iScdoct(dt

Research www.AJOG.org

OBSTETRICS

Clinically accurate fetal ECG parameters acquiredfrom maternal abdominal sensorsGari Clifford, PhD; Reza Sameni, PhD; Jay Ward; Julian Robinson, MBChB; Adam J. Wolfberg, MD

OBJECTIVE: We sought to evaluate the accuracy of a novel system formeasuring fetal heart rate (FHR) and ST-segment changes using nonin-vasive electrodes on the maternal abdomen.

STUDY DESIGN: Fetal electrocardiograms were recorded using ab-dominal sensors from 32 term laboring women who had a fetal scalpelectrode (FSE) placed for a clinical indication.

RESULTS: Good-quality data for FHR estimation were available in91.2% of the FSE segments and 89.9% of the abdominal electrode

segments. The root mean square error between the FHR data calcu-

2011;205:47.e1-5.

0002-9378/$36.00 • © 2011 Mosby, Inc. All rights reserved. • doi: 10.1016

lated by both methods over all processed segments was 0.36 beats perminute. ST deviation from the isoelectric point ranged from 0–14.2% ofR-wave amplitude. The root mean square error between the ST changecalculated by both methods averaged over all processed segments was3.2%.

CONCLUSION: FHR and ST change acquired from the maternal abdo-men is highly accurate and, on average, is clinically indistinguishablefrom FHR and ST change calculated using FSE data.

Key words: fetal electrocardiogram, fetal monitoring, labor

Cite this article as: Clifford G, Sameni R, Ward J, et al. Clinically accurate fetal ECG parameters acquired from maternal abdominal sensors. Am J Obstet Gynecol

oageestact

Continuous fetal heart rate (FHR)monitoring during labor is utilized

n �85% of labor episodes in the Unitedtates and represents the standard ofare,1 although there is scant evidence toemonstrate that the use of the technol-gy improves newborn or maternal out-omes.2 Encouraging data demonstratehat intrapartum fetal electrocardiogramfECG) analysis can reduce newborn aci-emia, hypoxic ischemic encephalopa-hy,3 and cesarean deliveries.4 How-

ever, the only clinically available devicefor fECG analysis—the STAN monitorfrom Neoventa (Moindal, Sweden)—re-

From the Department of Engineering ScienceUK (Dr Clifford), and Department of ElectricMassachusetts Institute of Technology, CambClifford); Mindchild Medical Inc (Drs ClifforE-TROLZ Inc (Mr Ward), North Andover, MRobinson), MA; and Tufts Medical Center, BoUniversity, Shiraz, Iran (Dr Sameni).

Presented at the 30th Annual Meeting of the SoFeb. 1-6, 2010.

Received Dec. 7, 2010; revised Feb. 4, 2011; ac

Reprints not available from the authors.

This study was supported by The Safe Fetus ProjeTechnology, Boston, MA; Obstetrix Inc, Sunrise, FDevelopment (NICHD) (Noninvasive Observation oInnovation Research Grant no. 1R43HD059263-01K23HD060662-01A1).

quires an invasive fetal scalp electrode(FSE), limiting its use to a subsetof pregnant women who are laboringwith ruptured membranes and a dilatedcervix.

The potential utility of noninvasivefECG for fetal evaluation is significant.However, to date, there has been no sys-tematic study proving that fECG can beextracted noninvasively without distort-ing important clinical parameters, suchas the ST segment. Prior reports haveshown the capacity to measure FHR us-ing electrodes on the maternal abdomen,but none have demonstrated the capac-

e University of Oxford, Oxford, England,ngineering and Computer Science,ge, and Harvard University, Boston (Drameni, and Wolfberg and Mr Ward), and

ewton-Wellesley Hospital, Newton (Drn (Dr Wolfberg), MA, and Shiraz

y for Maternal-Fetal Medicine, Chicago, IL,

ted Feb. 24, 2011.

Center for Integration of Medicine and Innovativeational Institute of Child Health and Humantal Activity Garment Project, Small Business); and NICHD (The Safe Fetus Project, Grant no.

/j.ajog.2011.02.066

JULY 2011 Ameri

ity to accurately record the fECG wave-form with sufficient fidelity to evaluatethe morphology.5 Recently, we devel-

ped a novel real-time signal processingpproach for extracting fECG that miti-ates many of the issues involved inxtracting an accurate and clinically rel-vant electrocardiogram (ECG). Thistudy evaluates the performance of ourechnique in extracting FHR variationsnd ST levels from laboring patients andompares them to invasive scalp elec-rode data.

MATERIALS AND METHODSWe recorded data from 32 term laboringwomen who had an FSE placed for a clin-ical indication and consented to partici-pate in this study. Enrollment occurred se-quentially, and there were no exclusioncriteria. The study was conducted duringthe first and second stages of labor. Demo-graphic information about the studiedsubjects is summarized in the Table.

Data were recorded using an E-TROLZ(North Andover, MA) physiologic monitor-ing platform, which samples 32 channels at1 kHz. Standard gel-adhesive ECG elec-trodes (Red-Dot; 3M, St. Paul, MN) wereused in a standard configuration devel-oped to maximize the chances of havingelectrodes adjacent to the fetal heart. The

, thal Eridd, SA; Nsto

ciet

cep

ct,L; Nf Na1A1

general configuration of the abdominal

can Journal of Obstetrics & Gynecology 47.e1

ultf

wm

Research Obstetrics www.AJOG.org

electrodes is illustrated in Figure 1. Elec-trodes were placed based on anatomiclandmarks (the umbilicus, xyphoid pro-cess, pubic symphysis, axilla, and spineare used to locate electrodes), and as aconsequence, the distance between elec-trodes varied with the maternal abdom-inal girth. The number of electrodes was

FIGURE 1Abdominal electrode location

Locations of abdominal electrodes and corre-sponding signal quality on each electrode (red �high, black � low).Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

TABLEDemographic dataof studied subjects

Demographics(n � 32) Median Range

Maternal age atdelivery, y

32 19–40

...........................................................................................................

Gestational ageat delivery, wk

40 35–41

...........................................................................................................

Nonwhitemother, %

41

...........................................................................................................

Multiparous, % 47...........................................................................................................

Birthweight, g 3459 1840–4110...........................................................................................................

Female baby, % 50...........................................................................................................

1-min Apgarscore

8 2–9

...........................................................................................................

5-min Apgarscore

9 8–9

...........................................................................................................

Body massindex

30.4 21.7–45.1

...........................................................................................................

Epidural duringstudy, %

97

...........................................................................................................

Pushing duringstudy, %

9.4

...........................................................................................................

Clifford. Noninvasive fetal ECG. Am J ObstetGynecol 2011.

47.e2 American Journal of Obstetrics & Gynecolog

arbitrarily chosen based on the capacityof the recording device, and allows forexcellent coverage of the maternal abdo-men, sides, and back. The specific loca-tion of the electrodes is unimportant, asthe analysis is done based on the physio-logic signal alone, without considerationto the location of each electrode. No pa-tient skin preparation was done prior toelectrode placement. Both the abdomi-nal ECG data and the scalp ECG werepreprocessed for removal of interferencefrom maternal ECG (mECG), as well aspower line contamination and othersources of background electrical noise,including maternal muscle artifact.

Our fetal extraction method was thenapplied to the abdominal data. For theextracted and filtered abdominal data,each beat was located using a standardQRS detector.6 Each beat was segmented�20 milliseconds around the fiducialpoint (R-peak). FHR was calculatedfrom the reciprocal of the median RR in-terval scaled by a factor of 60. MedianFHRs were calculated from 1012 10-sec-ond segments from the processed ab-dominal data and gold standard scalpfECG. Data were recorded for between9 –28 minutes from each subject, and weanalyzed the first 10 seconds of every 30-second epoch of data for all 32 subjects.As a consequence, between 17–56 10-second segments were analyzed for eachsubject. Data were analyzed after deliv-ery had occurred, and no research datawere available to clinicians managing thesubjects’ labor.

We used a new method of extractingfECG from a mixture of mECG andfECG that uses a priori information con-cerning the cardiac signals, includingtheir pseudoperiodic structure, to im-prove the performance of existing tech-niques and to design novel filtering tech-niques specific to fetal cardiac signal. Byusing a realistic model of an individualfetus’ ECG,7-9 which is tracked over time

sing a Kalman filter framework,10 aow-distortion representation of the ma-ernal beat could be extracted from theECG/mECG mixture.8-14 Although the

Kalman filter and its extensions are nowconsidered to be classic tools in the signalprocessing community, our approach

involves customizing these techniques E

y JULY 2011

for fetal cardiac signals, to develop real-istic dynamic models that are able to fol-low the temporal variations of the fECGin highly noisy environments. We haveextended these frameworks and devel-oped novel methods for tracking nonsta-tionarities in both the noise and signal.Unlike other computational approaches,this method successfully cancels mECG,with minimal distortion (in the relevantclinical parameters) of the fECG.

The color coding in Figure 1 is an ex-ample that illustrates the signal qualityfrom a given set of electrodes for 1 pa-tient at a particular instant in time. Thesignal quality will vary across patients(and over time for a particular patient).This is often due to changes in maternalor fetal position and hence, not all sen-sors are used at any given point in time.However, it is difficult to predict whichsensor locations are optimal in advance.We therefore, used an over-complete setof electrodes and a series of signal qualitymeasures to determine automaticallywhich sensors contribute most of the in-formation at any given point in time.

Isoelectric levels and ST levels were es-timated from a subset of 271 10-secondsegments from the same data. Adult STanalysis was performed on an averagewindow, rather than on individualbeats,15-17 and therefore, an average beat

as calculated for each 10-second seg-ent for use as an analysis template.

FIGURE 2Histogram of heart ratesof studied fetalelectrocardiograms

(1012 electrocardiograms segments used in his-togram.)Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

ach beat was also cross correlated with

sp

b

www.AJOG.org Obstetrics Research

the template. Beats with a correlationwith the template of �0.9 were rejected.If the number of beats left was �15 beats,or �40% of the beats were rejected, thesegment was rejected. Otherwise, the re-maining beats were reaveraged. ST analy-is was then performed using techniquesreviously described,16 except that the

thresholds15,17 were scaled to allow for dif-ferences between fetal and adult beats.These criteria were applied independentlyon each channel of data, so no informationfrom the scalp electrode was used to alterthe abdominal data or select abdominaldata for comparison.

The ST-segment amplitude and theisoelectric level were computed as themedian of signal segments of length 20milliseconds surrounding the J-pointand the isoelectric segment to avoidmeasurement jitter due to amplitudescatter of the original signal samples. Toensure objectivity in the analysis, the ab-dominal fetal separation algorithm wasimplemented by 1 author and the ST-segment analysis algorithms were devel-oped and run independently by anotherauthor without collaboration.

This study was approved by the institu-tional review boards at our institutions.

RESULTS

In Figure 2, the histogram of the averageFHRs calculated from 1012, 10-secondsegments is shown. For these data, theminimum, maximum, and average FHRwere 91.4 beats per minute (BPM), 187.8BPM, and 141.6 BPM, respectively, withSD (�) of 15.4 BPM. This range covers a

road range of typical FHRs.5

Data of sufficient quality for FHR esti-mation were 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 be-tween the FHR data calculated by bothmethods over all processed segments was0.36 BPM. ST elevation from the isoelec-tric level ranged from 0–10.6% of R-waveamplitude. ST depression from the iso-electric level ranged from 0 –14.2% of R-wave amplitude. The root mean squareerror between the ST change calculatedby both methods averaged over all pro-

cessed segments was 3.2% and the mean

FIGURE 3ECG waveforms extracted from abdomen

Segment of abdominal signals A, before and B, after maternal electrocardiogram (ECG) subtraction.Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

FIGURE 4Extracted fetal ECG waveforms

Several fetal electrocardiogram beats (from Figure 3, B), before (gray lines) and after (black line)postprocessing using our Kalman filter approach, together with 68% (��) and 99.7% (�3�) con-fidence intervals (upper and lower dashed envelopes).ECG, electrocardiogram.

Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

JULY 2011 American Journal of Obstetrics & Gynecology 47.e3

9

Research Obstetrics www.AJOG.org

absolute difference was �0.4%, indicat-ing a very small positive bias.

There was no relationship between thesubjects’ body mass index and the fidelity ofthe fECG waveform that was extracted.

A typical ECG waveform can be seenin Figure 3. Figure 4 illustrates the ex-tracted fECG in Figure 3, B after postpro-cessing together with its 68% (��) and

9.7% (�3�) confidence intervals (signalenvelopes). Figure 5 illustrates a compari-son between 10-second averaged heart-beats from the invasive scalp electrode andthe extracted fECG from the abdominalelectrodes (without using the FSE). Notethe similarity in morphology, includingthe isoelectric levels, theSTlevel,andthePRinterval.TheaveragecorrelationbetweenthefECG extracted from the abdominal andscalp leads was 0.96 over a complete beat and0.69 over the ST segment.

Figure 6 presents the results of themedian FHR estimations from thescalp and abdominal electrodes. Whenplotted as a function of each other, theyclearly demonstrate a strong correla-

FIGURE 5Typical fetal ECG waveform

Typical fetal heart beat recorded from scalp elefetal electrocardiogram (ECG) (blue).Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

tion, with almost every point lying on

47.e4 American Journal of Obstetrics & Gynecolog

the line of identity. To quantify this ob-servation, the Kolmogorov-Smirnovgoodness-of-fit hypothesis test wasperformed on the scalp and abdominalheart rate time series. The null hypoth-esis, ie, the hypothesis that the scalpand abdominal heart rate time series of10-second segments have identical dis-tributions, was rejected in only 5.5% ofthe study segments (P � .01).

COMMENTOur results indicate that extraction of non-invasive fECG without distorting clinicalparameters is possible using a novel signalprocessing approach. FHR variability andST deviation from ECG acquired fromthe maternal abdomen can be estimated,which are clinically indistinguishable fromFHR variability and ST deviation derivedfrom the FSE in our patient population.

This report is the first to compare thefECG waveform measured using nonin-vasive electrodes to the ECG waveformmeasured using FSE. To the extent that

de (magenta) and overlaid extracted abdominal

the FSE represents the gold standard in

y JULY 2011

fetal monitoring, both for fECG and forFHR monitoring, this is an importantevaluation of our technique’s accuracy.

In 10.1% of abdominal ECG segmentsevaluated, we failed to extract a usefulECG waveform. However, the currentstudy was conducted on data recordedduring labor but analyzed after delivery.Therefore, human errors, such as sensormisplacements, sensor detachments, orloose connections, have been accumu-lated in the 10% figure. In the future,real-time analysis of data, with immedi-ate display of signal quality and ECGwaveform, will allow for the clinician tomake minor adjustments that will in-crease accuracy of the technology overthe 89.9% rate reported here.

Due to the small number of patientsstudied and the fortunate absence of intra-partum hypoxic or ischemic events, wewere unable to watch the change in the STsegment during periods of fetal ischemia.Because of this, we do not demonstratethat our algorithm maintains ST-segmentfidelity between the FSE and the abdomi-nal leads during conditions of hypoxia orischemia. Larger clinical studies or animalresearch will be needed to confirm the fi-delity of our algorithms under all clinical

FIGURE 6Extracted fetal heart rate

Comparison of median fetal heart rate calculatedfrom scalp vs abdominal electrode for 10-secondsegments. (1012 electrocardiogram segmentsused in histogram.)BPM, beats per minute.

Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

ctro

circumstances.

eeaasqoes

sfwv

wat

tcw

pwdopthcmetmwgo

www.AJOG.org Obstetrics Research

There are no guidelines as to exactly howmuch of an elevation or depression wouldbe clinically significant in a fetal popula-tion. Figure 7 presents the distribution ofthe differences in ST elevation or depres-sion between the scalp and abdominal ex-tracted data. Note that the largest differ-ence was �14.5%. This difference can bexplained due to the morphological differ-nce of the fECG extracted from the scalpnd abdominal leads, similar to differentdult ECG morphologies. Moreover, con-idering that ST-segment elevation re-uires a continued elevation for several ep-chs, the actual ST-level sensitivities arexpected to be better than the results pre-ented in this study.

Future research will extend these analy-es to larger samples of patients at high riskor ischemia, as well as to animal modelshere fetal oxygen content and umbilicalascular flow can be modulated.

Although our methods performedell in the second stage of labor and

mong obese women, additional studies

FIGURE 7Fetal ST deviation

Empirical distribution of differences in ST devia-tion estimated from extracted abdominal andscalp electrodes. (271 electrocardiogram seg-ments used in histogram.)Clifford. Noninvasive fetal ECG. Am J Obstet Gynecol 2011.

hat include larger numbers of women in

hese clinical situations that challengeonventional monitoring technology,ill be needed to extend our results.Limitations of this study include the

otential for selection bias. Only womenho had an FSE placed for a clinical in-ication were included in the study, andnly women who consented to partici-ate in the study were included. Addi-ionally, nearly all women in the studyad an epidural in place when data wereollected. It may be that the epidural di-inished patient activity and made it

asier to extract the fECG signal. Addi-ional research will be needed to deter-

ine whether our technique performs asell in a group of women without re-ional anesthesia as it does in the groupf subjects presented here.

ACKNOWLEDGMENTSWe would like to thank Profs Roger Mark andFranc Jager for helpful discussions and insight.

REFERENCES1. American College of Obstetricians and Gy-necologists. ACOG practice bulletin no. 106:intrapartum fetal heart rate monitoring: no-menclature, interpretation, and general man-agement principles. Obstet Gynecol 2009;114:192-202.2. Vintzileos AM, Nochimson DJ, Guzman ER,Knuppel RA, Lake M, Schifrin BS. Intrapartumelectronic fetal heart rate monitoring versus in-termittent auscultation: a meta-analysis. ObstetGynecol 1995;85:149-55.3. Amer-Wahlin I, Bordahl P, Eikeland T, et al. STanalysis of the fetal electrocardiogram during la-bor: Nordic observational multicenter study. JMatern Fetal Neonatal Med 2002;12:260-6.4. Noren H, Blad S, Carlsson A, et al. STAN inclinical practice—the outcome of 2 years of reg-ular use in the city of Gothenburg. Am J ObstetGynecol 2006;195:7-15.5. Sameni R, Clifford GD. A review of fetal ECG sig-nal processing; issues and promising directions.Open Pacing Electrophysiol Ther J 2010;3:4-20.6. Hamilton PS, Tompkins WJ. Quantitative in-

vestigation of QRS detection rules using the

JULY 2011 Ameri

MIT/BIH arrhythmia database. IEEE TransBiomed Eng 1986;33:1157-65.7. McSharry PE, Clifford GD, Tarassenko L,Smith LA. A dynamical model for generatingsynthetic electrocardiogram signals. IEEE TransBiomed Eng 2003;50:289-94.8. Clifford GD. A novel framework for signal rep-resentation and source separation: applicationsto filtering and segmentation of biosignals.J Biol Syst 2006;14:169-83.9. Clifford GD, Shoeb A, McSharry PE, JanzBA. Model-based filtering, compression andclassification of the ECG. Int J Bioelectromagn2005;7:158-61.10. Sameni R, Shamsollahi MB, Jutten C, Clif-ford GD. A nonlinear Bayesian filtering frame-work for ECG denoising. IEEE Trans BiomedEng 2007;54:2172-85.11. Sameni R, Vrins F, Parmentier F, et al.Electrode selection for noninvasive fetal elec-trocardiogram extraction using mutual infor-mation criteria. Proceedings of the 26th Inter-national Workshop on Bayesian Inference andMaximum Entropy Methods in Science andEngineering (MaxEnt 2006). Vol 872. Paris,France: CNRS; 2006.12. Sameni R, Jutten C, Shamsollahi MB. Mul-tichannel electrocardiogram decomposition us-ing periodic component analysis. IEEE TransBiomed Eng 2008;55:1935-40.13. Sameni R, Clifford GD, Jutten C, Sham-sollahi M. Multichannel ECG and noise mod-eling: application to maternal and fetal ECGsignals. EURASIP J Adv Signal Process2006;2007:1-14.14. Sameni R. Extraction of fetal cardiac signalsfrom an array of maternal abdominal recordings[doctoral thesis]. Grenoble, France: ElectricalEngineering and Computer Science, InstitutNational Polytechnique; 2008.15. Jager F. Automated detection of transientST-segment changes during ambulatory ECGmonitoring [doctoral dissertation]. Ljubljana,Slovenia: Faculty of Electrical and ComputerEngineering, University of Ljubljana; 1994.16. Jager F, Moody GB, Mark RG. Detection oftransient ST-segment episodes during ambula-tory ECG monitoring. Comput Biomed Res 1998;31:305-22.17. Jager F, Taddei A, Moody GB, et al. Long-term ST database: a reference for the devel-opment and evaluation of automated isch-emia detectors and for the study of thedynamics of myocardial ischemia. Med Biol

Eng Comp 2003;41:172-82.

can Journal of Obstetrics & Gynecology 47.e5