Term and Preterm prediction by Uterine EMG Analysis

By:

Kishor P. Kshirasagar Electrical and Electronics Engineering National Institute of Technology Karnataka, Surathkal

Sharukh S. Shaikh Electronics and Communication Engineering National Institute of Technology Karnataka, Surathkal

Guided by: Prof. Kumar Sankar Ray Electronics and Communication Sciences Unit Indian Statistical Institute, Kolkata

Abstract:
Preterm labor is an urgent challenge in current healthcare. Many handicaps and 85% of prenatal deaths are a because of preterm birth. 10% of births each year are preterm with birth-weights of less than 5.5lbs [2].Complications of preterm birth includes neurological, mental, behavioral, and pulmonary problems. In survivors of preterm birth, neurological impairment varies from 10% to 20% and growth is restricted 20%.The ability to identify preterm labor is crucial in the survival of an infant born prematurely. It is widely accepted that uterine contractions are generated by the electrical activity originating from the depolarization-repolarization of billions of smooth muscle myometrial cells. Electromyography (EMG) is a way to monitor the electrical activity in a muscle or groups of muscles. In the case of a pregnant woman, uterine contractility used to be monitored by placing electrodes directly on the uterus. Most recently, the same results are trying to be obtained using noninvasive procedures. The electrodes are placed directly on the abdominal surface. The data, once obtained, must be analyzed to obtain a diagnostic result. To analyze the data, Artificial Neural Networks (ANNs) are used. ANNs are ideal for the classification of objects (e.g., patients) based upon one or more input variables (e.g., uterine EMG variables) An ANN is made up of a circuit or computer program (equivalent of neurons in the brain of living creatures) that has three main parts. The three parts are the inputs, the hidden layers to process the data, and the outputs. In the case of the uterine electromyography, the inputs are multiple

parameters calculated using statistical analysis.

Introduction:
Premature labor prediction is an extremely difficult task. This is due to the lack of knowledge regarding the exact physiology of the uterus [1, 10] and parturition. Premature labor prediction so far has mostly been based on calculating the risk factors. Although many risk factors [1] were identified such as diabetes, conization, hypertension, smoking, abnormalities of the uterus, short cervix, a positive fibronectin test and others, premature labor prediction is far from certain. Analysis of uterine Electromyogram (EMG), termed as Electrohysterogram (EHG), records is one such technique. The EHG records correspond to the activity of the uterine muscles and might therefore be used to predict the premature onset of labor [6]. The signal acquisition is both noninvasive and relatively simple and could therefore easily be introduced into hospital practice. Using the EHG it is possible to detect uterine activity related to contractions during both gestation and active labor. The EHG could therefore supplement the tools currently used to monitor labor [6]. Uterine electromyography (EMG) has shown great promise for monitoring patients during pregnancy [3, 4, 5]. Early studies established that the electrical activity of the myometrium, or uterine muscle, is responsible for myometrial contractions. Many experiments have been performed in the last several decades in an attempt to monitor uterine contractility using the electrical activity measured from electrodes placed directly on the uterus. The latest work indicates that uterine

EMG activity can be monitored noninvasively from the abdominal surface. Once uterine EMG data are obtained, however, they must be assimilated in some fashion to get a diagnostic result.

data, signal loss, or broken connection between the skin and the electrodes, or no

Database description:
The EHG records [11] used in this research were collected from 1997 until 2006 at the Department of Obstetrics and Gynecology, Medical Centre Ljubljana, Ljubljana [1]. Records were collected from the general population as well as from the patients admitted to the hospital with the diagnosis of impending pre-term labor. One record per pregnancy was recorded. The records are of 30-min duration and consist of three channels. The sampling frequency, fs, was 20 Hz. The records were collected from the abdominal surface using four AgCl2 electrodes (see Fig. 1). The electrodes were placed in two horizontal rows, symmetrically under and above the navel, spaced 7 cm apart. A special protocol was used during the attachment of the electrodes in order to improve the quality of the measurements. According to the protocol, the resistance between the electrodes had to be lower than 100 k. The first acquired signal was measured between the topmost electrodes (E2E1), the second signal between the leftmost electrodes (E2E3) and the third signal between the lower electrodes (E4E3). Prior to sampling the signals were filtered using an analog three pole Butterworth filter with the bandwidth from 0 to 5 Hz. The resolution of the scanning system was 16 bits with the amplitude range 2.5 mV. Due to the large scope of the research, some recording errors were inevitable, e.g., missing accompanying

Fig: 1. The placement of the electrodes on the abdomen, above the uterine surface. Signal 1: E2E1, signal 2: E2E3, signal 3: E4E3 electrical activity.

Fig: 2. The placement of the electrodes on the abdomen, above the uterine surface.

After a careful visual inspection, and after rejecting those records of pregnancies containing no electrical activity or containing excessive noise, those ended in C-sections and those ended in induced delivery, 300 EHG records (300 pregnancies) ending in term or preterm deliveries out of a total of 1,211 records were chosen for further analysis. The times of deliveries versus recording times for the 300 EHG records are shown in Fig. 3. The following groups of EHG records were formed: 1. 262 records from pregnancies where the deliveries were term (pregnancy duration C37 weeks) of which: (a) 143 records were recorded early, before the 26th week of gestation; (b) 119 records were recorded later, during or after the 26th week of gestation. 2. 38 records from pregnancies which ended prematurely (pregnancy duration\37 weeks) of which: (a) 19 records were recorded early, before the 26th week of gestation; (b) 19 records were recorded later, during or after the 26th week of gestation. We put special attention to records recorded early. The frequency of contractions early in the pregnancy is relatively low. Therefore, we selected only those records showing some visually detected electrical activity different from noises.

Butterworth digital filter), 0.054 Hz, 0.2 4 Hz, and filtering methods including wavelets were used.

Fig. 3. The times of delivery in relation to the times of recording for the 300 EHG records. Open circles term delivery records, filled circles pre-term delivery records, the dashed horizontal line indicates the boundary (37th week of gestation) between the term and preterm delivery records, upper left group of term delivery records recorded early (before the 26th week of gestation), upper right group of term delivery records recorded later (during or after the 26th week of gestation), lower left group of pre-term delivery records recorded early, lower right group of pre-term delivery records recorded later.

Preprocessing:
The selection of digital filters to remove noise from signals before the processing may greatly influence the results. A bandpass filter is needed. Various frequency bands, such as 0.084 Hz (using a

It was recognized that the uterine EMG content ranges from 0 to <5 Hz. We chose digital Butterworth filters which have a smooth frequency response and are computationally non-intensive. Their major drawback, the phase-shifting, is especially troublesome when using highpass filtering. Fortunately, the phase-shift can be eliminated by filtering the whole signal twice in different directions, forward and then again backward, thus obtaining a well filtered signal with zero phase shift. The chosen four pole Butterworth filters were applied bidirectionally to each signal. We used the band-pass filter 0.33 Hz [6].

delivery record, it seems that only one physiologic mechanism is involved and the muscle activities are more predictable. The main goal of this study was to explore the possibility of classification of term and pre-term delivery EHG records [9].

Feature extraction:
Cross correlation coefficients: The correlation coefficients are the zero lag of the normalized covariance function [12]. Two sets of term and preterm sample recordings are used as training samples. The testing sample is concatenated with the term and preterm training datasets separately and then correlation coefficients are calculated as below, Variance is the measure of the spread of data in a data set. In fact it is almost identical to the standard deviation. The formula is:

Fig. 4. Examples of EHG signals. A Signal 1 of a term delivery record (recorded in 30th week, delivery in 39th week), b signal 1 of a preterm delivery record (recorded in 30th week, delivery in 32nd week).

The figure 4 shows two examples of EHG signals from different records. Both records were recorded in the 30th week of gestation. The upper signal is of a term delivery EHG record (delivery in 39th week of gestation), while the lower signal is of a pre-term delivery EHG record (delivery in 32nd week). The signals were filtered using digital a filter with the bandwidth from 0.3 to 3 Hz. Looking at the signals of uterine muscle activities of both signals, one may argue that different physiologic mechanisms are involved. In the signal of the term delivery record, it seems that multiple physiologic mechanisms or non-linear processes are involved in the background, and the muscle activities are less predictable; while in the signal of the pre-term

The formula for covariance is very similar to the formula for variance. The formula is:

The covariance is always measured between 2 dimensions. If we have a data set with more than 2 dimensions, there is more than one covariance measurement that can be calculated. For example, from a 3 dimensional data set (dimensions x, y, z) could calculate cov(x, y), cov(x, z), and cov(y, z). The covariance matrix is given by

The matrix R of correlation coefficients calculated from an input matrix X whose rows are observations and whose columns are the test data concatenated with the training data set. The matrix R is related to the covariance matrix C by the formula

The median frequency is defined as the frequency just above where the sums of the parts above and below in the frequency-power spectrum, P, are the same. Autocorrelation zero-crossing: The autocorrelation zero-crossing is defined as the first zero-crossing starting at the peak in the autocorrelation, Rxx(s), of the signal x(t):

If the absolute value of R (i , j) is close to 1 then the ith data of input matrix has more resemblance to jth data. Extracting the first row of the matrix R excluding its first element gives the cross correlation coefficients of the test data with respect to each training data set (Term and Preterm datasets separately for early and late cases). The mean square value of the cross correlation row vector extracted from the matrices R corresponding to the preterm and term training datasets are found and compared. The vectors corresponding to preterm and term training datasets with mean square value more close to unity dictates the class to which the test sample data belongs. This correlation test gives a very high overall accuracy of about 91.34%. And when considered separately for Preterm and Term database has an accuracy of 89.47% and 93.51% respectively. During this computation, the number of training samples used is 15, each from the set of early and late preterm database and 50 samples, each from the set of early and late term database. Median frequency of the signal power spectrum: The median frequency, fmed, is calculated as:

Sample entropy: The sample entropy, sampEn, is a measure of the regularity of finite length time series. Less predictable time series exhibit higher sample entropy. Given a time series x(t) of length N, and patterns aj(0,,m-1) of length m, m<N, where the patterns aj are taken from the time series x(t), aj(i) = x(i + j), i = 0,,m-1, j = 0,,N-m; the part of the time series x(t) at time t = ts, x (ts,,ts+m-1) is considered as a match for a given pattern aj if |x(ts + i)-aj(i)| r for each 0i<m. The number of pattern matches (within a margin of r), cm, is constructed for each m. The sample entropy, sampEn, is then defined as: SampEnm, r(x) =

During evaluation of the technique, we varied the parameter m from 2 to 4 in steps of 1, and the parameter r from 0.1 to 0.2 in steps of 0.125.

Skewness: Skewness is a measure of the asymmetry of the data around the sample mean. If skewness is negative, the data are spread out more to the left of the mean than to the right. If skewness is positive, the data are spread out more to the right. The skewness of the normal distribution (or any perfectly symmetric ymmetric distribution) is zero. The he skewness of a distribution is defined as

pregnant patients into those who w have labor. . The high percentage of correctly classified patients, and the significant difference in values of the electrical parameters ters for ANN-sorted ANN groups, is proof that the method is effective. The confusion matrix is shown in the figure below.

where is the mean of x, is the standard deviation of x, and E (t) t) represents the expected value of the quantity t. Kurtosis: Kurtosis is a measure of how outlieroutlier prone a distribution is. The kurtosis of the normal distribution is 3. Distributions that are more outlier-prone prone than the normal distribution have kurtosis greater than 3; distributions that are less outlieroutlier prone rone have kurtosis less than 3. The kurtosis of a distribution is defined as
Fig. 5. Confusion Matrix

where is the mean of x, is the standard deviation of x, and E (t) represents the expected value of the quantity t. t

Conclusion:
As in previous studies, we established herewith that non-invasive invasive transtrans abdominal uterine EMG measurements can be used to effectively monitor pregnant patients. Artificial neural networks, in conjunction with uterine EMG data, seem to be an effective method for classifying both term and preterm

The class 1 refers to preterm dataset and the class 2 belongs to the term dataset. The accuracy for preterm dataset is 86.8% and the accuracy for term dataset is 97.3%. And the overall accuracy of about 96%. The best uterine EMG classification input parameters parame for the ANN in this study were the correlation coefficients, skewness, autocorrelation first zero crossing and sample entropy. entropy These resulted in a sufficiently high classification rate. These parameters have previously been linked to contraction strength. The remaining parameters like RMS value of signal and peak frequency

investigated actually reduced the predictive capability of the ANN when processed by the classification algorithm. We suppose that this is because they have little, if any, physiological significance, or at least they seem to have little diagnostic relevance. However, other sophisticated and perhaps "less-traditional" calculated uterine EMG parameters not considered in this study (e.g., propagation velocity, fractal dimension, wavelet energy, or Lyapunov exponents) should also be investigated for patient classification capabilities, using them as input variables for the ANN in future work.

References:
[1] A comparison of various linear and non-linear signal processing techniques to separate uterine EMG records of term and pre-term delivery groupsG. Fele-Zorz G. Kavsek Z. NovakAntolic F. Jager. [2] Identification of Human Term and Preterm labor using Uterine ElectromyographyJenna Marcus, Biomedical Engineering Seminar III, April 16, 2007. [3] Comparing uterine electromyography activity of antepartum patients versus term labor patientsRobert E. Garfield, PhD,* William L. Maner, BS, BA, Lyn B. MacKay, BA, Dietmar Schlembach, MD, George R. Saade, MD. [4] Classification of uterine EMG signals using supervised classification methodMohamad O. Diab, Amira El-Merhie, Nour El-Halabi, Layal Khoder. [5] Classification for Uterine EMG Signals: Comparison between AR Model and Statistical Classification MethodMohamad O. Diab, Catherine Marque and Mohamad A. Khalil.

[6] Identification of Human Term and Preterm Labor using Artificial Neural Networks on Uterine Electromyography Data.-WILLIAML MANER and ROBERT E. GARFIELD, Department of Obstetrics and Gynecology, University of Texas Medical Branch, 301 University, Route 1062, Galveston, TX 77555, USA (Received 20 April 2006; accepted 7 December 2006; published online 17 January 2007). [7] Use of the Electrohysterogram Signal for Characterization of Contractions during Pregnancy-Hel`ene Leman,* Catherine Marque, and Jean Gondry. [8] Biophysical methods of prediction and Prevention of preterm labor: uterine Electromyography and cervical lightinduced fluorescence new obstetrical diagnostic techniques Robert E Garfield and William L Maner [9] -Non-invasive transabdominal uterine Electromyography correlates with the strength of intrauterine pressure and is predictive of labor and deliveryH. Maul, W. L. Maner, G. Olson, G. R. Saade and R. E. Garfield Division of Reproductive Sciences, Department of Obstetrics and Gynecology, The University of Texas Medical Branch, Galveston, Texas, USA [10]The physiology of uterine contractions- Holger Maul, MD, MMS, William L. Maner, BSc, George R. Saade, MD, Robert E. Garfield, PhD. [11]http://www.physionet.org/physioba nk/database/tpehgdb/ [12]Fundamental Statistics tutorials