Clinically accurate fetal ECG parameters acquired from maternal abdominal sensors

Abstract

OBJECTIVE

To evaluate the accuracy of a novel system for measuring fetal heart rate and ST-segment changes using non-invasive electrodes on the maternal abdomen.

STUDY DESIGN

Fetal ECGs were recorded using abdominal sensors from 32 term laboring women who had a fetal scalp electrode (FSE) placed for a clinical indication.

RESULTS

Good quality data for FHR estimation was available in 91.2% of the FSE segments, and 89.9% of the abdominal electrode segments. The root mean square (RMS) error between the FHR data calculated by both methods over all processed segments was 0.36 beats per minute. ST deviation from the isoelectric point ranged from 0 to 14.2% of R-wave amplitude. The RMS error between the ST change calculated by both methods averaged over all processed segments was 3.2%.

CONCLUSION

FHR and ST change acquired from the maternal abdomen is highly accurate and on average is clinically indistinguishable from FHR and ST change calculated using FSE data.

Introduction

Continuous fetal heart rate monitoring during labor is utilized in over 85% of labor episodes in the United States, and represents the standard of care, [1] although there is scant evidence to demonstrate that the use of the technology improves newborn or maternal outcomes. [2] Encouraging data demonstrate that intrapartum fetal ECG analysis can reduce newborn acidemia, hypoxic ischemic encephalopathy, [3] and cesarean deliveries. [4] However, the only clinically-available device for fetal ECG analysis – the STAN monitor from Neoventa -- requires an invasive fetal scalp electrode, limiting its use to a subset of pregnant women who are laboring with ruptured membranes and a dilated cervix.

The potential utility of non-invasive fetal ECG (fECG) for fetal evaluation is significant. However, to-date, there has been no systematic study proving that fetal ECG can be extracted non-invasively without distorting important clinical parameters such as the ST-segment. Prior reports have shown the capacity to measure fetal heart rate using electrodes on the maternal abdomen, but none have demonstrated the capacity to accurately record the fECG waveform with sufficient fidelity to evaluate the morphology [5]. Recently we developed a novel real-time signal processing approach for extracting fetal ECG that mitigates many of the issues involved in extracting an accurate and clinically relevant ECG. This study evaluates the performance of our technique in extracting FHR variations and ST-levels from laboring patients and compares them to invasive scalp electrode data.

Methods

We recorded data from 32 term laboring women who had a fetal scalp electrode placed for a clinical indication and consented to participate in this study. Enrollment occurred sequentially, and there were no exclusion criteria. The study was conducted during the first and second stages of labor. Demographic information about the studied subjects is summarized in Table 1.

Table 1.

Demographic data of the studied subjects

Demographics (n=32)	Median	Range
Maternal age at delivery (years)	32	19–40
Gestational age at delivery (weeks)	40	35–41
Non-white mother (%)	41
Multiparous (%)	47
Birthweight (grams)	3459	1840–4110
Female baby (%)	50
1-minute Apgar score	8	2–9
5-minute Apgar score	9	8–9
Body Mass Index	30.4	21.7–45.1
Epidural during study (%)	97
Pushing during study (%)	9.4

Data were recorded using an E-TROLZ physiologic monitoring platform, which samples 32 channels at 1 kHz. Standard gel-adhesive EKG electrodes (Red-Dot^™ by 3M^™) were used in a standard configuration developed to maximize the chances of having electrodes adjacent to the fetal heart. The general configuration of the abdominal electrodes is illustrated in figure 1. Electrodes were placed based on anatomic landmarks (the umbilicus, xyphoid process, pubic symphisis, axilla, and spine are used to locate electrodes), and as a consequence, the distance between electrodes varied with the maternal abdominal girth. The number of electrodes was arbitrarily chosen based on the capacity of the recording device, and allows for excellent coverage of the maternal abdomen, sides, and back. The specific location of the electrodes is unimportant, as the analysis is done based on the physiologic signal alone without consideration to the location of each electrode. No patient skin preparation was done prior to electrode placement. Both the abdominal ECG data and the scalp ECG were preprocessed for removal of interference from maternal ECG as well as power-line contamination and other sources of background electrical noise, including maternal muscle artifact.

Figure 1.

Locations of abdominal electrodes and corresponding signal quality on each electrode (red high, black low).

Our fetal extraction method was then applied to the abdominal data. For the extracted and filtered abdominal data, each beat was located using a standard QRS detector.[6] Each beat was segmented ±20ms around the fiducial point (R-peak). FHR was calculated from the reciprocal of the median RR interval scaled by a factor of 60. Median FHRs were calculated from 1012 10-second segments from the processed abdominal data and gold standard scalp fECG. Data were recorded for between 9 and 28 minutes from each subject, and we analyzed the first 10 seconds of every 30-second epoch of data for all 32 subjects. As a consequence, between 17 and 56 10-second segments were analyzed for each subject. Data were analyzed after delivery had occurred, and no research data were available to clinicians managing the subjects’ labor.

We used a new method of extracting fECG from a mixture of maternal ECG (mECG) and fECG that uses a priori information concerning the cardiac signals, including their pseudo-periodic structure, to improve the performance of existing techniques and to design novel filtering techniques specific to fetal cardiac signal. By using a realistic model of an individual fetus’ ECG [7–9] which is tracked over time using a Kalman Filter framework, [10] a low-distortion representation of the maternal beat could be extracted from the fECG/mECG mixture. [8–14] Although the Kalman filter and its extensions are now considered to be classic tools in the signal processing community, our approach involves customizing these techniques for fetal cardiac signals, to develop realistic dynamic models that are able to follow the temporal variations of the fECG in highly noisy environments. We have extended these frameworks and developed novel methods for tracking nonstationarities in both the noise and signal. Unlike other computational approaches, this method successfully cancels mECG with minimal distortion (in the relevant clinical parameters) of the fECG.

The color coding in figure 1 is an example that illustrates the signal quality from a given set of electrodes for one patient at a particular instant in time. The signal quality will vary across patients (and over time for a particular patient). This is often due to changes in maternal or fetal position and hence not all sensors are used at any given point in time. However, it is difficult to predict which sensor locations are optimal in advance. We therefore used an over-complete set of electrodes and a series of signal quality measures to determine automatically which sensors contribute most of the information at any given point in time.

Isoelectric-levels and ST-levels were estimated from a subset of 271 10-second segments from the same data. Adult ST analysis was performed on an average window, rather than on individual beats, [15–17] and therefore an average beat was calculated for each 10-second segment for use as an analysis template. Each beat was also cross correlated with the template. Beats with a correlation with the template of less than 0.9 were rejected. If the number of beats left was less than 15 beats, or more than 40% of the beats were rejected, the segment was rejected. Otherwise, the remaining beats were re-averaged. ST analysis was then performed using techniques previously described, [16] except that the thresholds [15, 17] were scaled to allow for differences between fetal and adult beats. These criteria were applied independently on each channel of data, so no information from the scalp electrode was used to alter the abdominal data or select abdominal data for comparison.

The ST segment amplitude and the isoelectric level were computed as the median of signal segments of length 20 ms surrounding the J-point and the isoelectric segment in order to avoid measurement jitter due to amplitude scatter of the original signal samples. To ensure objectivity in the analysis the abdominal fetal separation algorithm was implemented by one author and the ST segment analysis algorithms were developed and run independently by another author without collaboration.

This study was approved by the institutional review boards at our institutions.

Results

In Figure 2, the histogram of the average fetal heart rates calculated from 1012 10-second segments is shown. For these data, the minimum, maximum, and average FHR were 91.4 beats per minute (BPM), 187.8 BPM, and 141.6 BPM, respectively, with a standard deviation (σ) of 15.4 BPM. This range covers a broad range of typical fetal FHRs.[5]

Figure 2.

The histogram of the heart rates of the studied fetal ECGs. Red line illustrates a Gaussian curve with same mean and σ as empirical distribution. (Number of ECG segments used in histogram=1012.)

Data of sufficient quality for FHR estimation was available in 91.2% of the 10-second FSE segments, and 89.9% of the 10-second abdominal electrode segments. The average (root mean square) error between the FHR data calculated by both methods over all processed segments was 0.36 BPM. ST elevation from the isoelectric level ranged from 0 to 10.6% of R-wave amplitude. ST depression from the isoelectric level ranged from 0 to 14.2% of R-wave amplitude. The root mean square error between the ST changes calculated by both methods averaged over all processed segments was 3.2% and the mean absolute difference was +0.4%, indicating a very small positive bias.

There was no relationship between the subjects’ body mass index and the fidelity of the fetal ECG waveform that was extracted.

A typical ECG waveform can be seen in figure 3. Figure 4 illustrates the extracted fECG in figure 1b after post-processing together with its 68% (±σ) and 99.7% (±3σ) confidence intervals (signal envelopes). Figure 5 illustrates a comparison between 10 second averaged heart-beats from the invasive scalp electrode and the extracted fECG from the abdominal electrodes (without using the fetal scalp electrode). Note the similarity in morphology, including the isoelectric-levels, the ST-level and the PR-interval. The average correlation between the fECG extracted from the abdominal and scalp leads was 0.96 over a complete beat and 0.69 over the ST segment.

Figure 3.

Figure 3a and 3b. Segment of abdominal signals, (a) before, and (b) after maternal ECG subtraction

Figure 4.

Several fECG beats (from figure 3b), before (grey lines) and after (black line) the post-processing using our Kalman filter approach, together with the 68% (±σ) and 99.7% (±3σ) confidence intervals (upper and lower dashed envelopes).

Figure 5.

A typical fetal heart beat recorded from the scalp electrode (in magenta) and overlaid extracted abdominal fetal ECG (in blue).

Figure 6 presents the results of the median FHR estimations from the scalp and abdominal electrodes. When plotted as a function of each other, they clearly demonstrate a strong correlation, with almost every point lying on the line of identity. To quantify this observation, the Kolmogorov-Smirnov (KS) goodness-of-fit hypothesis test was performed on the scalp and abdominal heart rate time-series. The null hypothesis, i.e., the hypothesis that the scalp and abdominal heart rate time-series of 10s segments have identical distributions, was rejected in only 5.5% of the studies segments (p<0.01).

Figure 6.

Comparison of the median FHR calculated from the scalp vs. abdominal electrode for 10s segments. (Number of ECG segments used in histogram=1012.)

Comment

Our results indicate that extraction of non-invasive fetal ECG without distorting clinical parameters is possible using a novel signal processing approach. FHR variability and ST-deviation from ECG acquired from the maternal abdomen can be estimated, which are clinically indistinguishable from FHR variability and ST-deviation derived from the fetal scalp electrode in our patient population.

This report is the first to compare the fetal ECG waveform measured using non-invasive electrodes to the ECG waveform measured using a FSE. To the extent that the FSE represents the gold standard in fetal monitoring, both for fECG and for fetal heart-rate monitoring, this is an important evaluation of our technique’s accuracy.

In 10.1% of abdominal ECG segments evaluated, we failed to extract a useful ECG waveform. However, the current study was conducted on data recorded during labor but analyzed after delivery. Therefore, human errors such as sensor misplacements, sensor detachments or loose connections have been accumulated in the 10% figure. In the future, real-time analysis of data with immediate display of signal quality and ECG waveform will allow for the clinician to make minor adjustments that will increase accuracy of the technology over the 89.9% rate reported here.

Due to the small number of patients studied and the fortunate absence of intrapartum hypoxic or ischemic events, we were unable to watch the change in the ST-segment during periods of fetal ischemia. Because of this, we do not demonstrate that our algorithm maintains ST-segment fidelity between the FSE and the abdominal leads during conditions of hypoxia or ischemia. Larger clinical studies or animal research will be needed to confirm the fidelity of our algorithms under all clinical circumstances.

There are no guidelines as to exactly how much of an elevation or depression would be clinically significant in a fetal population. Figure 7 presents the distribution of the differences in ST elevation or depression between the scalp and abdominal extracted data. Note that the largest difference was less than 14.5%. This difference can be explained due to the morphological difference of the fECG extracted from the scalp and abdominal leads, similar to different adult ECG morphologies. Moreover, considering that ST segment elevation requires a continued elevation for several epochs, the actual ST level sensitivities are expected to be better than the results presented in this study.

Figure 7.

Empirical distribution of differences in ST deviation estimated from extracted abdominal and scalp electrodes. (Number of ECG segments used in histogram=271.)

Future research will extend these analyses to larger samples of patients at high risk for ischemia as well as to animal models where fetal oxygen content and umbilical vascular flow can be modulated.

Although our methods performed well in the second stage of labor and among obese women, additional studies that include larger numbers of women in these clinical situations that challenge conventional monitoring technology will be needed to extend our results.

Limitations of this study include the potential for selection bias. Only women who had a fetal scalp electrode placed for a clinical indication were included in the study, and only women who consented to participate in the study were included. Additionally, nearly all women in the study had an epidural in place when data were collected. It may be that the epidural diminished patient activity and made it easier to extract the fetal ECG signal. Additional research will be needed to determine whether our technique performs as well in a group of women without regional anesthesia as it does in the group of subjects presented here.