Blind extraction of fetal and maternal components from the abdominal electrocardiogram: An ICA implementation for low-dimensional recordings
Introduction
The fetal electrocardiogram (FECG), a bioelectric signal produced by the cardiac activity of the fetus, allows to measure the fetal heart rate (FHR) and its variability, which are significant parameters for fetal surveillance [2,3]. Contrary to Doppler ultrasonography, that irradiates the fetus with ultrasonic energy, the electrocardiography does not apply any form of energy to the fetal tissues, making it possible the continuous monitoring of the FHR during pregnancy and labor. The technique is completely noninvasive; it has low power requirements and therefore can be used over extended periods (e.g., 24 h). Thus, by positioning electrodes on the maternal womb, the FECG can be detected, but highly contaminated by interfering sources such as the maternal ECG and EMG (MECG and MEMG, respectively), artefactual movements and line-noise [3,4]. The recorded signal is referred to as the abdominal ECG (AECG) and, because of (1) the usual larger energy of the interfering sources and (2) the temporal and spectral overlapping between such sources and the FECG, the recovery of the latter from the AECG cannot be achieved by traditional signal processing approaches like digital filtering [2,[4], [5], [6]].
During decades, there have been many efforts to implement the extraction of the FECG from the AECG by techniques such as adaptive filtering, template subtraction, spatial filtering, time-frequency analysis, Blind Source Separation (BSS) based methods, neural network-based methods, non-negative matrix factorization, empirical mode decomposition, Kalman filtering, tensor decomposition, empirical methods and by the fusion of some of them [2,[4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]]. Some of these approaches were presented in response to the Computing in Cardiology Challenge 2013 (CinC13) [15], where more than 50 teams attempted the extraction of the FECG from a large database of AECGs [15,16]. According to [16], while most teams followed a two-step approach to first remove the maternal information and then to detect the FQRS, “most of the top-performing entries in the challenge made use of strategies that differentiated them from their competitors”. Certainly, after reading the approaches followed by the four best performing entries (both official and unofficial in events 4 and 5 [15]), it was noticed that the strategies of each team were mainly based (1) on developing a hybrid-multistage method (with more than 5 stages) and, most interestingly to this work, that, in one stage (at least), (2) the signals were processed by a BSS method such as Independent Component Analysis (ICA) [1,17,18] or Principal Component Analysis (PCA) [19].
In their work, [1,[17], [18], [19]] mention that most of the conventional approaches have not totally managed to solve the extraction problem since they may fail on the elimination of the MECG and thus, produce a number of positive false when detecting the R wave for FHR measuring. The BSS approach by means of ICA, on the other hand, has been successfully used to separate out the sources underlying the AECG (on datasets between 6 and 55 channels) [[20], [21], [22], [23], [24], [25], [26]], making it suitable to detect the R wave in the estimate cardiac traces and then, to calculate the fetal and maternal heart rates. Thus, results based on ICA implementations have been promising [21,24], however, since the technique performance is highly dependent on the dataset dimension (i.e., the number of observations, x, has to be larger or equal than the number of sources [27]), its application on recordings composed by a low number of channels (i.e., low-dimensional recordings) like the AECGs in the CinC13 challenge is not always reliable (especially in presence of noisy data).
On this matter, [1,[17], [18], [19]] state that, because of the interfering sources, in ICA-based methods, the two basic assumptions of (1) instantaneous invariant linear mixing and (2) the number of mixtures being larger or equal than the number of sources might not be fully satisfied. As a result, when processing such AECGs by ICA, the FECG might be mostly contaminated by the MECG that would reduce the FHR detection accuracy, especially in low signal to noise situations [27,28]. Because of that, the top-performing entries in the CinC13 challenge developed frameworks whose particular stages would be dealing with the possible FECG extraction problems due to ICA, if any [1,17,18]. This gave rise to multi-stage hybrid algorithms that (in a very general description), preprocess the signals to reduce the number of sources, apply FastICA or JADE to extract the MECG, automatically detect the MQRS in a single or multi-channel configuration (e.g., adaptively or by a Pan and Tompkins detector), cancel the MECG by different methods (e.g., template adaptation, Kalman filtering, singular value decomposition or template subtraction), conduct the FECG enhancement using different approaches (e.g., ICA again or template subtraction) and, finally, automatically detect and correct the FQRS in a single or multi-channel configuration (iteratively conducted in some cases) [1,17,18], all of them tuned by the expert knowledge of their developers and programmed to automatically produce a single vector with the FQRS positions (as requested by the CinC13 challenge).
Alternatively to these multi-stage MECG canceling approaches, the present study proposes that ICA would manage to process the AECGs of the CinC challenge (or other low-dimensional biomedical data) and successfully separate good-quality traces corresponding to the fetal and maternal ECGs (that are useful to obtain reliable references of the R wave positions per channel), as soon as the data dimensionality problem is suitably solved. Thus, to meet the terms of high dimensionality in the recordings for ICA to work, this paper followed the strategy presented in [[29], [30], [31]], where a low-dimensional dataset was projected into a higher dimensional space in order to build up a robust multidimensional representation suitable to be analyzed by ICA. The approach, referred to as Space-Time ICA (ST-ICA), has been applied on preselected channels of ictal scalp EEG recordings, where the source corresponding to the ictal process was successfully retrieved [[29], [30], [31]]. Therefore, the purpose of this work was to study the performance of ST-ICA when working on low-dimensional recordings of abdominal ECGs, this based on a popular and fast to converge implementation known as FastICA [27,32]. To this end, and aiming to develop a systematic evaluation, we studied the quality and reliability of the FECG trace estimated by ST-FastICA as a function of (1) the number of channels used to construct the higher-dimensional representation (i.e., 4 channels, 3 channels and 2 channels) and (2) the influence of both, the working modes (i.e., deflation and symmetric) and the nonlinear functions used by FastICA (i.e., pow3, tanh and gauss) [27,32,33]. These by means of a set of 24 indexes that quantified the amount of fetal information recovered in the estimate FECG, the residual maternal information left in the FECG and the effect of the FECG quality on the calculation of the RR intervals and the FHR [4,[34], [35], [36]]. In addition, as requested by the paper reviewers, the performance of ST-ICA was compared against the performance of two of the top-performing entries in the CinC13 challenge [1,18].
This paper is organized as follows: Section 2 details the implementation of ST-FastICA for the AECG analysis and describes the indexes used to evaluate its performance. Section 3 illustrates the retrieve maternal and fetal cardiac traces as well as the performance figures. Finally, Sections 4 and 5 present the discussion and conclusions, respectively.
Section snippets
Dataset
The recordings used in this work were downloaded from the “Abdominal and Direct Fetal Electrocardiogram Database” in Physionet [37], which contains five recordings obtained from five different women in labor, between 38 and 41 weeks of gestation. Each recording has four AECG channels and a direct FECG (FECGD, registered on the fetal head and accompanied by annotations about the fetal R wave positions). The signals were recorded during 5 min using a sampling rate of 1 kHz and a resolution of 16
Results
Fig. 1 presents a segment (5 s length) of subject 2 recording (i.e., both the direct and abdominal channels), the output of the standard implementation of FastICA (i.e., without stages 2.2.2 and 2.2.4, and configured as sym-gauss), and the fetal and maternal ECGs finally extracted by ST-FastICA (using 4 and 2 channels in stage 2.2.2). In (a), from top to bottom, the pass-band versions of the FECGD and the four AECGs, in (b), from top to bottom, the pass-band version of the FECGD and the four
Discussion
This work studied the performance of ST-FastICA on the separation of fetal and maternal ECGs from a dataset of low-dimensional abdominal recordings (i.e., composed by 4 channels). This by using 24 indexes to test the quality of the FECGs extracted and their effect on the RR intervals and the FHR. To this end, ST-FastICA was implemented by combining the Method of Delays with FastICA and an automatic classifier and applied to the “Abdominal and Direct Fetal Electrocardiogram Database” in
Conclusions
This work studied ST-FastICA on the separation of fetal and maternal ECGs from a dataset of low-dimensional abdominal recordings. To this end, 24 indexes were used to qualify the FECGs retrieved by using 4, 3, 2 channels and combinations of them (e.g., channels 1–3–4 or channels 2–4) with different configurations of FastICA (e.g., deflation and pow3 or symmetric and tanh). The study, conducted by statistically comparing 22 indexes in five subjects, gave rise to three findings. First, that the
Author contributions statement
Both authors equally contributed to the development of this work with the conceptualization, formal analysis, methodology, software, supervision, validation, visualization, writing and editing the original draft. AJG reviewed and edited this revised version.
Acknowledgements
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (45)
- et al.
Fetal heart rate monitoring based on independent component analysis
Comput. Biol. Med.
(2006) - et al.
Independent component analysis: algorithms and applications
Neural Netw.
(2000) - et al.
Extracting qualitative dynamics from experimental data
Phys. D Nonlinear Phenom.
(1986) - et al.
Source separation using single channel ICA
Signal Processing
(2007) - et al.
A multi-step approach for non-invasive fetal ECG analysis national research council of Italy, Institute Of Clinical Physiology, Pisa, Italy University Of Pisa, Italy
Comput. Cardiol.
(2010) - et al.
Detection and processing techniques of FECG signal for fetal monitoring
Biol. Proced. Online
(2009) - et al.
Review article: non-invasive fetal heart rate monitoring techniques
Annu. ORNL Biomed. Sci. Eng. Cent. Conf.
(2014) - et al.
A practical guide to non-invasive foetal electrocardiogram extraction and analysis
Physiol. Meas.
(2016) - et al.
A review of fetal ECG signal processing: issues and promising directions, Open Pacing
Electrophysiol. Ther. J.
(2010) - et al.
Non-adaptive methods for fetal ECG signal processing: a review and appraisal
Sensors
(2018)
QRStree: A prefix tree-based model to fetal QRS complexes detection
PLoS One
Recovery of the Fetal Electrocardiogram for Morphological Analysis From Two Trans-abdominal Channels Via Optimal Shrinkage
Separation of Fetal-ECG from single channel abdominal ECG using activation scaled non-negative matrix factorization
IEEE J. Biomed. Heal. Informatics
Extraction of foetal ECG from abdominal ECG by nonlinear transformation and estimations
Comput. Methods Programs Biomed.
Noninvasive fetal ECG extraction using doubly constrained block-term decomposition
Math. Biosci. Eng.
An open-source framework for stress-testing non-invasive foetal ECG extraction algorithms
Physiol. Meas.
Feasibility of non-invasive fetal electrocardiographic interval measurement in the outpatient clinical setting
Pediatr. Cardiol.
Home-based mobile fetal/maternal electrocardiogram acquisition and extraction with cloud assistance
2019 IEEE MTT-S Int. Microw. Biomed. Conf., IEEE
Noninvasive Fetal ECG - The PhysioNet Computing in Cardiology Challenge
Noninvasive fetal ECG: the PhysioNet/Computing in cardiology challenge 2013
Comput. Cardiol.
Maternal signal estimation by Kalman filtering and Template Adaptation for fetal heart rate extraction
Comput. Cardiol.
Non-invasive FECG extraction from a set of abdominal sensors
Comput. Cardiol.
Cited by (17)
Graphene coated textile based highly flexible and washable sports bra for human health monitoring
2020, Materials and DesignA Nonlinear Functional Link Multilayer Perceptron Using Volterra Series as an Adaptive Noise Canceler for the Extraction of Fetal Electrocardiogram
2024, Annals of Biomedical EngineeringInvestigation of fetal ECG signal using textile-based electrodes
2024, Journal of the Textile InstituteDeep learning with fetal ECG recognition
2023, Physiological MeasurementComparative Assessment of BSS Techniques for Non-Invasive Extraction of Fetal ECG from Abdominal ECG Signal
2023, Journal of The Institution of Engineers (India): Series B