Professional Documents
Culture Documents
Selective Heart Rate Variability Analysis To Account For Uterine Activity
Selective Heart Rate Variability Analysis To Account For Uterine Activity
Abstract— Cardiotocography (CTG) is currently the most of- response of the fetal ANS to stabilize the cardiovascular sys-
ten used technique for detection of fetal distress. Unfortunately, tem under fluctuating conditions of the exterior (i.e. uterus).
CTG has a poor specificity. Recent studies suggest that, in If HRV features are calculated without distinguishing periods
addition to CTG, information on fetal distress can be obtained
from analysis of fetal heart rate variability (HRV). However, of contractions from rest periods, this effect is averaged over
uterine contractions can strongly influence fetal HRV. The aim the rest periods and is no longer visible.
of this study is therefore to investigate whether HRV analysis In this paper we examine whether detection of fetal
for detection of fetal distress can be improved by distinguishing distress by HRV analysis can be improved by separating
contractions from rest periods. Our results from feature selec- contractions from rest periods. Therefore, in addition to
tion indicate that HRV features calculated separately during
contractions or during rest periods are more informative on HRV features calculated over the entire heart rate, HRV
fetal distress than HRV features that are calculated over the features were also calculated separately during contractions
entire fetal heart rate. Furthermore, classification performance and during rest periods. A sequential feature selection al-
improved from a geometric mean of 69.0% to 79.6% when gorithm was implemented to determine the most informative
including the contraction-dependent HRV features, in addition subset of HRV features, both with and without the additional
to HRV features calculated over the entire fetal heart rate.
contraction-dependent HRV features. Using these subsets as
I. I NTRODUCTION input, performance of classification of healthy fetuses and
fetuses with adverse outcome was evaluated by Support
During labor, oxygen supply to the fetus can temporally Vector Machines (SVMs).
be reduced due to contractions. If the reduction in oxygen First, data acquisition is described in Section II-A. Then,
supply is severe and prolonged this can lead to neonatal Sections II-B to II-F describe the signal processing tech-
morbidity and mortality. In clinical practice, cardiotocogra- niques used in this paper: pre-processing, HRV analysis,
phy (CTG) is currently the most often used technique for HRV analysis during contractions and rest periods, feature
fetal monitoring. However, CTG is interpreted visually and selection, and classification. In section II-G, our validation
has a high inter- and intra-observer variability. Besides, the method is explained. Finally, results are presented in Section
poor specificity of CTG leads to unnecessary interventions. III and discussion in Section IV.
Therefore, additional information is required to reduce the
number of unnecessary operative deliveries. II. M ETHOD
Recent studies have shown interest in analysis of fetal This section describes the different processing techniques
heart rate variability (HRV) for detection of fetal distress that were used in this paper. The overall procedure is
[1]–[6]. Since the heart rate is regulated by the autonomic presented in Fig. 1.
nervous system (ANS), variations in the heart rate reflect
changes in ANS control and could indirectly provide infor- A. Data acquisition
mation on fetal distress. In the literature, many HRV features For this study, fetal heart rate (FHR) and uterine activity
have been developed that describe different properties of the (UA) signals were obtained from CTG registrations that
heart rate. However, most HRV features have been developed were recorded on a Neoventa STAN
R
(Mölndal, Sweden).
for adults. Unlike for adults, it is not possible to control the We only used beat-to-beat FHR signals that were measured
fetal cardiovascular system. Moreover, during labor contrac- by a scalp electrode. The UA was recorded either with a
tions can strongly influence the fetal cardiovascular system. tocodynamometer or intra-uterine pressure catheter. Based
In [6] and [7] it was shown that for healthy fetuses, spectral on the UA signal, contractions were annotated by a clinical
power is significantly higher during contractions as compared expert that had no knowledge of the FHR signal or any other
to rest periods (i.e. the periods in between contractions). The medical information. An example of a recording is shown in
increased spectral power during contractions could indicate a Fig. 2.
1 G.J.J. Warmerdam, R. Vullings, and J.W.M. Bergmans are with the Fetal outcome was defined using the arterial acid-base
Faculty of Electrical Engineering, Eindhoven University of Technology, status: adverse outcome was defined as a pH below 7.05 and
Eindhoven, The Netherlands g.j.j.warmerdam@tue.nl healthy as a pH above 7.20 [8]. Pregnancies complicated by
2 J.O.E.H. Van Laar,, M.B. Van der Hout-Van der Jagt, and S.G. Oei
intra-uterine growth restriction or fetal congenital anomalies
are with the Department of Obstetrics and Gynecology, Máxima Medical
Center, Veldhoven, The Netherlands were excluded. In total, 22 cases were included with adverse
3 L. Schmitt is with Philips Research, Eindhoven, The Netherlands outcome, that were matched with 110 healthy cases. Note
150
calculate the power in the frequency bands [10]. Power
100 in the low frequency band (LF , 0.04-0.15 Hz) represents
50 both sympathetic and parasympathetic activity, and power
100
in the high frequency band (HF , 0.4-1.5 Hz) represents
UA (mmHg)
75
50
parasympathetic activity only. Besides absolute LF and HF ,
25 also normalized frequency powers were calculated (LFn and
0
0 2 4 6 8 10 HFn ), by dividing LF and HF by the total power (T P ,
time (min)
0.04-1.5 Hz).
Fig. 2. Example of a FHR and UA signal for a healthy fetus. 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
that certain medication can cross the placenta and influence healthy blood pressure control by the ANS. Complexity in
fetal HRV (e.g. anti-hypertensives or pain killers). We ap- short time signals can be quantified by Sample Entropy
proximately matched the prevalence of these medications in (SampEn) [11]. SampEn estimates the conditional proba-
both groups, in accordance with the difference in size of the bility that if patterns are similar for length m − 1, they will
two groups. also be similar for length m. A tolerance parameter r is used
to define a threshold for similarity between two patterns. In
B. Pre-processing this study we used m = 2 and r = 0.2SD, similar to settings
HRV features are calculated on interbeat interval (R-R used in [2] and [3].
interval) signals. Since beat-to-beat R-R intervals can only Finally, Detrended Fluctuation Analysis (DFA) was used
be determined at the instance a heartbeat occurs, the R- to describe the scaling properties of the HRV over different
R intervals are not equidistantly sampled. Therefore, linear time scales [12]. The scaling exponent α was calculated
interpolation was used to obtain an equidistantly sampled over a range of 1 to 60 seconds, corresponding to the
heart rate. When the difference between two consecutive R- physiological range of the baroreflex.
R intervals was more than 200 ms or more than 25 Beats Per
Minute (BPM), we defined both R-R intervals as artifacts. D. HRV analysis during contractions and rest periods
Artifacts were corrected by linear interpolation. To calculate HRV features separately during contrac-
For each CTG registration, a segment of 10 minutes was tions and during rest periods, the expert annotations of the
selected with less than 20% loss of FHR signal and at least contractions were used. Rest periods were defined as the
2 minutes of FHR signal during contractions and during rest periods in between contractions. If the time between two
periods. Segments were selected closest to birth, only during contractions was more than one minute, we only considered
the second stage of labor (stage of active pushing) and within the final minute preceding the second contraction as rest
a range of 45 minutes before birth. period. Values of HRV features during contractions were
calculated based on R-R values that were recorded during
C. HRV features contractions. Similarly, values of HRV features during rest
For each 10 minute FHR signal, conventional HRV fea- were calculated based on R-R intervals that were recorded
tures were calculated that describe different properties of during rest periods. For each HRV feature, the average value
the heart rate. An overview of the used HRV features is of all contraction periods and all rest periods within the 10
presented in Table I. Due to limited space, we only briefly minute segments (indicated as f eatureuc and f eaturerest ,
summarize the HRV features used in this study. For a more respectively) is then used for classification. We also con-
thorough description of used HRV features, the authors refer sidered the ratio between HRV features calculated during
the interested reader to [6]. contractions and during rest periods (f eatureratio ).
To describe the statistical properties of the heart rate we Note that computation of LF , LFn , HFn , and α requires
used the standard-deviation (SD) and root mean square approximately one minute of consecutive FHR signal. Since
2951
TABLE I boundaries. For implementation of the SVMs, we used the
HRV FEATURES standard implementation fitcsvm in the Statistics and machine
Learning toolbox of Matlab
R
.
HRV features entire FHR contractions rest ratio
SD x x x x Due to the imbalance in the distribution of our dataset
RM SSD x x x x over the classes, training an SVM with a similar cost for
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 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
during the second stage contractions and rest periods are class, in line with the class distribution.
often less than one minute, we could not extract these HRV An SVM rbf classifier requires optimization of two pa-
features separately during contractions and rest periods. We rameters; the width of the Gaussian kernel function (σ) and
thus limited our analysis during contractions and rest periods a penalty parameter for misclassification (C). Since feature
to SD, RM SSD, HF , and SampEn. selection by SFS uses the performance of the classifier to
In total, 21 HRV features were used: 9 calculated over score the candidate subsets, σ and C should ideally be
the entire FHR, 4 during rest, 4 during contractions, and 4 optimized for each candidate subset. However, this would
ratio features. The feature set that consists only of HRV fea- computationally not be feasible. Instead, for feature selection
tures calculated over the entire FHR, without distinguishing we therefore used the default settings of the SVM (σ = 1
contractions from rest periods, is defined as S1 ∈ R9 . The and C = 1). Then, after feature selection, σ and C were
feature set consisting of HRV features that are calculated optimized for the selected HRV feature subset.
over the entire FHR and separately during contractions and
G. Validation
rest periods is defined as S2 ∈ R21 .
From S1 and S2 , subsets of three HRV features were
E. Feature selection selected by SFS. Then, SVM parameters (σ and C) were
Many of the HRV features described in Sections II-C and optimized with the selected subsets as input. We used 10-fold
II-D are correlated and might contain redundant information. cross validation to find the optimal subset of HRV features
A feature is considered redundant if the information about the and for optimization σ and C.
two classes is already captured by other features. Moreover, Due to our relatively small dataset containing only 22
we have a small dataset with only 22 fetuses with adverse cases of fetuses with adverse outcome, feature selection and
outcome and there is a risk of overfitting when too many classification could be influenced by the location where the
HRV features are used for classification. dataset was split into ten fold. To obtain a more objective
To select the best combination of HRV features we used result, the above procedure was repeated ten times. For both
a sequential feature selection algorithm. Since the number S1 and S2 we thus generated ten subsets of three HRV
of HRV features in S1 and S2 is relatively small, we use a features and ten combinations of σ and C.
In case of an imbalanced dataset, the overall classification
Sequential Forward Selection (SFS) algorithm. Starting from
accuracy is not the best metric to evaluate the performance
an empty feature set, SFS creates candidate feature subsets
of the classifier [13]. Instead we used the geometric mean
by sequentially adding features that have not been selected,
(g) as metric. The metric g provides a balance between the
such that the new candidate subset maximizes an objective
specificity (Sp, classification accuracy for the majority class)
function. For implementation of SFS, we used sequentialfs
and sensitivity (Se, classification
√ accuracy for the minority
in the Statistics and Machine Learning toolbox of Matlab
R
2952
10
classifier was trained to perform well on a balance between
# selected Se and Sp. Depending on how the classifier is used in
5
clinical practice, it might be better to define a different
0
penalty for mis-classification. This different penalty will
SD LFn HFn LF TP SampEn
α
increase either Se or Sp. It is difficult to compare our results
(a) to other work in the literature due to differences in the used
10
dataset. We only included FHR signals that were recorded
# selected
2953