sound, fetal heart sound, and high frequency (HF) noise, respectively. Also, maternal S1 (red circles) and S2 (green circles) in second extracted source are separable. Fetal S1 (downward arrows) and S2 (upward arrows) in third extracted source are well separated by the algorithm. The result has shown that the SSA-ICA technique reveals the superiority in estimating the sources in high and low SNR signals.
Fig. 3.
The abdominal signal with SNR of −23.8 dB and its PSD
Fig. 4.
The extracted sources by SSA-ICA and their corresponding PSDs for SNR of −23.8 dB
Application of the wavelet-ICA on simulated signals
Next, the Wavelet-ICA method was applied on signals of the previous section to extract the sources. For comparison purpose, the number of decomposition levels was set to five and the data matrix was constructed using detail coefficients of the entire levels and approximation coefficients of the last level. Based on previous studies, db4 mother wavelet was chosen because of its similarity to the heart sounds [3]. For high SNR signal, the number of ICs to be extracted is set to three, as the SSA-ICA method. The decomposed signals and their corresponding PSDs are shown in Fig. 5. It is obvious that the first estimated source is related to the fPCG. However, upon careful observation of second and third sources, it is clear that the remnants of maternal components still present in the low frequency noise. The result is that the above method has not been able to extract the sources accurately.
Fig. 5.
The extracted sources by Wavelet-ICA and their corresponding PSDs for SNR of −4.4 dB
The application of Wavelet-ICA for very low SNR signal decomposition was also studied. The number of sources in the ICA algorithm was set to four, as previous. Figure 6 shows that Wavelet-ICA performed worse than SSA-ICA in extracting sources from low SNR signal. The first and second extracted sources, which can be related to low frequency noise and maternal heart sound, have similar time components. In other words, it seems that a combination of both has been appeared in two extracted sources. The Wavelet-ICA algorithm performed clearly worse than SSA-ICA in both simulations.
Fig. 6.
The extracted sources by Wavelet-ICA and their corresponding PSDs for SNR of −23.8 dB
Application of the EEMD-ICA on simulated signals
In this section, the results of applying the EEMD-ICA method for separation of sources are presented. The six rows data matrix was constructed by five extracted IMFs and residual signal. The EEMD-ICA performance is significantly affected by the parameters and . As the extracted sources had low frequency contents, was set to a large value 2 [17].
Three estimated source signals from the single-channel high SNR signal are shown in Fig. 7. According to the frequency of peaks in the PSD of first and second extracted sources, the fPCG has been presented in both sources. The third source with a broadband spectrum, probably contains mother heart sound, high frequency and low frequency noises.
Fig. 7.
The extracted sources by EEMD-ICA and their corresponding PSDs for SNR of −4.4 dB
Figure 8 shows the performance of the EEMD-ICA method in estimating four sources of very low SNR signal. Considering the PSD peaks, the method has not been able to extract the sources accurately and only high frequency noise can be detected. Therefore, EEMD-ICA method has poor performance in resource allocation.
Fig. 8.
The extracted sources by EEMD-ICA and their corresponding PSDs for SNR of −23.8 dB
FHR estimation using SSA-ICA method
For estimation of the FHR, it is mandatory to detect S1 peaks in the extracted fPCG signal using a threshold level. A dynamic threshold value was selected based on the average of absolute amplitude of the signal. To test the performance of the SSA-ICA method, a low SNR signal was considered. A graphical representation of the extracted fPCG signal is shown in Fig. 9. The red line indicates the threshold level and the black curve indicates the position of the S1 peaks. Then, FHR is estimated easily by counting the number of peaks and substitution in the equation below:
| 8 |
where is number of peaks and is time length of the extracted fPCG. The obtained FHR is 148 bpm for this case. It is identified that the method is able to detect all S1 peaks even from low SNR abdominal signal. Thus, reliable FHR is obtained, which can be used for further analysis.
Fig. 9.

Estimation of the FHR in a low SNR abdominal signal
Application of the SSA-ICA method on real PCG signal
To illustrate the performance of the SSA-ICA algorithm on real-life applications, we considered the following case. A 10 s signal recorded from the abdominal of a 39 weeks pregnant woman is selected, as shown in Fig. 10. The number of ICs to be extracted was set to four. Figure 11 indicates extracted sources and their PSD. As previous, estimated sources include low frequency noise, fetal heart sound, maternal heart sound, and high frequency noise, respectively.
Fig. 10.
Recorded abdominal signal at 39th week gestation
Fig. 11.
The extracted sources by SSA-ICA and their corresponding PSDs for real fPCG signal
Discussion
In order to examine whether the extracted sources were mutually independent or not, the cross correlation was computed as a source separability measure. The small value indicates that the two components are efficiently separated and the large value represents the components are poorly separated. Ten signals with different SNRs were selected and all three methods were applied on them. Figure 12 shows the cross correlation values between pairs of estimated sources for a high SNR signal. The horizontal axis represents different combinations of two sources from the four extracted sources and vertical axis represents the cross correlation value. It is obvious that in EEMD-ICA and Wavelet-ICA methods, the sources (1, 2) and (3 and 4) have the largest correlation. On the other hand, the extracted sources by the SSA-ICA method has the least pairwise similarity.
Fig. 12.

Cross correlation values between pairs of estimated sources by three method
To compare the three methods more precisely, the average of two maximum cross correlation values between the extracted sources of a signal was calculated. Figure 13 illustrates the obtained values for signals with different SNRs. As shown, the SSA-ICA outperforms the other two especially for high noise-to-signal ratios. In other words, the SSA-ICA method is more efficient in dealing with more noisy signals.
Fig. 13.

The average of two maximum cross correlation values between the extracted sources for signals with different SNRs
Conclusion
Continuous monitoring of FHR in high risk pregnancies is still a challenging task. The fPCG signal can be considered as a totally passive, low cost and noninvasive tool for FHR estimation. Due to the low SNR of sound signal recorded through maternal abdomen, accurate estimation of the fetal component of the signal is a difficult task.
An overview of previous studies is conducted which shows the ICA method performs well in extracting independent sources of fPCG signal. In this paper, combinations of ICA with Wavelet, EEMD and SSA methods were presented to decompose the single-channel signal. The results indicated that the SSA-ICA method out performs both Wavelet-ICA and EEMD-ICA in extracting four sources of fetal heart sound, mother heart sound, low and high frequency noises in simulated signals with different SNRs and real data. The cross correlation between two extracted sources shows a high capability of the method in sources separation. The advantage of the SSA-ICA method is that the decomposition process is data-based. In contrast, the decomposition in the Wavelet-ICA method is dependent on the selection of the mother wavelet. The EEMD-ICA method also needs to adjust the parameters correctly for proper sources separation.
Compliance with ethical standards
Conflict of interest
There is no conflict of interest.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Sartwelle TP. Electronic fetal monitoring: a bridge too far. J Leg Med. 2012;33:313–379. doi: 10.1080/01947648.2012.714321. [DOI] [PubMed] [Google Scholar]
- 2.Cesarelli M, Ruffo M, Romano M, Bifulco P. Simulation of foetal phonocardiographic recordings for testing of FHR extracting algorithms. Comput Methods Programs Biomed. 2012;107:513–523. doi: 10.1016/j.cmpb.2011.11.008. [DOI] [PubMed] [Google Scholar]
- 3.Adithya PC, Sankar R, Moreno AM, Hart S. Trends in fetal monitoring through phonocardiography: Challenges and future directions. Biomed Signal Process Control. 2017;33:289–305. doi: 10.1016/j.bspc.2016.11.007. [DOI] [Google Scholar]
- 4.Ruffo M, Cesarelli M, Romano M, Bifulco P, Fratini A. An algorithm for FHR estimation from foetal phonocardiographics signals. Biomed Signal Process Control. 2010;5:131–141. doi: 10.1016/j.bspc.2010.02.002. [DOI] [Google Scholar]
- 5.Clifford G. Chapter 15: Blind source separation using PCA and ICA. Biomedical Signal and Image Processing; 2008.
- 6.Davies M, James CJ. Source separation using single channel ICA. Signal Process. 2007;87:1819–1832. doi: 10.1016/j.sigpro.2007.01.011. [DOI] [Google Scholar]
- 7.Jimenez-Gonzalez A, James CJ. Blind source separation to extract foetal heart sounds from noisy abdominal phonograms: A single-channel method. Conference: Advances in Medical, Signal and Information Processing; 2008.
- 8.Jimenez-Gonzalez A, James CJ. Source separation of foetal heart sounds and maternal activity from single-channel phonograms: A temporal independent component analysis approach. Computers in Cardiology. 2008; 949–952.
- 9.Jiménez-González A, James CJ. Time-structure based reconstruction of physiological independent sources extracted from noisy abdominal phonograms. IEEE Trans Biomed Eng. 2010;57:2322–2330. doi: 10.1109/TBME.2010.2051226. [DOI] [PubMed] [Google Scholar]
- 10.Jiménez-González A, James CJ. De-noising the abdominal phonogram for foetal heart rate extraction: blind source separation versus empirical filtering. Int. Con. of the IEEE EMBS. 2013; 1358–1361. [DOI] [PubMed]
- 11.Maddirala AK, Ahamed R. Removal of EMG artifacts from single-channel EEG signal using singular spectrum analysis. IEEE International Circuits and Systems Symposium (ICSyS). 2015; 111–115.
- 12.Maddirala AK, Ahamed R. Removal of EOG artifacts from single-channel EEG signals using combined singular spectrum analysis and adaptive noise canceler. IEEE Sens J. 2016;16:8279–8287. [Google Scholar]
- 13.Maddirala AK, Ahamed R. Separation of sources from single-channel EEG signals using Independent Component Analysis. IEEE Trans Instrum Meas. 2018;67:382–393. doi: 10.1109/TIM.2017.2775358. [DOI] [Google Scholar]
- 14.Hyvarinen A, Oja E. Independent component analysis: algorithms and applications. Neural Netw. 2000;13:411–430. doi: 10.1016/S0893-6080(00)00026-5. [DOI] [PubMed] [Google Scholar]
- 15.Mijovic B, De Vos M, Gligorijevic I, Taelman J, Huffel S. Source separation from single-channel recordings by combining empirical-mode decomposition and independent component analysis. IEEE Trans Biomed Eng. 2010;57:2188–2196. doi: 10.1109/TBME.2010.2051440. [DOI] [PubMed] [Google Scholar]
- 16.Goldberger AL, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation. 2000;101(23):215–220. doi: 10.1161/01.CIR.101.23.e215. [DOI] [PubMed] [Google Scholar]
- 17.Wu Z, Huang NE. Ensemble empirical mode decomposition: A noise-assisted data analysis method. Adv Adaptive Data Anal. 2009;1:1–41. doi: 10.1142/S1793536909000047. [DOI] [Google Scholar]










