Selective heart rate variability analysis to account for uterine activity
during labor and improve classification of fetal distress
G.J.J. Warmerdam1 , R. Vullings1 , J.O.E.H. Van Laar2 , M.B. Van der Hout-Van der Jagt2 , J.W.M. Bergmans1 ,
L. Schmitt3 , and S.G. Oei 2
Abstract— Cardiotocography (CTG) is currently the most of-
ten used technique for detection of fetal distress. Unfortunately,
CTG has a poor specificity. Recent studies suggest that, in
addition to CTG, information on fetal distress can be obtained
from analysis of fetal heart rate variability (HRV). However,
uterine contractions can strongly influence fetal HRV. The aim
of this study is therefore to investigate whether HRV analysis
for detection of fetal distress can be improved by distinguishing
contractions from rest periods. Our results from feature selec-
tion indicate that HRV features calculated separately during
contractions or during rest periods are more informative on
fetal distress than HRV features that are calculated over the
entire fetal heart rate. Furthermore, classification performance
improved from a geometric mean of 69.0% to 79.6% when
including the contraction-dependent HRV features, in addition
to HRV features calculated over the entire fetal heart rate.
I. I NTRODUCTION
During labor, oxygen supply to the fetus can temporally
be reduced due to contractions. If the reduction in oxygen
supply is severe and prolonged this can lead to neonatal
morbidity and mortality. In clinical practice, cardiotocogra-
phy (CTG) is currently the most often used technique for
fetal monitoring. However, CTG is interpreted visually and
has a high inter- and intra-observer variability. Besides, the
poor specificity of CTG leads to unnecessary interventions.
Therefore, additional information is required to reduce the
number of unnecessary operative deliveries.
Recent studies have shown interest in analysis of fetal
heart rate variability (HRV) for detection of fetal distress
[1]–[6]. Since the heart rate is regulated by the autonomic
nervous system (ANS), variations in the heart rate reflect
changes in ANS control and could indirectly provide infor-
mation on fetal distress. In the literature, many HRV features
have been developed that describe different properties of the
heart rate. However, most HRV features have been developed
R
for adults. Unlike for adults, it is not possible to control the
fetal cardiovascular system. Moreover, during labor contrac-
tions can strongly influence the fetal cardiovascular system.
In [6] and [7] it was shown that for healthy fetuses, spectral
power is significantly higher during contractions as compared
to rest periods (i.e. the periods in between contractions). The
increased spectral power during contractions could indicate a
1 G.J.J. Warmerdam, R. Vullings, and J.W.M. Bergmans are with the
Faculty of Electrical Engineering, Eindhoven University of Technology,
Eindhoven, The Netherlands g.j.j.warmerdam@tue.nl
2 J.O.E.H. Van Laar,, M.B. Van der Hout-Van der Jagt, and S.G. Oei
are with the Department of Obstetrics and Gynecology, Máxima Medical
Center, Veldhoven, The Netherlands
3 L. Schmitt is with Philips Research, Eindhoven, The Netherlands
978-1-4577-0220-4/16/$31.00 ©2016 IEEE 2950
response of the fetal ANS to stabilize the cardiovascular sys-
tem under fluctuating conditions of the exterior (i.e. uterus).
If HRV features are calculated without distinguishing periods
of contractions from rest periods, this effect is averaged over
the rest periods and is no longer visible.
In this paper we examine whether detection of fetal
distress by HRV analysis can be improved by separating
contractions from rest periods. Therefore, in addition to
HRV features calculated over the entire heart rate, HRV
features were also calculated separately during contractions
and during rest periods. A sequential feature selection al-
gorithm was implemented to determine the most informative
subset of HRV features, both with and without the additional
contraction-dependent HRV features. Using these subsets as
input, performance of classification of healthy fetuses and
fetuses with adverse outcome was evaluated by Support
Vector Machines (SVMs).
First, data acquisition is described in Section II-A. Then,
Sections II-B to II-F describe the signal processing tech-
niques used in this paper: pre-processing, HRV analysis,
HRV analysis during contractions and rest periods, feature
selection, and classification. In section II-G, our validation
method is explained. Finally, results are presented in Section
III and discussion in Section IV.
II. M ETHOD
This section describes the different processing techniques
that were used in this paper. The overall procedure is
presented in Fig. 1.
A. Data acquisition
For this study, fetal heart rate (FHR) and uterine activity
(UA) signals were obtained from CTG registrations that
were recorded on a Neoventa STAN
(Mölndal, Sweden).
We only used beat-to-beat FHR signals that were measured
by a scalp electrode. The UA was recorded either with a
tocodynamometer or intra-uterine pressure catheter. Based
on the UA signal, contractions were annotated by a clinical
expert that had no knowledge of the FHR signal or any other
medical information. An example of a recording is shown in
Fig. 2.
Fetal outcome was defined using the arterial acid-base
status: adverse outcome was defined as a pH below 7.05 and
healthy as a pH above 7.20 [8]. Pregnancies complicated by
intra-uterine growth restriction or fetal congenital anomalies
were excluded. In total, 22 cases were included with adverse
outcome, that were matched with 110 healthy cases. Note
 Pre-processing
Segment selection
HRV analysis
Feature selection
Classification
Validation
Fig. 1. Flowchart of signal processing.
200
FHR (BPM)
150
100
50
100
UA (mmHg)
75
50
25
0
0 2 4 6
time (min)
Fig. 2. Example of a FHR and UA signal for a healthy fetus.
that certain medication can cross the placenta and influence
fetal HRV (e.g. anti-hypertensives or pain killers). We ap-
proximately matched the prevalence of these medications in
both groups, in accordance with the difference in size of the
two groups.
B. Pre-processing
HRV features are calculated on interbeat interval (R-R
interval) signals. Since beat-to-beat R-R intervals can only
be determined at the instance a heartbeat occurs, the R-
R intervals are not equidistantly sampled. Therefore, linear
interpolation was used to obtain an equidistantly sampled
heart rate. When the difference between two consecutive R-
R intervals was more than 200 ms or more than 25 Beats Per
Minute (BPM), we defined both R-R intervals as artifacts.
Artifacts were corrected by linear interpolation.
For each CTG registration, a segment of 10 minutes was
selected with less than 20% loss of FHR signal and at least
2 minutes of FHR signal during contractions and during rest
periods. Segments were selected closest to birth, only during
the second stage of labor (stage of active pushing) and within
a range of 45 minutes before birth.
C. HRV features
For each 10 minute FHR signal, conventional HRV fea-
tures were calculated that describe different properties of
the heart rate. An overview of the used HRV features is
presented in Table I. Due to limited space, we only briefly
summarize the HRV features used in this study. For a more
thorough description of used HRV features, the authors refer
the interested reader to [6].
To describe the statistical properties of the heart rate we
used the standard-deviation (SD) and root mean square
of successive differences (RM SSD) [9]. Since SD is a
measure for the overall variability, it is related to both
sympathetic and parasympathetic activity. RM SSD on the
other hand is mostly sensitive to beat-to-beat variations,
hence reflecting parasympathetic activity. Although SD and
RM SSD are actually defined on the beat-to-beat FHR, we
calculated these features on the resampled FHR so that all
HRV features used in this study were extracted from the
same signal.
Spectral analysis was performed to examine the heart rate
at specific frequency bands that are related to sympathetic
and parasympathetic activity [1]. We used a Continuous
Wavelet Transform with a fifth order symlet wavelet to
calculate the power in the frequency bands [10]. Power
in the low frequency band (LF , 0.04-0.15 Hz) represents
both sympathetic and parasympathetic activity, and power
in the high frequency band (HF , 0.4-1.5 Hz) represents
parasympathetic activity only. Besides absolute LF and HF ,
also normalized frequency powers were calculated (LFn and
8 10 HFn ), by dividing LF and HF by the total power (T P ,
0.04-1.5 Hz).
Heart rate signals often show complex and irregular fluc-
tuations that cannot be explained by spectral analysis. The
presence of these irregularities in the heart rate is seen as
healthy blood pressure control by the ANS. Complexity in
short time signals can be quantified by Sample Entropy
(SampEn) [11]. SampEn estimates the conditional proba-
bility that if patterns are similar for length m − 1, they will
also be similar for length m. A tolerance parameter r is used
to define a threshold for similarity between two patterns. In
this study we used m = 2 and r = 0.2SD, similar to settings
used in [2] and [3].
Finally, Detrended Fluctuation Analysis (DFA) was used
to describe the scaling properties of the HRV over different
time scales [12]. The scaling exponent α was calculated
over a range of 1 to 60 seconds, corresponding to the
physiological range of the baroreflex.
D. HRV analysis during contractions and rest periods
To calculate HRV features separately during contrac-
tions and during rest periods, the expert annotations of the
contractions were used. Rest periods were defined as the
periods in between contractions. If the time between two
contractions was more than one minute, we only considered
the final minute preceding the second contraction as rest
period. Values of HRV features during contractions were
calculated based on R-R values that were recorded during
contractions. Similarly, values of HRV features during rest
were calculated based on R-R intervals that were recorded
during rest periods. For each HRV feature, the average value
of all contraction periods and all rest periods within the 10
minute segments (indicated as f eatureuc and f eaturerest ,
respectively) is then used for classification. We also con-
sidered the ratio between HRV features calculated during
contractions and during rest periods (f eatureratio ).
Note that computation of LF , LFn , HFn , and α requires
approximately one minute of consecutive FHR signal. Since
2951
 TABLE I
HRV FEATURES
R
HRV features entire FHR contractions rest ratio
SD x x x x
RM SSD x x x x
LF x misclassification of both classes would result in a classifier
HF x x x x
TP x that tends to correctly classify the majority class (healthy
LFn x fetuses) but performs poorly on classification of the minority
HFn x class (fetuses with adverse outcome) [13]. To account for
SampEn x x x x
α x
during the second stage contractions and rest periods are
often less than one minute, we could not extract these HRV
features separately during contractions and rest periods. We
thus limited our analysis during contractions and rest periods
to SD, RM SSD, HF , and SampEn.
In total, 21 HRV features were used: 9 calculated over
the entire FHR, 4 during rest, 4 during contractions, and 4
ratio features. The feature set that consists only of HRV fea-
tures calculated over the entire FHR, without distinguishing
contractions from rest periods, is defined as S1 ∈ R9 . The
feature set consisting of HRV features that are calculated
over the entire FHR and separately during contractions and
rest periods is defined as S2 ∈ R21 .
E. Feature selection
Many of the HRV features described in Sections II-C and
II-D are correlated and might contain redundant information.
A feature is considered redundant if the information about the
two classes is already captured by other features. Moreover,
we have a small dataset with only 22 fetuses with adverse
outcome and there is a risk of overfitting when too many
HRV features are used for classification.
To select the best combination of HRV features we used
a sequential feature selection algorithm. Since the number
of HRV features in S1 and S2 is relatively small, we use a
Sequential Forward Selection (SFS) algorithm. Starting from
an empty feature set, SFS creates candidate feature subsets
by sequentially adding features that have not been selected,
such that the new candidate subset maximizes an objective
function. For implementation of SFS, we used sequentialfs
in the Statistics and Machine Learning toolbox of Matlab
R
(The Mathworks, Inc. Natick, MA).
SFS can be categorized as a wrapper method for feature
selection. Wrapper methods typically use the performance of
a classifier as objective function to score candidate feature
subsets. In our study we used the performance of an SVM
classifier that is explained in Section II-F.
F. Classification
In this paper, SVMs were used to classify healthy fetuses
and fetuses with adverse outcome. The SVM constructs
a set of hyperplanes in feature space that minimizes the
misclassification error while maximizing the margin between
the hyperplanes and the data points. A Gaussian radial
based function (rbf) was used to enable non-linear decision
2952
boundaries. For implementation of the SVMs, we used the
standard implementation fitcsvm in the Statistics and machine
Learning toolbox of Matlab
.
Due to the imbalance in the distribution of our dataset
over the classes, training an SVM with a similar cost for
the imbalance in the distribution of the dataset we therefore
defined the cost for misclassification of the minority class
five times higher than for misclassification of the majority
class, in line with the class distribution.
An SVM rbf classifier requires optimization of two pa-
rameters; the width of the Gaussian kernel function (σ) and
a penalty parameter for misclassification (C). Since feature
selection by SFS uses the performance of the classifier to
score the candidate subsets, σ and C should ideally be
optimized for each candidate subset. However, this would
computationally not be feasible. Instead, for feature selection
we therefore used the default settings of the SVM (σ = 1
and C = 1). Then, after feature selection, σ and C were
optimized for the selected HRV feature subset.
G. Validation
From S1 and S2 , subsets of three HRV features were
selected by SFS. Then, SVM parameters (σ and C) were
optimized with the selected subsets as input. We used 10-fold
cross validation to find the optimal subset of HRV features
and for optimization σ and C.
Due to our relatively small dataset containing only 22
cases of fetuses with adverse outcome, feature selection and
classification could be influenced by the location where the
dataset was split into ten fold. To obtain a more objective
result, the above procedure was repeated ten times. For both
S1 and S2 we thus generated ten subsets of three HRV
features and ten combinations of σ and C.
In case of an imbalanced dataset, the overall classification
accuracy is not the best metric to evaluate the performance
of the classifier [13]. Instead we used the geometric mean
(g) as metric. The metric g provides a balance between the
specificity (Sp, classification accuracy for the majority class)
and sensitivity (Se, classification
√ accuracy for the minority
class), and is defined as g = Sp · Se.
III. R ESULTS
Figure 3 shows all HRV features that were selected in the
ten runs by SFS. From S1 (only HRV features that are calcu-
lated over the entire FHR) α, SD, LFn , and HFn were most
frequently selected, as shown in Fig. 3a. In case S2 was used
(also including HRV features that were calculated separately
during contractions and rest periods), SDratio , RM SSDuc ,
and SampEnratio were most frequently selected, as shown
in Fig. 3b.
Results of the classification performance with the selected
subsets as input are shown in Table II. The average value of
the Se, Sp, and g for the ten runs is presented. The average
g for S1 and S2 were 69.0% and 79.6%, respectively.
 10
# selected
5
0
SD LFn
α
(a)
10
# selected
5 during the second stage of labor. Moreover, we also included
0
SDratio RMSSDuc SampEnratio SDuc
(b)
Fig. 3. Results feature selection by SFS. (a) Selected HRV features from
S1 . (b) Selected HRV features from S2 .
TABLE II
R ESULTS CLASSIFICATION
Feature set g (%)
S1 69.0
S2 79.6
IV. D ISCUSSION
Analysis of fetal HRV provides information on fetal dis-
tress and could assists clinicians in their decision making.
However, during labor the fetal cardiovascular system is
strongly influenced by contractions. In this paper we showed
that separating contractions from rest periods improves HRV
analysis for the detection of fetal distress during labor.
We used SFS to select the best combination of HRV
features for detection of fetal distress. Although SFS usually
provides an optimal subset for a particular classifier, it uses
the classification performance to score the candidate subsets.
As a consequence, selected HRV features might perform
differently for other classifiers. Furthermore, the best subset
was determined by 10-fold cross validation. Since splitting
the dataset into ten folds influences the feature selection, SFS
was repeated ten times. The influence of splitting the dataset
is reflected by the number of HRV features that SFS selected
at least once in the ten runs. As can be seen from Fig. 3,
feature selection from S2 was influenced less by splitting
the dataset than feature selection from S1 . Moreover, the
most frequently selected HRV feature from S1 (α) was only
selected 8 times, while the most frequently selected HRV
feature from S2 (SDratio ) was selected in all runs.
As shown in Fig. 3b, HRV features calculated over the
entire FHR are generally not selected from S2 . This indicates
that HRV features calculated separately during contractions
and during rest periods contain more information about the
two classes then HRV features calculated over the entire
FHR. Similar to results obtained in [6], our result suggest that
in particular ratio features are informative on fetal distress.
Further study using a larger dataset and other classifiers
is required to gain insight in which combination of HRV
features is the most informative.
The average classification performance improved from
g = 69.0% for HRV features calculated over the entire
FHR to g = 79.6% when HRV features were also calculated
separately during contractions and rest periods. Note that the
classifier was trained to perform well on a balance between
Se and Sp. Depending on how the classifier is used in
clinical practice, it might be better to define a different
penalty for mis-classification. This different penalty will
HFn LF TP SampEn
increase either Se or Sp. It is difficult to compare our results
to other work in the literature due to differences in the used
dataset. We only included FHR signals that were recorded
several types of medications that could influence fetal HRV.
HF
For future work it would be interesting to evaluate the
performance of the classifier for different types of medication
that are prescribed during labor. Furthermore, we selected
FHR segments near birth, but it might be interesting to
examine the classification performance over time. Moreover,
inclusion of clinical information such as gestational age or
maternal temperature might also improve the classification.
Se (%) Sp (%) ACKNOWLEDGMENT
62.7 76.4
77.7 81.6 This research was performed within the framework of the
IMPULS perinatology program. The authors would like to
thank A. Kwee (University Medical Center Utrecht, The
Netherlands) and K.G. Rosén (Boras University, Sweden)
for provision of their STAN registrations.
R EFERENCES
[1] J. van Laar, C. Peters, R. Vullings, S. Houterman, J. Bergmans, and
S. G. Oei, “Fetal autonomic response to severe acidaemia during
labour.” BJOG, vol. 117, no. 4, pp. 429–437, Mar 2010.
[2] H. Gonçalves, A. Rocha, D. Ayres-de Campos, and J. Bernardes,
“Linear and nonlinear fetal heart rate analysis of normal and acidemic
fetuses in the minutes preceding delivery,” Med Biol Eng Comput.,
vol. 44, no. 10, pp. 847–855, 2006.
[3] J. Spilka, V. Chudáček, M. Kouckỳ, L. Lhotská, M. Huptych, P. Janků,
G. Georgoulas, and C. Stylios, “Using nonlinear features for fetal heart
rate classification,” Biomedical Signal Processing and Control, vol. 7,
no. 4, pp. 350–357, 2012.
[4] A. Georgieva, S. J. Payne, M. Moulden, and C. Redman, “Artificial
neural networks applied to fetal monitoring in labour,” Neural Com-
puting and Applications, vol. 22, no. 1, pp. 85–93, 2013.
[5] P. Abry, S. Roux, V. Chudácek, P. Borgnat, P. Goncalves, and
M. Doret, “Hurst exponent and intrapartum fetal heart rate: Impact
of decelerations,” in Proc. of the IEEE 26th Int. Symp. on Computer-
Based Medical Systems. IEEE, 2013, pp. 131–136.
[6] G. Warmerdam, R. Vullings, J. Van Laar, J. Bergmans, L. Schmitt,
S. Oei et al., “Using uterine activity to improve fetal heart rate
variability analysis for detection of asphyxia during labor,” Physiol
Meas, vol. 37, no. 3, pp. 387–400, 2016.
[7] M. Romano, P. Bifulco, M. Cesarelli, M. Sansone, and M. Bracale,
“Foetal heart rate power spectrum response to uterine contraction,”
Med Biol Eng Comput., vol. 44, no. 3, pp. 188–201, 2006.
[8] A. Sundström, D. Rosén, and K. Rosén, “Fetal surveillance,” Gothen-
burg: Neoventa Medical AB, 2000.
[9] U. Schneider, E. Schleussner, A. Fiedler, S. Jaekel, M. Liehr,
J. Haueisen, and D. Hoyer, “Fetal heart rate variability reveals dif-
ferential dynamics in the intrauterine development of the sympathetic
and parasympathetic branches of the autonomic nervous system,”
Physiological Measurement, vol. 30, no. 2, p. 215226, Jan 2009.
[10] C. Peters, R. Vullings, M. Rooijakkers, J. Bergmans, S. Oei, and
P. Wijn, “A continuous wavelet transform-based method for time-
frequency analysis of artefact-corrected heart rate variability data.”
Physiol Meas, vol. 32, no. 10, pp. 1517–1527, Oct 2011.
[11] J. Richman and J. Moorman, “Physiological time-series analysis using
approximate entropy and sample entropy,” Am J Physiol Heart Circ
Physiol., vol. 278, no. 6, pp. H2039–H2049, 2000.
[12] E. Ihlen, “Introduction to multifractal detrended fluctuation analysis
in matlab,” Front. Physio., vol. 3, 2012.
[13] H. He and E. A. Garcia, “Learning from imbalanced data,” IEEE
Trans. Knowledge and Data Eng., vol. 21, no. 9, pp. 1263–1284, 2009.
2953