ISSNIP Biosignals and Biorobotics Conference 2010
Theme: Biosignals and Robotics for Better and Safer Living
Vitoria, Brazil - 4-6 January 2010
A Wavelet Packet-based Energetic Approach for the
Analysis of the Uterine EMG
Bassam Moslem 1,2, Mohamad Khalil 2, Catherine Marque 1 , Mohamad O. Diab3
1
University of Technology of Compiègne,CNRS UMR 6600 Biomécanique et Bio-ingénierie, Compiègne, Cedex, France.
Corresponding author: +33 3 44 23 44 23, [email protected], [email protected]
2
Azm Center for Research in Biotechnology, Lebanese University, Tripoli, Lebanon, [email protected]
3
Hariri Canadian University (HCU), College of Engineering Bio-electronics department, P.O. Box: 10 - Damour, Chouf 2010,
Meshref, Lebanon. Office: 00961 560 1386 ext. 512, [email protected]
frequencies. More recently, Diab et al. [9] used wavelet
transform to explore the frequency content of the uterine
EMG then to extract the parameters of the contractions.
Results showed that the frequency content changes from one
woman to another, and during pregnancy. They finally
concluded that uterine contractions may be classified into
independent groups according to their frequency content.
Abstract— In respect to the main goal of our ongoing work for
predicting preterm birth, we investigate in this paper whether
the spectral content energy distribution of the transabdominal
uterine electromyography (EMG) signals is modified during
pregnancy and in parturition. We analyzed recordings of uterine
electrical activity on eleven women at different increasing
pregnancy term. We introduced a method based on the evolution
of the energy distribution of the uterine EMG along gestation
using wavelet packet transform (WPT). The results obtained
from the analyzed data indicate that energy emerges significantly
towards higher frequencies along gestation. Furthermore, we
also observed a noticeable difference in the energy distribution
between two studied classes (pregnancy, labor). These results are
very promising for monitoring pregnancy and detecting labor. It
may help in identifying preterm labor.
The main aim of this paper is to describe how energy
proportion, in different frequency bands, is modified for
different pregnancy terms and during labor. Unfortunately, the
uterine EMG is a highly non-stationary signal. Therefore,
classical methods which are based on the assumption of
stationarity, such as the Welch periodogram and the
correlogram methods cannot be used. Hence, most recent
methods for the analysis of uterine EMG are based on modern
techniques for processing non-stationary signals. One of the
most adapted techniques is the wavelet packet transform
(WPT). This technique has the capability to separate the signal
energy among different frequency bands (i.e. different scales)
realizing a good compromise between temporal and frequency
resolution. Over recent years, WPT has emerged as an
advanced tool for the investigation of biomedical signals:
Chendeb et al. [10] used the WPT for event detection and
segmentation in real signals like uterine EMG. Leman et al.
[11] proposed a denoising method based on the WPT and used
it to denoise uterine EMG signals while Hitti et al. [12] used
the WPT to detect abrupt changes in noisy multi-component
signals.
Keywords: Uterine EMG, energy distribution, Wavelet Packet
analysis, Prediction, monitoring.
I.
INTRODUCTION
Parturition, or labor, is the culminant point of pregnancy
by which the fetus is expelled from the uterus [1]. Knowing
that labor has begun as well as predicting when it will start is
important for both normal and complicated pregnancies. In
that later case, it will allow clinicians to start treatment early
in women with true labor. Since the uterine electrical activity
gradually changes throughout pregnancy, the uterine EMG
can be regarded as one of the most promising methods for
pregnancy and parturition monitoring. It may provide a noninvasive, more objective and highly accurate method for the
prediction of labor [2-3].
In our work, we describe how we used the wavelet packet
decomposition to monitor the energy distribution of the
uterine EMG in different frequency bands and for several
increasing pregnancy terms. Additionally, in an attempt to
differentiate between true and false labor, we made a
comparison of the energy distribution between two classes of
contractions (pregnancy, labor).
Over the last years, uterine EMG has become the object of
many research studies [4-6]. Temporal and spectral
characteristics of the uterine EMG were defined in two
physiological states: pregnancy and parturition. Energy was
analysed in many researches. Schlembach et al. [7] reported
that there are measurable increases in energy and a rise in the
high-frequency content of the uterine electrical activity with
the progression of pregnancy and the onset of labor. Marque
et al. [8] observed spectral changes of the uterine EMG from
pregnancy to parturition and concluded that pregnancy
contractions are mainly characterized by low frequencies
whereas labor contractions are related to the presence of high
II. MATERIALS AND METHODS
A. Data
We studied 11 women: 5 recorded during pregnancy (33 –
39 week of gestation, WG) and 6 during labor (39 – 42 week
98
ISSNIP Biosignals and Biorobotics Conference 2010
Theme: Biosignals and Robotics for Better and Safer Living
Vitoria, Brazil - 4-6 January 2010
of gestation, WG). Among the 5 women recorded during
pregnancy, two women were recorded for different WG. After
manual segmentation of the bursts of uterine electrical activity
corresponding to contractions with the help of the
tocodynamometer trace, we obtained 12 labor contractions
and 25 pregnancy contractions. The measurements were made
at the Landspitali University hospital in Iceland and FSA
Akureyri, Iceland, by using a protocol approved by the ethical
committee (VSN 02-0006-V2). They were performed by using
a 16 electrode grid, arranged in a 4x4 matrix positioned on the
women’s abdomen (Fig.1). In this study, we considered
vertical bipolar signals (Vbi) in order to increase the signal to
noise ratio. Our signals form thus a rectangular 3x4 matrix.
All the bursts presented a good signal to noise ratio on all
bipolar channels. Signals were sampled at 200 Hz. The
recording device has an anti-aliasing filter with a cut-off
frequency of 100 Hz. The simultaneous tocodynamometer
paper trace was digitized to ease the segmentation of the
bursts.
packets were introduced by Coifman and Wickerhauser [17]
as a generalized family of multiresolution orthogonal or
biorthognal basis. Unlike wavelet transform which is realized
only by a low-pass filter bank, WPT is implemented by a
basic two-channel filter bank which can be iterated over either
the low-pass or the high-pass branch. Therefore, with WPT,
the information in high frequencies as well as that in low
frequencies can be analyzed with WPT. As a result, finer
frequency bands can be obtained by wavelet packet transform
than by wavelet transform. We start with h(n) and g (n) , the
two impulsive responses of low-pass and high-pass analysis
filters, corresponding to the scaling function and the wavelet
function, respectively. The sequence of functions is defined by:
2 N 1
W2n ( x)  2
 h(k )W (2x  k )
n
k 0
2 N 1
W2n1 ( x)  2
 g (k )W (2x  k )
n
k 0
where, n is the time localization parameter, k is the oscillation
parameter and j is the scale parameter. W0 ( x)   ( x) is a
scaling function which corresponds to a low-pass filter. The
filtered signal is an approximation of the analyzed signal. The
function W1 ( x)   ( x) is the mother wavelet function that
corresponds to a high-pass filter. The filtered signal is a detail
of the analyzed signal. Consequently, the approximation and
detail can be further decomposed by dyadic decomposition by
using the dilated and translated scale functions and wavelet
functions. In other words, the three indexed family of
analyzing functions can be reached by:
W j ,n,k ( x)  2 j 2Wn (2 j x  k )
W j ,n,k ( x) roughly analyzes the fluctuations of the signal
around the position 2 j.n . The wavelet packets coefficients at
each node (j, n) are written as:
Fig. 1. Electrode configuration on the woman’s abdominal wall. The Vbi
(i=1-12) represent the derived bipolar signals.
C j,n (k ) 
B. Normalizing
Since the uterine EMG is highly variable and is dependent
upon electrode application and placement [13], normalization
is necessary to facilitate comparison between different
subjects or between days within a single subject. It is a
prerequisite for any comparative analysis of uterine EMG or
for documenting changes over days. Therefore, all the 12
bipolar signals recorded on each woman’s abdomen were first
normalized before the analysis by dividing each signal by its
variance.
f (t ), 2 j 2Wn (2  j t  k )
Figure 2 illustrates the wavelet packet decomposition tree
with three levels.
C. WAVELET PACKETS TRANSFORM
1) Introduction:
The multiresolution analysis was first proposed in 1989 by
Mallat [14]. Since then, the advanced research and
development in wavelet analysis gained more and more
attention and have found numerous applications in fields such
as signal processing and pattern recognition [15-16]. Wavelet
Fig. 2. The wavelet packet decomposition tree of three levels
99
ISSNIP Biosignals and Biorobotics Conference 2010
Theme: Biosignals and Robotics for Better and Safer Living
Vitoria, Brazil - 4-6 January 2010
Each node of the WP tree is indexed with the pair of
integers (j,n): j represents the level and n the number of nodes
per level with n=0, 1,…, 2 j  1 . A vector of WP coefficients
C j ,n corresponds to each node (j,n), according to the basic
step
procedure.
The
length
of
a
vector
C j ,n
increase of the energy proportions in each of the remaining
packets representing the rest of the bandwidth of the studied
signal. Therefore, we grouped all these packets into one
representing the frequency band where it was shown that the
energy proportion increases.
is
approximately N 2 j . The frequency range corresponding to
each node of the WP tree is approximately FS 2 j 1 .
2) Wavelet and scale choices:
In previous studies [7-9], bursts of electrical activity
corresponding to contractions were analyzed with power
spectral density analysis. It was demonstrated that the main
energy of the uterine EMG is located in the 0.34-1Hz range.
Since the sampling frequency of the signals is 200 Hz, and in
order to reduce the number of decomposition levels, all the
signals were first downsampled. Therefore, the bandwidth of
each signal was reduced to 3.125Hz. In this case, we are only
interested to study the packets of level 3, where the signal is
decomposed into 8 frequency bands and the bandwidth of
each one is approximately 0.39Hz.
Figure 3.a illustrates a measurable decrease in the energy
proportion of the low frequency content of the bursts
throughout pregnancy while Figure 3.b shows a relocation of
the energy of the uterine EMG represented by an increase of
the energy proportion of the higher frequency bands along
gestation, for both women. We can evidence an obvious
longitudinal evolution of the values of energy percentage in
these frequency bands of the signal throughout pregnancy.
Moreover, the choice of the mother wavelet is a critical
problem in signal processing theory. It must be well adapted
to the characteristics of the studied signals [9,18,19]. We state
that the choice of the wavelet can be determined from the
information included in the interesting part of the signal. In
previous works related to the uterine EMG signal, the Symlet
wavelet was used. It has been demonstrated that it gives the
best results in terms of analyzing uterine EMG signals [20].
III. RESULTS
First, we studied the temporal variations of the energy
distribution of the uterine EMG along gestation. Fig. 3
illustrates the evolution of the percentage of energy in two
different frequency bands corresponding to the packet (3,0),
(3,1), (3,2) and (3,3) representing different frequency bands
covering the total bandwidth of the uterine EMG signals.
These packets will allow us to monitor the variation of the
energy distribution in the frequency bands [0-0.39 Hz] and
[0.39-0.78 Hz], [0.78-1.17 Hz] and [1.17-1.56 Hz]
respectively. Data used were recorded on two women (w1 and
w2) at two and three increasing pregnancy terms respectively.
First, we compute the percentage of energy corresponding to
each node of the packet tree over the matrix of signals related
to each contraction. Then we determine their average over all
the contractions during the recording sessions: 2 contractions
at 30 WG (G1) and 2 contractions at 32 WG (G2) for the first
women W1 and 4 contractions at 33 WG (G’1), 2 contractions
at 35 WG (G’2) and 3 contractions at 37 WG (G’3) for the
second woman W2. The mean values were determined for
each woman at each term and plotted against the woman’s
weeks of gestation interval (fig.3). The results show that there
energy distribution changes along gestation: there is a
noticeable decrease in the energy proportion in the first packet
representing the lowest frequency band whereas, there is an
Fig. 3. Evolution of the energy distribution for the 2 women (W1 and W2) at
different increasing weeks of gestation: a) energy located in the low
frequency band of the signals: packet (3,0), b) energy located in the rest of
the bandwidth of the signal representing higher frequencies band: packets
(3,1), (3,2) and (3,3)
Furthermore, our study was conducted to investigate
whether these values undergo significant changes from
pregnancy to labor. We studied 12 contractions randomly
chosen for each class: one class recorded during pregnancy
contraction (32 WG) and the other during labor (41 WG).
First, we compute the percentage of energy corresponding to
each node of the packet tree over the matrix of signals related
to each of the 12 contractions of each type. Then we compute
the mean of these values over the contractions associated to
each class. In this part of our study, we also chose to compare
the energy distribution in the same packets (i.e. the same
100
ISSNIP Biosignals and Biorobotics Conference 2010
Theme: Biosignals and Robotics for Better and Safer Living
Vitoria, Brazil - 4-6 January 2010
frequency bands). The results are illustrated in fig.4a for the
comparison of the energy proportion in the low frequency
band and fig.4b for the rest of the bandwidth. These results
indicate that the energy distribution undergoes noticeable
changes from pregnancy to parturition.
(low and high) of the signals changes gradually during
pregnancy and exhibits a longitudinal evolution along
gestation. As for now, we only have longitudinal data for two
women and for one woman we only have two data points. The
evidences provided by this scanty dataset are not sufficient to
deduce a general model for pregnancy evolution, since they do
not permit us to give a quantitative value for the change.
However, they can contribute to the conception of this model.
Additionally, noticeable differences in the energy
distribution of the electrical signals that originate from
pregnancy and labouring patients have been shown to
distinguish the two types of recordings (pregnancy or labor)
from one another. This information may provide important
evidence for the onset of labor and it may help indentify
premature labor.
V. CONCLUSION
In the described work, we studied the energy distribution
of the uterine EMG along gestation. The study demonstrated
that the energy distribution, similarly to other extracted
features from the uterine EMG signals, gradually changes
during pregnancy and energy emerges towards higher
frequencies. Thus, it was concluded that labor is not a direct
transition from an inactive to an active state of the uterus. The
results may help improve our understanding of the progression
of pregnancy as well as how labor is initiated. We finally,
concluded that the results discussed in our study may help
identify patients who are likely to deliver prematurely.
Fig. 4. Average of the values of the energy distribution obtained from the two
studied classes a) energy located in the low frequency band of the signals, b)
energy located in the rest of the bandwidth of the signal
IV. DISCUSSION
From our study as well as those of others, it is clear that
changes in uterine EMG activity are associated with
progression of pregnancy and the onset of labor. Changes of
the characteristics of the uterine activity along gestation are
induced by the changes that occur in the electrical properties
of the uterus during the gestational period and labor. In the
first part of our work, we observed that the uterine EMG
energy, which spreads across the bandwidth of the signal,
undergoes spectral changes throughout pregnancy as well as
from pregnancy to parturition. Our study indicates that the
energy distribution follows a very particular way along
gestation. As pointed earlier, it was reported that the energy of
the uterine activity increases with the progression of
pregnancy and the onset of labor. In our study, we showed
that, as the time of delivery approaches, the energy of the
signal emerges from low frequencies into higher frequencies.
This is coherent with the results presented in different studies
either on women [7,8], or on animals [20]. Our results
revealed that the percentage of energy in two frequency bands
Thus, this study may allow in the future a more accurate
labor/non-labor patient classification. It may help identifying
women who will imminently deliver. Also, with a larger
database of this type of signals, this work may lead to the
conception of a general model of pregnancy monitoring by
means of uterine EMG processing.
ACKNOWLEDGMENT
The authors wish to thank M. Brynjar Karlsson (School of
science and Engineering, Reykjavik University, Iceland) for
providing the women uterine EMG recordings.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
101
Marieb E., Anatomie et physiologie humaine, 3/e, Pearson Education,
2005.
Maner W., Garfield RE., Maul H., Olson G., Saade G., Predicting tem
and preterm delivery with transabdominal uterine electromyography,
The American College of Obsterticians and Gynecologists, June 2003,
Vol. 101,No.6.
Verdenik I, Panjntar M, Leskošek B., Uterine electrical activity as
predictor of preterm birth in women with preterm contractions,
European Journal of Obstetrics & Gynecology and Reproductive
Biology, 95, 2001, 149-153.
Duchêne J., Khalil M., Hewson D., Current advances in uterine SEMG
processing methods, Recent Res. Devel. Biomed. Eng., 2002,:149-170.
Garfield RE., Maner W., Physiology and electrical activity of uterine
contractions, Seminars in Cell & Developmental Biology 18, 2007, 289295.
Doret M., Pasquier J.-C., Gharib
C., Gaucherand P.,
L’électromyogramme utérin : principes et intérêt pour le diagnostic de
ISSNIP Biosignals and Biorobotics Conference 2010
Theme: Biosignals and Robotics for Better and Safer Living
Vitoria, Brazil - 4-6 January 2010
[7]
[8]
[9]
[10]
[11]
[12]
[13]
travail prématuré, Journal de Gynécologie Obstétrique et Biologie de la
Reproduction, 2008, 37, 24—32
Schlembach D., Maner W, Garfield RE., Maul H., Monitoring the
progress of pregnancy and labor using electromyography, European
Journal of Obstetrics & Gynecology and Reproductive Biology, 144S,
2009, S33-S39.
Marque C., Duchêne J., Leclercq S., Panczer G., Chaumont J., Uterine
EHG processing for obstetrical monitoring, IEEE Transactions on
Biomedical Engineering, Vol. BME-33, No. 12, December 1986.
Diab M.O., Marque C., Khalil M., An unsupervised classification
method of uterine electromyography signals: Classification for detection
of preterm deliveries, J. Obstet. Gynaecol. Res. Vol. 35, No. 1: 9–19,
February 2009.
Chendeb M., Khalil M., Duchêne J., The use of wavelet packets for
event detection, Proceedings of the 13th European Signal Conference
EUSIPCO, Antalya, Turkey, 4-8 September 2005.
Leman H., Marque C., Rejection of the maternal electrocardiogram in
the electrohysterogram signal, IEEE Transactions on Biomedical
Engineering, vol. 47, August 2000.
Hitti E., Lucas M. F., Wavelet-packet basis selection for abrupt changes
detection in multicomponent signals, in Proc. EUSIPCO 1998, Island of
Rhodes, Greece, September 8-11, 1998, pp. 1841-1844.
Lehman G., McGill S.M., The Importance of Normalization in the
Interpretation of Surface Electromyography: A Proof of Principle,
Journal of Manipulative and Physiological Therapeutics, Volume 22,
September 1999.
[14] Mallat S.G., A theory for multiresolution signal decomposition: the
wavelet representation, IEEE Trans. Pattern Anal. Mach. Intell. 11
(1989) 674—693.
[15] Khan Y.U., Gotman J., Wavelet based automatic seizure detection in
intracerebral electroencephalogram, Clinical Neurophysiology 114,
2003, 898-908.
[16] Diab M.O., Marque C., Khalil M., An unsupervised classification
method of uterine electromyography signal using wavelet
decomposition, Proceedings of the 26th Annual International
Conference of the IEEE EMBS San Francisco, CA, USA, September 1-5,
2004.
[17] Coifman R. R., Wickerhauser M. V., Entropy based algorithms for best
basis selection, IEEE Trans Inform Theory, vol. 38, pp.1241-1243,
1992.
[18] Dinh H A N, Kumar D K, Pah N D, Burton P, Wavelets for QRS
Detection, Proceedings – 23rd Annual Conference – IEEE/EMBS,
Istanbul, TURKEY, October 2001
[19] Pah ND, Kumar DK, Burton P, Adding Wavelet Decomposition to
Neural Networks for the Classification of Fatigue SEMG, Proceedings
of the IEEE Engineering in Medicine and Biology Society Conference,
Victoria, Australia, February2001
[20] Khalil M., Une approche de la détection et de la classification dans les
signaux non stationnaires. Application à l’EMG utérin, Thèse de
l’Université de Technologie de Troyes (1999).
[21] Mansour S., Devedeux D., Germain G., Marque C., Duchêne J., Uterine
EMG spectral analysis and relantioship in pregnant monkeys, Medical
& Biological Engineering & Computing, March 1996.
102