J. Biomedical Science and Engineering, 2010, 3, 837-842
doi:10.4236/jbise.2010.39113 Published Online September 2010 (http://www.SciRP.org/journal/jbise/
JBiSE
).
Published Online September 2010 in SciRes. http:// www.scirp.org/journal/jbise
Classification of uterine EMG signals using supervised
classification method
Mohamad O. Diab, Amira El-Merhie, Nour El-Halabi, Layal Khoder
Hariri Canadian University, College of Engineering, Bioelectronics Engineering, Damour, Meshref, Lebanon.
Received 5 July 2010; revised 19 July 2010; accepted 5 August 2010.
ABSTRACT
Aim: The main purpose of this ar ticle is to det ect any
risk of preterm deliveries at an early gestation period
using uterine electromyography signals. Detecting
such uterine signals can yield a promising approach
to determine and take actions to prevent this poten-
tial risk. Methods: The best position for the detection
of different uterine signals is the median vertical axis
of the abdomen. These signals differ from each other
by their frequency content. Initially, simulation is
done for the real detected EMG signals: preterm de-
liveries (PD) EMGs and deliveries at term (DT)
EMGs. This is performed by applying autoregressive
model (AR) of specific order to estimate AR coeffi-
cients of these real EMG signals. Finally, after calcu-
lation of the AR parameters of the two types of de-
liveries, we generate two types of simulated uterine
contractions by using White Gaussian Noise (WGN).
Frequency parameter extraction and classification
are first applied on simulated signals to test the limits
and performance of the used methods. The last re-
maining step is the classification of the contractions
using supervised classification method. Results: Res-
ults show that uterine contractions may be classified
using the Artificial Neural Networks (ANNs). The Si-
mple Perceptron ANN is applied on the signals for
their supervised classification into independent grou-
ps: preterm deliveries (PD) and deliveries at term
(TD) according to their frequency content.
Keywords: Uterine EMG Signals; AR Model; PSD; ANN
1. INTRODUCTION
A most urgent challenge in healthcare currently is the phe-
nomenon of preterm labor, or labor prior to 37 comple-
ted weeks of gestation.
Although the majority of risk factors for preterm birth
have been identified, the prediction models exhibit a re-
latively poor performance. On the other hand, the effec-
tiveness of tocolytic agents depends on early introduc-
tion of therapy, therefore, timely recognition of the pro-
cess leading to active labor is of prime importance [1].
Preterm labor is known as the primary cause of neon-
atal morbidity and mortality, and it remains a major pro-
blem in obstetrical practice. The incidence of preterm
delivery varies between 5% and 11% [2].
In the developed world, the rate in general has been ri-
sing slowly over the past 10 to 20 years. In New Zea-
land, the singleton preterm birth rate rose from 4.3% in
1980 to 5.9% in 1999, a rise of 37%. Interestingly, the
rate rose by 72% in high socioeconomic groups but only
by 3.5% in the most deprived groups. This is due to the
effects of delayed childbearing in affluent career-women,
and to the increase in assisted reproduction in that group.
Preterm labor also leaves serious impacts on economy
and society as a whole. Besides, it affects the develop-
ment and health of a new born as well as the health of a
birth giving woman. The complications of preterm birth
include significant neurological, mental, behavioural and
pulmonary problems in later child’s life.
Significant progress has been made in the care of pre-
mature infants, but not in reducing the prevalence of pre-
term birth. The cause for preterm birth is in many situati-
ons elusive and unknown. The development of effective
methods to prevent or reduce the incidence of preterm
birth depends largely upon finding indicators for the pro-
blem.
Uterine electromyography (EMG) has shown great pr-
omise for monitoring patients during pregnancy, so it’s
of great interest in parturition and pregnancy abnormal-
ity identification.
Effect of the recording electrode position on the spec-
tral content of the signal has been investigated by using a
mathematical model of the women's abdomen. Then, the
simulated results have been compared to actual record-
ings.
On signals with noise reduced with a dedicated algo-

M. O. Diab et al. / J. Biomedical Science and Engineering 3 (2010) 837-842
838
rithm, the main frequency components of the signal spe-
ctrum have been characterized in order to compute para-
meters indicative of different situations: preterm contrac-
tions resulting nonetheless in term delivery (i.e. normal
contractions) and preterm contractions leading to preterm
delivery (i.e. high-risk contractions). A diagnosis system
permitted to discriminate between these different catego-
ries of contractions [3].
The best position for the detection of different uterine
signals is the median vertical axis of the abdomen. The
frequency content of the uterine contractions changes fr-
om one woman to another and during pregnancy. The ob-
tained signals can be classified into different groups de-
pending on their frequency because the main analyzed
parameters are extracted from the frequency content of
the uterine activity bursts [4]. The difference in frequen-
cies is a key to determination and detection of the type
of delivery.
Early detection of preterm deliveries will help to red-
uce costs of care and will allow taking all the suitable
precautions; such as, the use of tocolytic treatments to
prolong the gestation i.e. fetal development in utero [5].
Another proposed method for the detection of preterm
delivery is: “Uterine EMG Analysis: Time-Frequency
Based Techniques for Preterm Birth Detection”.
For the purpose of detection, two steps are required:
the first step aims to detect all events in this signal and to
identify these events by allocating them to physiological
classes: contractions, fetus motions, Alvarez or Long Du-
ration Low Frequency (LDBF) waves. The second step
consists of the identification of contractions between no-
rmal contractions and preterm birth contractions.
Recently, a method based on the comparison betw-
een AR and statistical classification models is being used.
The frequency content of the contraction changes from
one woman to another and during pregnancy. First, an
AR model is applied on the Uterine EMG signal for the
calculation of the ai parameters. Then, wavelet decom-
position is used to extract the parameters of each simu-
lated contraction, and an unsupervised statistical classi-
fication method based on Fisher test is used to classify
the signals. A principal component analysis projection is
then used to evidence the groups resulting from this cla-
ssification. Results show that uterine contractions may
be classified into independent groups according to their
frequency content and according to term (at the record-
ing, or at delivery) [6].
We propose in this article a method based on AR mo-
del and ANNs to characterize the uterine contractions re-
corded by abdominal EMG. The Simple Perceptron ANN
is applied on the signals for their supervised classi- fica-
tion into independent groups: preterm deliveries (PD)
and deliveries at term (TD) according to their frequency
content (Figure 1).
2. METHODS
2.1. UEMG Signals’ Extraction
The uterine EMG signals were recorded by means of
two Ag-AgCl Beckman electrodes, and one reference el-
ectrode located on the patient’s hip. After careful prepa-
ration of the skin (cleaning with an abrasive paste and
degreasing with a mixture of ether, alcohol, and acetone),
which decreases the interelectrode impedance, the elec-
trodes are aligned directly above the median axis of the
uterine muscle, on the epidermis, midway between sym-
physis and uterus fundus [6].
The resulting uterine EMG is amplified, band pass fi-
ltered by using a compact battery operating acquisition
system.
After a thorough briefing to insure best interpatient co-
nsistency, a hand held event marker was used to record
simultaneously the subject’s perception of contractions
and of fetal movements [6].
The information obtained by EMG recording was then
analyzed on a computer with a MATLAB.
UEMG Signals AR Model Simulation of UEMG Si
g
nals
PSD
Extraction of Fre
q
uenc
y
Parameters
Artificial Ne ural Network
(ANN)
Supervised Classification
of PD & DT
Figure 1. Block diagram showing sequence of work from obtaining uterine EMG signals, AR Modelling, simulation of original
signals, extraction of frequency parameters and finally supervised classification of term and preterm labor using ANNs.

M. O. Diab et al. / J. Biomedical Science and Engineering 3 (2010) 837-842
839
t
2.2. Modeling and Generation of Simulated
UEMG Signals
EMG(t) is the result of the integration of filtered elem-
entary activities. Filtering induces a correlation between
successive samples. An autoregressive modeling allows
the detection process to be applied to the prediction error
(white process) instead that to the original samples [7].
2.2.1. Autoregress ive Model (AR)
The autoregressive model is one of a group of linear pre-
diction formulas that attempt to predict an output y[n] of
a system based on the previous outputs (y[n-1], y [n-2]...)
and inputs (x[n], x[n-1], x[n-2]...). Deriving the linear
prediction model involves determining the coefficients
a1, a2, .., and b0, b1, b2, ... in the equation:
[]()1 [1]2[2]
0[]1[ 1]
yenestimateday nay n
bxnbxn
  



Several methods and algorithms exist for calculating
the coefficients of the AR model, all of which are imple-
mented by the matlab command ‘ar’.
The definition being used is the following:
1
N
titi
i
xax





where ai is the autoregression coefficients, xt is the series
under investigation, N is the order (length) of the filter
which is generally very much less than the length of the
series. The noise term or residue, epsilon in the ab- ove,
is almost always assumed to be Gaussian white noise
[8].
Several AR Models are available, in our study, we use
a Linear Prediction Filter Coefficients (lpc). Lpc uses the
autocorrelation method of autoregressive (AR) modeling
to find the filter coefficients. It determines the coeffici-
ents by minimizing the prediction error in the least squa-
res sense. It has applications in filter design and speech
coding.
In our model and after testing many orders of AR mo-
del, we used the order 16 to estimate the AR coefficients
of the two types of real EMG signals: preterm deliveries
(PD) and deliveries at term (DT) (Figure 2).
After obtaining the 16 lpc coefficients for each UE-
MG signal, the mean average was calculated for these lpc
coefficients of the first group (PD UEMG signals) as well
as for the second group (DT UEMG signals).
After calculation of the mean average of the AR para-
meters (coefficients) of two groups of UEMG signals,
we generated two types of simulated uterine contractions
by using White Gaussian Noise (WGN) (Figure 3).
2.3. Parameters’ Extraction and Classification
After simulation the EMG signals, power spectrum dens-
ity (PSD) has to be applied on them for further extraction
of frequency parameters from them.
2.3.1. Power Spectrum Density (PS D)
Power spectral density function shows the strength of the
variations (energy) as a function of frequency. In other
words, it shows at which frequencies variations are str-
ong and at which frequencies variations are weak. Com-
putation of PSD is done directly by the method called
FFT or computing autocorrelation function and then tra-
nsforming it. The goal of spectral estimation is to descri-
be the distribution (over frequency) of the power conta-
ined in a signal, based on a finite set of data. Estimation
of power spectra is useful in a variety of applications.
The power spectral density (PSD) of a stationary ran-
dom process xn is mathematically related to the autocor-
relation sequence (RXX) by the discrete-time Fourier tr-
ansform. In terms of normalized frequency, this is given
by:
 
1
2
j
m
xx xx
m
PRm




e
This can be written as a function of physical frequent-
cy f (e.g., in hertz) by using the relation ω = 2πf/fs, where
fs is the sampling frequency [9].
 
2
8
1x
j
fm f
xx xx
m
Pf Rme
f




In our study, we used periodogram. After that, frequ-
ency parameters of a signal have to be extracted. Then,
we used to extract ten frequency parameters (using dec-
ile method) from each simulated signal and stored them
successively in a matrix for precise and accurate classi-
fication with minimum error.
Uterine EMG
Signals
Calculation of Preterm
Deliveries Parameters
Calculation of Term
Deliveries Parameters
PT AR Model
DT AR Model
Figure 2. Creation an AR model for the PD and DT.

M. O. Diab et al. / J. Biomedical Science and Engineering 3 (2010) 837-842
840
Wavelet
Transform
Parameters
Extraction
Results
Inverse AR
Model
White Gaussian
Noise
Simulated
EMG Signals
Figure 3. Block diagram for the classification of a new data.
2.3.2. Artificial Neural Networks (ANNs)
For the classification to be properly performed, we used
artificial neural networks (ANNs).
ANNs are mathematical algorithms that are ideal for
the classification of objects (e.g., patients) based upon
one or more input variables (e.g., uterine EMG variables)
[10]. “Artificial intelligence” is the field of computer sc-
ience that attempts to give computers humanlike thought.
One of the primary means by which the computers are
endowed with such capability is through the use of an
ANN. An ANN is composed of fundamental compone-
nts, usually a circuit or computer program, which are
designed to be the machine equivalent of neurons in the
brains of living creatures. ANNs are made up of inputs
(hidden layers) and outputs.
Most recently work has progressed which will utilize
various biological and clinical markers for evaluating the
risks of preterm labor using ANNs [11]. Values calcu-
lated for a number of uterine EMG parameters were used
as inputs for a part for an ANN, and the outputs, speci-
fically patient classifications, were compared to clinical
assessments. The joining of uterine EMG with ANNs in
this way may produce a powerful, objective tool for as-
sessing labor.
The ANN that we used in our study is the simple per-
ceptron neural network.
The network of the simple perceptron adapts as foll-
ows: change the weight by an amount proportional to the
difference between the desired output and the actual out-
put.
As an equation, it is represented in this way:

.
ii
WDY   I
where η is the learning rate, D is the desired output, and
Y is the actual output.
Thus, we created a new simple perceptron that was tr-
ained on the frequency parameters of twenty UEMG sig-
nals representing preterm delivery and delivery at term
respectively with a 100 iterations. After the successful
training, the remaining extracted frequency parameters
of the simulated signals (30 simulated signals) were intr-
oduced for testing, it was found that almost all of them
were classified correctly (minimum error was obtained)
showing that artificial neural network is an effective cla-
ssification method.
Therefore, the Simple Perceptron ANN is applied on
the signals for their supervised classification into indep-
endent groups: preterm deliveries (PD) and deliveries at
term (TD) according to their frequency content.
3. FINAL RESULTS
A successful classification of preterm labor and labor at
term was achieved. We notice this classification clearly
on the Figure 4:
The displayed results correspond to the simulated UE-
MG signals that were introduced to the simple percep-
tron ANN to test its validity of classification. First 15
signals correspond to group G1 which are the preterm
labor contractions of the uterus while the last 15 signals
are from the group G2 which are the uterus contractions
of labor at term. It is clear from Figure 4 that the classi-
fication was achieved successfully with a minimal error.
The confusion matrix clearly demonstrates the confu-
sion error which is equal to 3.33% which is a minimum
05 1015 20 2530
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 4. Results of classification of simulated UEMG sig-
nals.

M. O. Diab et al. / J. Biomedical Science and Engineering 3 (2010) 837-842
841
error value (Tabl e 1). Moreover, we have run this progr-
am several times and found the mean average of the co-
nfusion errors; it was found to be equal to 1.332% for 10
runs.
After trying the extraction of 2 frequency parameters,
we obtained the following Figure 5:
The confusion error is 100%. Therefore, 2 frequency
parameters are not enough for the classification of UE-
MG signals (Table 2).
4. DISCUSSION
In our research, we studied the uterine EMG signals of
the pregnant women to be able to classify them accord-
ing to the term of the delivery. The method of classifica-
tion used in our study is the supervised classification me-
thod based on Artificial Neural Networks (ANNs). From
the results obtained, we were able to proof that ANN (Si-
mple Perceptron) is a good method of classification pro-
viding minimum confusion error.
Table 1. Confusion matrix: result of classification on G1 and
G2.
G1 G2
G1 14 0
G2 1 15
05 10 15 20 25 30
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Figure 5. Results of classification of simulated UEMG signals
using 2 frequency parameters.
Table 2. Confusion matrix: result of classification on G1 and G2
using 2 frequency parameters.
G1 G2
G1 0 15
G2 0 15
In addition to this, we varied the number of extracted
frequency parameters from each UEMG signal. We have
found that as the number of parameters decreases, conf-
usion error increases indicating that the decile extraction
(10 frequency parameters) is the best of them (minimal
confusion error 3.33). We have tried extracting only five
and two frequency parameters from each signal; this ga-
ve less accurate results.
5. CONCLUSIONS
In this article, we have presented an approach for superv-
ised classification method by using Artificial Neural Ne-
tworks Simple Perceptron. Classification is based on the
extraction of frequency parameters from each UEMG si-
gnal because frequency characterizes each signal. Then,
signals are being classified according to differences and
similarities in their frequency parameters as viewed by
simple perceptron neuron.
The classification method was applied on the simul-
ated signals, and it gave very good results. Therefore, we
have found that our supervised method of classification
is an efficient way to distinguish between preterm labor
and labor at term.
Various clinical techniques used for this classification
such as the use of fetal fibronectin and tocodynamometer
(TOCO), all have a limited range of usefulness in this re-
gard [12]. Even noticeable dynamic cervical change, long
thought to be indicative, may not always be an accurate
identifier of true labor, since a large percentage of wom-
en with established cervical dilation do not actually de-
liver preterm, even when they are not treated with labor-
inhibiting, or tocolytic agents [13]. Moreover, these me-
thods do not lead predictive value above 65%. Therefore,
we can suggest the use of ANN as a more effective
method of classification (96.7% predictive value).
As a perspective, we believe that an improvement
might be attained by using multi layer perceptron neuron
(MLP).
REFERENCES
[1] Mckean, M., Walters, A.W.W. and Smith, R. (1993)
Prediction and early diagnosis of preterm labor: A criti-
cal review. Obstetrical & Gynecological Survey, 48(4),
209-225.
[2] Senat, M.V., Tsatsaris, V., Ville, Y. And Fernandez, H.
(1999) Menaced’accouchement prématuré. Encycl Méd
Chir (Elsevier, Paris), Urgences, 17.
[3] Marque, C.K., Terrien, J., Rihana, S. and Germai, G.
(2007) Preterm labour detection by use of a biophysical
marker: The uterine electrical activity. BMC Pregnancy
and Childbirth.
[4] Newman, R.B., Gill, P.J., Campion, S. and Katz, M.
(1987) Antepartum ambulatory tocodynamometry: The
significance of low-amplitude, high-frequency contrac-
tions. Obstetrics & Gynecology, 70(5), 701-750.

M. O. Diab et al. / J. Biomedical Science and Engineering 3 (2010) 837-842
842
JBiSE
[5] Marque, C., Duchêne, J., Leclercq, S., et al. (1986) Uter-
ine EHG processing for obstetrical monitoring. IEEE
Transactions on Biomedical Engineering, 33(12), 1182-
1187.
[6] Diab, M.O., Marque, C. and Khalil, M.A. (2007) Classi-
fication for uterine EMG signals: Comparison between
AR model and statistical classification method. Interna-
tional Journal of computational cognition, 5(1), 8-14.
[7] Khalil, M. and Duchene, J. (1999) Detection and classi-
fication of multiple events in piecewise stationary signals:
Comparison between autoregressive and multiscale ap-
proaches. Signals processing, 75(3), 239-251.
[8] Hayes, M. (1996) Statistical digital signal processing and
modelling. John Wiley & Sons, Georgia Institute of
Technology.
[9] Kay, S. (1988) Modern spectral estimation theory and
application. Englewood Cliffs, Prentice-Hall, New Jersey.
[10] Gurney, K. (1997) An introduction to neural networks,
University College London Press.
[11] Lockwood, C.J. and Kuczynski, E. (2001) Risk stratifica-
tion and pathological mechanisms in preterm delivery.
Paediatric and Perinatal Epidemiology, 15(Suppl 2),
78-89.
[12] Iams, J.D. (2003) Prediction and early detection of pre-
term labor. American Journal of Obstetrics & Gynecol-
ogy, 101(2), 402-412.
[13] Linhart, J., Olson, G., Goodrum, L., Rowe, T., Saade, G.
and Hankins, G. (1990) Preterm labor at 32 to 34 weeks
gestation: Effect of a policy of expectant management on
length of gestation. American Journal of Obstetrics &
Gynecology, 178-179.