Non-invasive acquisition of fetal ECG from the maternal xyphoid process: a feasibility study in pregnant sheep and a call for open data sets

The INFANT study in ~47 000 pregnancies reported no evidence of benefit on neonatal outcomes associated with the use of decision-support software in conjunction with cardiotocography (CTG) compared with CTG alone (Group 2017). Recent studies have shown that the true predictive ability of CTG can only be determined once it is collected at a sampling rate that preserves the underlying physiological information (i.e. 1000 Hz rather than 4 Hz) (Durosier et al 2014, Li et al 2015). This requires fetal electrocardiogram, rather than ultrasound, to derive the fetal heart rate (FHR) (Frasch et al 2017).

To address this challenge, we developed an algorithm for low-cost, portable high quality maternal and fetal ECG monitoring that could be used in human and large animal studies to record fetal ECG non-invasively and in an open format (Li and Wu 2017, Li et al 2017). While the approach has been validated in human data sets, such data sets remain scarce.

The fetal sheep model is widely used to mimic various fetal conditions, yet ECG recordings in a public data set form are not available to test the predictive ability of ECG and FHR. Part of the challenge is the requirement to perform a sterile fetal surgical instrumentation with precordial ECG leads (Burns et al 2015). The ability to record fetal sheep ECG from the mother's surface, i.e. non-invasively, would reduce the complexity of the model, hopefully contributing to its wider utilization and increased availability of testable fetal ECG data sets recorded under well-controlled animal experimental conditions with translational relevance.

To explore the possibility of using fetal sheep model data, here we tested the ability of our algorithm to extract fetal ECG from a single channel maternal ECG signal recorded over the xyphoid process (maternal xyphoid ECG signal), but not the abdominal maternal ECG signal, the routine way maternal ECG is recorded in this animal model. In this study, the maternal ECG lead was placed just below the xiphoid process. Due to its proximity to the abdomen, we hypothesized that the algorithm could extract the fetal ECG from the maternal xyphoid ECG.

Animal care followed the guidelines of the Canadian Council on Animal Care and the approval by the University of Montreal Council on Animal Care (protocol #10-Rech-1560).

Anesthesia and surgical procedure

Data acquisition

Data analysis

The applied fetal ECG extraction algorithm for the single channel maternal xyphoid ECG signal is composed of three steps. (Li and Wu 2017) First, the de-shape short-time Fourier transform (STFT) (Lin et al 2017) is applied to estimate the maternal instantaneous heart rate, and hence the maternal R peaks via the beat tracking algorithm (Lin et al 2017). Second, the nonlocal Euclidean median (NLEM) is applied to recover the maternal ECG (Li and Wu 2017). By a direct subtraction, we obtain the rough fetal ECG. Finally, the fetal R peaks and hence the fetal ECG are obtained by applying the de-shape STFT, beat tracking, and the NLEM on the rough fetal ECG. In Li and Wu (2017), the algorithm is shown to successfully extract the fetal ECG from the maternal abdominal ECG signal. We refer the reader with interest to Li and Wu (2017) for details. The code is publicly available in https://sites.google.com/site/hautiengwu/home/download. See figure 1 for a typical example for the human maternal abdominal ECG signal.

Zoom In Zoom Out Reset image size

We also consider other available single channel fetal ECG extraction algorithms, particular the template subtraction (TS) approaches, including TS_C proposed by Cerutti et al (1986) and TS_PCA proposed by Kanjilal et al (1997). Here, we follow the nominations proposed in Andreotti et al (2016). The code is publicly available.

All animals were in good health during surgery with fetal pH, pO₂, pCO₂, O₂Sat, lactate and BE in physiological range at 7.29  ±  0.03, 33.2  ±  8.4, 56.0  ±  7.8, 78.3  ±  7.6, 2.8  ±  0.6, −0.3  ±  2.4, respectively. See tables 2 and 3 for details.

Four ewes carried singletons, seven ewes carried twins and one animal had triplets. Six were female and seven were male. See table 1 for details. Maternal body weight averaged 68  ±  6 kg and fetal body weight averaged 3.31  ±  0.40 kg for the instrumented foetuses and did not differ from that of the twins' weight at 3.04  ±  0.85 kg (p  =  0.42, two-tailed t test for unequal variances).

In all 13 animals, we select an 1 min interval of the highest quality for the analysis. Our single lead algorithm failed to extract any fetal ECG from the maternal xyphoid ECG signal. The F1 is 14.53  ±  7.28% (range: 0–25%). TS_C and TS_PCA also failed to extract the fetal ECG, with the F1 being 12.83  ±  6.78% (range: 0–21.6%) and 11.59  ±  6.98% (range: 0–22.54%). See table 4 for the case-by-case result. We found no difference in the performance of the algorithm dependent on the parity status (singleton versus non-singleton, two-sided rank sum test, p  =  0.82 for the proposed method, p  =  0.49 for TS_C, and p  =  0.85 for TS_PCA). See figure 2 for a typical failure example of the sheep fetal ECG derived from the maternal xyphoid ECG by the proposed algorithm. Clearly, no ECG pattern could be visualized in the rough fetal ECG.

Zoom In Zoom Out Reset image size

This paper represents an attempt of back-translation, bed-to-bench, of a long-standing tradition of deploying abdominal ECG to extract the fetal ECG signal non-invasively (Silva et al 2013). While the standard experimental design in fetal sheep model presumes installing ECG leads on the maternal xyphoid process, it is also feasible to install the ECG leads in the abdominal area, around the incision site where the uterotomy is performed to operate on the fetus. Future studies should focus on deploying abdominal ECG leads in pregnant ewes to acquire fECG non-invasively.

Our results do not support the initial hypothesis that maternal xyphoid ECG permits the extraction of fetal ECG signal by the proposed algorithm and other widely applied single lead fetal ECG extraction algorithms TS_C and TS_PCA, even when the lead is close to the uterus. There are several possible mutually non-exclusive reasons for this. First, the applied fECG extraction algorithm might not be suitable for the maternal xyphoid ECG signal. Second, the maternal xyphoid ECG signal does not contain strong enough fECG signal, although the lead is close to the mother's abdomen. Third, the sampling resolution of the data acquisition system, at 16 bits, may not suffice to capture the weaker fetal ECG signal when captured by aECG placed as far as over the maternal xyphoid. Specifically, while it is possible to sense the fetal ECG with the acquisition system at 16 bits amplitude resolution when the ECG electrode is placed on the maternal abdomen, in this study the ECG electrode was placed over the xyphoid process and the fetal ECG signal was weaker. We thus do not rule out the possibility that we can get fetal ECG from the maternal xyphoid process at higher ECG amplitude resolution. Further studies are needed to test these assumptions. Such studies could vary the following settings: testing other ECG extraction algorithms such as the scalogram approach (Khamene and Negahdaripour 2000), the phase space approach (Karvounis et al 2007), the sequential total variation denoising approach (Lee and Lee 2016), the adaptive noise cancellation (Zhang et al 2017), and the extended Kalman smoother (Panigrahy and Sahu 2017); placing maternal ECG over the abdomen instead of over the xyphoid process and using data acquisition systems with amplitude resolution higher than 16 bits while recording ECG from the maternal xyphoid process. From the fetal ECG extraction perspective, note that if we have more than one ECG channel close to maternal abdomen, the blind source separation algorithms could be applied; or if we have one more ECG channel recorded from maternal thorax, we could apply adaptive filter techniques to extract the fetal ECG (Andreotti et al 2016). However, in this study we only have single ECG signal recorded over the maternal xyphoid process, so these techniques could not be applied.

We have made this data set publicly available for those interested in fetal ECG extraction technology and invite fellow sheep model scientists to share their maternal/fetal ECG data sets in the efforts to enable a non-invasive fetal ECG acquisition in this crucial model of human physiology.

Ultimately, we hope to have highlighted the potential of launching such maternal-fetal ECG database as a benchmark for the fetal HRV analysis with high translational potential into clinical practice. The fetal sheep model is ideally suited for this endeavor, because the fetal precordial ECG quality is excellent and serves as the ground truth while permitting long-term recordings for hours to days and even weeks. As animal model, it naturally permits a more extensive control of physiological parameters than possible in a clinical setting.

Supported by CIHR and FRQS (MGF) and Sloan (HTW)