ENHANCED ANALYSIS OF UTERINE ACTIVTY USING SURFACE

ELECTROMYOGRAPHY

A. Herzog, L. Reicke, M. Kr¨oger

Institute of Dynamics and Vibration Research, Leibniz University Hannover, Germany

C. Sohn, H. Maul

Obstetrics and Gynecology, UniversityHospital Heidelberg, Germany

Keywords:

Uterine, electromyography, pulse detection, stochastic analysis, Karhunen-Lo`eve, principal component.

Abstract:

This contribution presents a new approach for the enhanced analysis of uterine surface electromyography

(EMG). First, a pulse detection separates the pulses, which contain the essential information about the uterine

contractibility, from the ﬂat line sections during relaxation. The functionality of this semi-automatic algorithm

is controlled by two comprehensible parameters. Subsequently, the mean frequency, which serves as a crite-

rion for imminent delivery, is evaluated from the extracted pulses. Although the pulse detection reduces the

deviation of the mean frequency signiﬁcantly, the results are still sensitive to parameter variations in the pulse

detection. A stochastic analysis based on the Karhunen-Lo`eve transform (KLT) derives generalised patterns,

the eigenforms, from the pulse ensemble. The mean frequency of the ﬁrst eigenform is less sensitive to pa-

rameter variations. Additionally, the correlation between the eigenforms of the left and right surface electrode

can serve as a criterion for the measurement’s quality.

1 INTRODUCTION

Even in modern obstetrics, the point of delivery can-

not be precisely predicted. Although the majority

of pregnancies passes without any complications, the

signiﬁcance of an enhanced analysis of uterine activ-

ity arises from the diagnosis of preterm labor as well

as the treatment of delayed delivery.

The uterine muscle (myometrium), which has

maintained a quiescent state during the majority of

pregnancy, is prepared for labor by local contractions.

These contractions, called training labors, improve

the synchronisation between the single muscle cells

in order to obtain a deﬁned contraction sequence dur-

ing delivery. Therefore, the identiﬁcation of imminent

labor requires an elaborate analysis and interpretation

of this preparatory phase.

Several methods for the evaluation of uterine con-

tractibility are commonly used: TOCO, IUPC and

EMG. Uterine contractions cause variations in the

abdomen’s contour, which can be detected by pres-

sure sensors. Due to the indirect measurement, this

so-called external tocodynamometry (TOCO) is not

sensitive and reliable enough. A more reliable ap-

proach consists of measuring the uterine’s internal

pressure (intrauterine pressure catheter, IUPC). The

surface electromyography (EMG) combines the non-

invasiveness property of TOCO with a sensitivity sim-

ilar to that of the IUPC (Maul et al., 2004). The

muscular activity is accompanied by variations of the

electric potential at the neuromuscular junction be-

tween nerve and muscle cells. This potential can

be picked up directly by needle electrodes and range

from −70mV (relaxation) up to +30mV during con-

traction. In case of surface electrodes, the voltage has

to be transmitted via the tissue to the skin, yielding to

lower peak values as well as deformations in the time

history of the voltage signal. For the measurement

of the uterine contractions two surface electrodes are

used. They are located on the right and left side of the

abdomen. The time-history of a single electromyo-

gram (EMG)-signal is displayed in Figure 1 above.

The pulses, which belong to uterine contractions, are

separated by ﬂat line sections. Up to now, the fre-

quency characteristics have been derived from large

sections of its time history by means of the Fourier

transform. Based on the assumption that ongoing syn-

chronisation leads to an increase of the pulse’s attack

377

Herzog A., Reicke L., Kröger M., Sohn C. and Maul H. (2008).

ENHANCED ANALYSIS OF UTERINE ACTIVTY USING SURFACE ELECTROMYOGRAPHY.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 377-384

DOI: 10.5220/0001062903770384

 SciTePress

0 100 200 300

level [mV]

time [s]

frequency [Hz]

log10 [1mV]

0 100 200 300

−3

−2

−1

Figure 1: Surface Electromyogram: Time history (above)

and its short-time Fourier transform (below).

and decay slope, the mean frequencyof the calculated

spectrum serves as a criterion to judge imminent de-

livery.

But a detailed analysis in the time-frequency do-

main reveals a strongly varying frequency content.

This analysis is done by a discrete short-time Fourier

transform (STFT) based on a Hann-Window 5s long.

An additional zero-padding and a logarithmic scal-

ing of the resulting amplitude-coefﬁcients unveils all

the signiﬁcant details. An introduction into the time-

frequencytransforms can be foundin (Mertins, 1999),

practical aspects are discussed e.g. in (Reicke et al.,

2006).

The logarithmic representation of the STFT co-

efﬁcients in Figure 1 does not only show the broad

frequency content of the pulses, it even unveils the

heart beat of the foetus at 1.6Hz as well as its harmon-

ics. Due to the fact that the pulses rather than the ﬂat

line sections contain the information about the uterine

contractibility, the authors suggest an enhanced anal-

ysis which is restricted to the EMG-pulses. This new

approach is supported by Figure 2. The diagram on

the top shows the instantaneous mean frequency

(t) =

∞

f ·

STFT

( f,t)|

||X

STFT

(t)||

df (1)

derived from the amplitude coefﬁcients X

STFT

( f,t) of

the STFT. The norm ||X

STFT

(t)|| denotes the instanta-

neous energy

STFT

( f,t)

|df of the STFT. The red

line represents the level of the original EMG-signal.

The lower diagram shows the evolution of the mean

frequency’s standard deviation

(t) =

∞

( f − f

)

STFT

( f,t)|

||X

STFT

(t)||

df (2)

over time. It underlines that a reliable estimation of

the mean frequency is restricted to the pulses. Only

in these time intervals, the standard deviation is less

than 3Hz.

0 100 200 300

mean freq. [Hz]

0 100 200 300

level [mV]

0 100 200 300

time [s]

std. dev. [Hz]

Figure 2: Instantaneous mean frequency (above) and its

standard deviation (below) of the signal shown in Fig. 1.

As only the pulses contain the relevant informa-

tion, it is convenient to analyse the pulses without

the intervals of relaxation. This contribution presents

a semi-automatic pulse detection, which extracts the

pulses out of the measured EMG-signal. The expres-

sion semi-automatic underlines that the operation is

controlled by the physician, whereas the algorithm

undertakes the time-consuming and exhausting work

of scanning through the signal searching for pulses.

Additionally, the use of surface electrodes causes de-

formations of the pulse shape. Therefore, the pulses

are processed by a stochastic method based on the

Karhunen-Lo`eve transform to evaluate a generalised

pattern.

2 PULSE DETECTION

2.1 Conditioning

The surface-EMG signals are distorted by noise and

a low frequency drift. A low-pass ﬁlter, which rejects

frequencies higher than 7.5Hz, is applied to attenu-

ate the noise. The low frequency drift is reduced by a

high-pass ﬁlter with a cut-off frequency of 0.1Hz and

a transition band of 0.2Hz. Both are implemented as

ﬁnite impulse response (FIR) ﬁlters based on a Kaiser

window design (Oppenheim and Schafer, 1999). An

additional noise-reduction is achieved by the pulse

detection: the ﬂat line intervals, which are charac-

terised by a low signal-to-noise ratio, are excluded

from the further analysis.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

378

2.2 Pulse Detection

The pulse detection extracts those parts of the sig-

nal which contain the relevant information about the

uterine contractibility. The localisation of the pulses

is done regarding the magnitude of the signal. First,

the global maximum and the mean value of the sig-

nal’s magnitude as an approximation of the noise level

are determined. All local peaks in the range between

these two values can be considered as potential pulse

centres. But only pulses whose peak values largely

exceed the noise limit offer a sufﬁcient signal-to-noise

ratio. Therefore, the ﬁrst parameter of the pulse detec-

tion, the level value, is introduced. This value deter-

mines the percentage of the range between noise level

and global maximum which is added to the noise level

in order to deﬁne the lower level limit. If the level

value is chosen equal to zero, the lower level limit is

identical to the noise level. Hence, any peak value

higher than the noise level is considered as a pulse

centre. In case of a level value equal to ”1”, the lower

level limit reaches the global maximum and any peak

except for the global maximum will be rejected. Fig-

ure 3 shows a particular lower level limit which cor-

responds to a level value of 0.3.

-2

1000

2000

3000

time[s]

level[mV]

global

maximum

noiselevel

lowerlevel

limit

Figure 3: EMG-signal (above) and its magnitude with

global maximum, noise level and a particular level value.

After the localisation of the pulse centres in the

signal, the initial and end point of each pulse are de-

termined. The pulse detection is based on the assump-

tion that a pulse begins and ends at roots. Therefore,

any low frequent drift has to be removed (cp. 2.1)

before the execution of the pulse detection algorithm.

Starting with the pulse centre, the adjoining roots tem-

porary describe the initial and end points. In the fol-

lowing, this part of the pulse between these two roots

is called the inner pulse. If these points were ﬁnally

considered as the pulse’s initial and end points, ad-

jacent over- and undershoots, which might belong to

the pulse and therefore contain valuable information,

would not be extracted. Hence, the surroundings of

the inner pulse have to be taken into consideration.

-1

3190

3195

3200

time[s]

level[mV]

inner

pulse

undershoot

Figure 4: Evaluation of inner and outer area.

Figure 4 displays a pulse with a preceding and

subsequent undershoot. To determine whether these

undershoots are part of the pulse, the roots before and

after the temporary initial and end points are consid-

ered. For example, the temporary end point and the

root located on its right enclose the subsequent under-

shoot. Now, the area of the undershoot is calculated

and related to the area of the inner pulse. In Figure 4,

the area of the undershoot (”outer area”) and the in-

ner pulse (”inner area”) are hatched in black and red,

respectively. If the ratio of the outer and inner area

exceeds a given value, the corresponding undershoot

belongs to the pulse. If the right undershoot in Figure

4 fulﬁls this area criterion, the inner pulse is expanded

by the right undershoot and the temporary end point

is shifted by one root to the right.

This given value is called the area value and can

be chosen anywhere between ”0”, which connects any

adjacent undershoot to the inner pulse, and ”1”. In

case of an area value equal to ”1”, only undershoots

exhibiting an area greater than the inner area are at-

tached to the pulse. The same procedure is done with

the undershoot on the left. This algorithm goes on

in both directions until the area of the current over-

or undershoot is less than that of the original inner

pulse. In this case, the temporary root becomes the

ﬁnal root, which borders the pulse to one side. As

soon as the left and right ﬁnal roots are determined,

the pulse can be extracted from the signal. The de-

tection of the next pulses follows the same algorithm.

In order to avoid overlapping of closely neighbouring

pulses, the extracted pulse data are replaced by zeros.

The level value inﬂuences the quantity of detected

pulses. The lower the level value, the more peaks

of the signal are regarded as pulse centres. The area

value controls the lengths of the pulses. The greater

the area value, the less over- and undershoots belong

to the inner pulse and therefore the less pulses are

ENHANCED ANALYSIS OF UTERINE ACTIVTY USING SURFACE ELECTROMYOGRAPHY

379

lengthened beyond their inner pulse. However, the

area value has an inﬂuence on the quantity of the

pulses, too. If the area value is very low, the pulses

extracted from the signal are so long that less pulses

can be detected in the remaining signal parts.

2.3 Characteristic Values

The pulse detection scans through the signal and cuts

out single time histories belonging to those pulses

whose shapes match the pattern speciﬁed by the level

and area value. The extracted pulses are described

in the time domain by their peak values and lengths.

Additionally, each pulse is analysed in the frequency

domain by the discrete Fourier transform (DFT). Con-

trary to the short-time Fourier transform X

STFT

( f,t)

of the entire signal, the spectrum X

pulse

( f) of an indi-

vidual pulse is not time-dependent. Hence, each pulse

is characterised by two values in the frequency do-

main, the mean frequency f

and the variance σ

Based on a measured EMG-signal, Figure 5 shows

the characteristic values of those pulses that fulﬁll a

level value equal to ”0.3”, i.e. 30 % of the global max-

imum, and an area value of ”0.4”. In the diagrams, the

horizontal line denotes the arithmetic mean. The third

diagram exhibits a strong variation in the pulse length.

Particularly, the 2nd and 4th pulse length strongly

deviate from the mean of ≈ 650samples. The 2nd

pulse’s mean frequency f

largely exceeds the mean

of ≈ 0.16Hz. The reason may be the short duration

of ≈ 200Samples, which also increases the variance

In order to demonstrate the inﬂuence of the two

parameters level value and area value on the num-

ber, length and mean frequencyof the pulses extracted

from the EMG-signal, the results of ﬁve different

pairs, shown in Table 1, are compared in Figures 6,

7 and 8.

Table 1: Pairs of parameters used for pulse detection.

parameter level area

settings value value

E1 0.05 0.1

E2 0.1 0.1

E3 0.1 0.3

E4 0.3 0.4

E5 0.3 0.7

For each pair, denoted with E1 up to E5, the left

and right bars represent the left and right channel of

the electromyogram, respectively. With increasing

level value, the number of pulses decreases because

pulses with lower peak values are now rejected. A

comparison between pair E2 and E3 as well as E4

0.2

0.1

0.04

0.02

1000

500

variance

[Hz]

length

[samples]

meanfrequency

[Hz]

peakvalues

[Hz]

2 3

pulses

Figure 5: Characteristic values of a pulse ensemble cut out

from a measured EMG-signal.

and E5 unveils the inﬂuence of the area value on the

length and numberof pulses. An increasing area value

leads to a shorter maximum pulse length. On the

other hand, a higher level value increases the mini-

mum pulse length because the pulses with low peak

values, which are obviously shorter, are not consid-

ered anymore. As it is mentioned in 2.2, a more re-

strictive area value can resolve and separate closely

neighbouring pulses into two individual pulse shapes.

E3 E4 E5

number

ofpulses

Figure 6: Inﬂuence of parameters in Table 1 on number of

pulses extracted.

200

100

length

[samples]

5000

Figure 7: Inﬂuence of parameters in Table 1 on minimum

and maximum pulse length.

0.2

0.1

E4 E5

frequency

[Hz]

Figure 8: Inﬂuence of parameters in Table 1 on mean fre-

quency.

The variation of the pulse length takes effect on

the mean frequency, which is shown in Figure 8. It

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

380

reveals a strong sensitivity of the mean frequency to

the given level and area values. This insufﬁcient un-

certainty motivated the authors to improve the anal-

ysis by a stochastic signal processing, which is de-

scribed in the next section. Additionally, a criterion is

required to give reliable information about the qual-

ity of the signal. A straightforward approach is the

consideration of the signal-to-noise ratio, but this will

not take into account any correlation between the two

channels .

3 STOCHASTIC ANALYSIS

3.1 Karhunen-Lo

eve Transform

The electric potential, which occurs at a neuromuscu-

lar junction, is transmitted via a quite complex elec-

tric network to the two electrodes at the abdomen’s

surface. This leads to an amplitude attenuation of

≈ −20dB and deformations in the pulse shape. In ad-

dition, the location of the contraction is randomly dis-

tributed over the entire uterine muscle, which implies

a random distortion of the EMG-signal with regard to

the peak value and pulse shape.

Due to the fact that the measuring time of about

30min is very short compared to the ongoing preg-

nancy, a stationary stochastic process is assumed. The

individual pulse shapes extracted by the pulse detec-

tion are considered as the realisations of this stochas-

tic process. The new approach uses the Karhunen-

Lo`eve transform (KLT), also referred to as Principal

Component Analysis (PCA), to determine a charac-

teristic pulse shape out of the pulse ensemble.

The Karhunen-Lo`eve transform is a signal-

depending decomposition based on the covariance

matrix

˜x˜x

= E



˜x ˜x



, (3)

in which E{ } denotes the statistic expectation and ˜x

the stochastic process. The decomposition requires

the eigenvectors u of the eigenvalue problem

˜x˜x

u = λ u. (4)

The eigenvectors u can be regarded as the characteris-

tic shapes of the stochastic process. The eigenvalue λ

represents the degree of similarity between the corre-

sponding eigenvector and the individual pulses. In the

following, the product of eigenvector and eigenvalue

is denoted as eigenform. The more similar the individ-

ual pulses of the ensemble are to each other, the more

dominant becomes the ﬁrst eigenform. If the ensem-

ble consists of identical pulse shapes, the ﬁrst eigen-

value will contain the whole variance of the stochastic

process, while all other eigenvalues are equal to 0. A

brief introduction into the Karhunen-Lo`eve transform

is given in (Mertins, 1999), a detailed description can

be found in (Jolliffe, 2002).

-1

-2

-3

time [s]

pulses

level[mV]

Figure 9: Ensemble of centred pulses.

For the stochastic analysis, a preprocessing of the

pulses is necessary. The calculation of the covariance

matrix requires an identical length of all the pulse

shapes. Therefore, the pulses are centered with re-

gard to their centres of area, followed by padding ze-

ros on both sides to obtain an identical pulse length.

The result of this preprocessing is shown in Figure 9,

in which the longest pulse, the blue one, speciﬁes the

dimension of the covariance matrix. The other pulse

shapes are shifted in such a way that all area centres

coincide.

The result of the KLT of Figure 9 is displayed

in Figure 10: The ﬁrst and second eigenforms (EF)

are dominant and contain ≈ 90% of the process’ vari-

ance. This can be seen from the time history on the

left side as well as the loadings on the right. In this

context, the loading denotes the normalised variance

of the stochastic process.

1.EF:67,13%

2.EF:22,51%

3.EF:4,66%

4.EF:2,67%

5.EF:1,81%

6.EF:0,82%

1,5

0,5

-0,5

level[mV]

2 4

8 10 12

time[s]

1.EF

2.EF

Figure 10: KLT: Eigenforms (left) and loadings (right).

Instead of deriving the mean frequency directly

from the pulse ensemble (cp. subsection 2.3), the

DFT of the ﬁrst eigenform yields to a mean frequency

which is less sensitive to parameter variations. On

top of Figure 11, the global mean frequency, which

is evaluated as the arithmetic mean of the individual

eigenforms’ mean frequencies, is displayed accord-

ing to the parameter settings shown in Table 1. While

ENHANCED ANALYSIS OF UTERINE ACTIVTY USING SURFACE ELECTROMYOGRAPHY

381

the global mean frequency is susceptible to parameter

variations, the mean frequency derived from the ﬁrst

eigenform seems to be less sensitive. Here, research

is in progress to conﬁrm this observation.

0.1

0.2

frequency

[Hz]

frequency

[Hz]

Figure 11: KLT: global mean frequency (above) and mean

frequency of the ﬁrst eigenform (below) for different pa-

rameter settings.

The Karhunen-Lo`eve transform does not only ex-

tract a characteristic pulse pattern, the ﬁrst eigenform,

out of the pulse ensemble. It can also provide a reli-

able criterion of the electromyogram’s reliability. If

the ﬁrst eigenvalue is dominant, the pattern of the ﬁrst

eigenform is similar to the shapes of the majority of

pulses in the ensemble, while the other eigenforms

represent the deformations in the pulse shapes.

3.2 Correlation Analysis

So far, the two channels of the electromyogram have

been analysed separately. In case of a dominant

eigenvalue (see subsection 3.1), the corresponding

eigenform characterises the pulse pattern of the in-

dividual channel’s pulse ensemble very well. There-

fore, the correlation between the left and right EMG-

channel can be evaluated by regarding their ﬁrst

eigenforms.

Even in case of identical pulse shapes, a time shift

between the left and right eigenform can occur. This

may be caused by different transmission delays from

the neuromuscular junction to the surface electrodes

in combination with the centring of the pulses before

the KLT is performed.

In order to evaluate the similarities between the

left and right eigenform u

ℓ

(t) and u

(t) the cross-

correlation function (CCF)

ℓ

(τ) =

∞

−∞

ℓ

(t) · u

(t + τ) dt (5)

is used. If the two eigenforms are of identical shape

but shifted to each other, u

(t) = u

ℓ

(t − ∆t), the CCF

resembles an autocorrelation function (ACF) whose

maximum value is shifted along the time axis. Due

to the fact that an ACF is symmetric to its origin τ =

0, the CCF of two identical but shifted eigenforms is

symmetric with regard to the time shift ∆t:

ℓ

(∆t − τ) = R

ℓ

(∆t + τ). (6)

Because the eigenvector’s orientation is not speciﬁed

by Equation 4, the left and right eigenforms can dif-

fer in their signs. Therefore, maximum correlation in

the CCF appears at its global maximum or minimum.

First of all, the time shift of the CCF is determined by

its global extremum. Subsequently the CCF is divided

into a symmetric

symm

(τ) =

ℓ

(∆t + τ) + R

ℓ

(∆t − τ)

(7)

and antimetric

anti

(τ) =

ℓ

(∆t + τ) − R

ℓ

(∆t − τ)

(8)

component. This decomposition is displayed in Fig-

ure 12. The extremum is located at ≈ 2000samples,

at which a vertical symmetry axis (dashed line) is

drawn. According to Equations 7 and 8 the cross-

correlation function (blue line) is decomposed into its

symmetric (green line) and antimetric (red line) com-

ponents.

0,02

0,01

-0,01

1000 2000

3000

CCF[Hz]

samples

CCF

antimetric

symmetric

Figure 12: Decomposition of the CCF (blue) at its ex-

tremum into a symmetric (green) and antimetric (red) com-

ponent.

Based on this decomposition a symmetry value

symm

= 1−

∞

−∞



ℓ

(τ) − R

symm

(τ)



dτ

∞

−∞

ℓ

(τ) dτ

(9)

can be speciﬁed as the square deviation of the CCF

from its symmetric component. In case of full axis

symmetry, the symmetric value in Equation 9 reaches

”1” or 100%. The cross-correlation function of the

left and right eigenforms for the parameter settings

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

382

in Table 1 are displayed in Figure 13. The CCF’s

oscillations are caused by the under- and overshoots

of the EMG-pulses. With an increasing area value,

these parts diminish in the pulse shapes and eigen-

forms. The CCFs are dominated by their symmetric

components, which is also conﬁrmed in Table 2 by

symmetry values close to 100%.

-50 0

0-10

0-5

lags[s]

-0.01

-0.02

CCF

0.02

CCF

0.02

CCF

-0.2

CCF

0.4

0.2

CCF

Figure 13: Cross-correlation functions for the parameter

settings of Table 1.

Table 2: Symmetry values for parameter settings in Table 1.

parameter symmetry

settings value in %

E1 99.79

E2 99.74

E3 99.43

E4 99.97

E5 99.96

Figure 14 displays the eigenform’s cross-

correlation functions of another EMG-signal. The

eigenforms are based on a pulse detection whose

parameters are shown in Table 3. Due to a poorer

signal-to-noise ratio, the minimum level value is set

to 0.2. Only CCF

, the CCF for the third parameter

set (level value of 0.2, area value equal to 0.7) seems

quite symmetric. This assumption is conﬁrmed by a

symmetry value C

symm

= 0.9975 in Table 3.

This means that the eigenforms of the second

EMG-signal are more sensitive to variations of the

pulse detection’s parameters. This may be caused by

an incorrect application of the surface sensors, which

induces additional noise and deformations. There-

fore, the combination of the eigenforms’ loadings and

the symmetry value indicates the quality of the mea-

surement.

-40 40

-10

-40

-10 10

lags[s]

0.02

CCF

-0.4

-0.8

CCF

-0.1

CCF

0.4

-0.4

CCF

0.1

CCF

E10

Figure 14: Cross-correlation function of another EMG-

signal with parameter settings of Table 3.

Table 3: Symmetry values of the second EMG-signal.

parameter level area symmetry

settings value value value in %

E6 0.2 0.1 86.29

E7 0.2 0.4 97.16

E8 0.2 0.7 99.75

E9 0.3 0.1 77.23

E10 0.3 0.4 96.19

ENHANCED ANALYSIS OF UTERINE ACTIVTY USING SURFACE ELECTROMYOGRAPHY

383

4 CONCLUSIONS

The existing methods for the analysis of EMG-signals

are not precise enough for a reliable prediction of

the point of delivery. The new approach presented

in this contribution is based on the distinction be-

tween pulses (muscular contraction) and ﬂat line sec-

tions during relaxation. A time-frequncy analysis re-

veals that only the pulses contain relevant information

whereas the ﬂat line sections can be neglected. For

this reason, a semi-automatic pulse detection is de-

veloped. The physician controls the functionality of

the pulse detection by adapting two comprehensible

parameters, while the time-consuming work of pulse

extraction is done automatically. The ﬁrst parame-

ter, the level value, inﬂuences the number of extracted

pulses, whereas the second parameter, the area value,

determines the length of the pulses. Therefore, the

physician integrates his current observations as well

as his medical experiences into the pulse detection.

The use of surface electrodes leads to deforma-

tions in the individual pulse shapes. In this ap-

proach, the pulses extracted by the pulse detection

are treated as realisations of a stationary stochastic

process. In order to derive a generalized pattern, a

stochastic analysis, the Karhunen-Lo`eve-Transform

(KLT), is carried out. The KLT is based on the eigen-

value/eigenvector problem of the covariance matrix.

While an eigenvector represents a generalised pattern,

the corresponding eigenvalue speciﬁes the degree of

similarity with regard to the pulse ensemble. Eigen-

value and eigenvector yield to the eigenform. The

more dominant the ﬁrst eigenform is, the better it rep-

resents the pulses of the ensemble.

Until now, the mean frequency has been used for

the prediction of the point of delivery. Although the

pulse detection reduces the frequency deviation sig-

niﬁcantly, the mean frequency remains sensitive to

variations of the pulse detection’s parameters because

the individual pulses are randomly distorted by con-

ductivity effects. The ﬁrst eigenform of the KLT is

less susceptible to parameter variations. Particular

in case of a dominant ﬁrst eigenform, the mean fre-

quency becomes a reliable criterion.

Furthermore, a new characteristic value is devel-

oped: the symmetry value. It is derived from the

cross-correlation function of the ﬁrst eigenforms of

the left and right EMG-channel. If the quality of

the electromyogram is high, the pulse ensembles of

the left and right channel will yield to quite identical

eigenforms and a symmetry value close to 100%. To-

gether with the eigenvalues of the KLT, the symmetry

value serves as a criterion for the measurement’s reli-

ability.

In the future, the pulse detection combined with

the stochastic analysis will be applied on a sufﬁciently

large amount of electromyograms taken from various

women during the last period of pregnancy. With

these results, the reliability of this new approach as

well as the improvement with regards to the present

methods will be be quantiﬁed.

REFERENCES

Jolliffe, I. T. (2002). Principal Component Analysis.

Springer, New York, 2nd edition.

Maul, H., Maner, W. L., Olson, G., Saade, G. R., and

Garﬁeld, R. E. (2004). Non-invasive transabdominal

uterine electromyography correlates with the strength

of intrauterine pressure and is predictive of labor

and delivery. In The Journal of Maternal-Fetal and

Neonatal Medicine. Parthenon Publishing.

Mertins, A. (1999). Signal analysis: wavelets, ﬁlter banks,

time-frequency transforms and applications. Wiley,

Chichester, 2nd edition.

Oppenheim, A. and Schafer, R. W. (1999). Discrete-time

signal processing. Prentice Hall, Upper Saddle River,

NJ, 2nd edition.

Reicke, L., Kaiser, I., and Kroeger, M. (2006). Identiﬁ-

cation of the running-state of railway wheelsets. In

ISMA2006, International Conference on Noise & Vi-

bration Engineering. Katholieke Universiteit Leuven,

Department of Mechanical Engineering.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

384