AUTOMATED EMG-SIGNAL PATTERN CLUSTERING BASED

ON ICA DECOMPOSITION

Abbas K. Abbas

and Rasha Bassam

Biomedical Engineering Dept., RWTH Aachen University, Germany

Biomedical Engineering Dept. Aachen University of Applied Sciences, Germany

Keywords: EMG clustering, Motor Unit Action Potentials (MUAPs), Independent Component Analysis (ICA), EMG

Entropy, ROC analysis.

Abstract: Adaptive independent component analysis is interactive method for processing and classifying EMG signals

pattern through short steps of ICA algorithms. In this work the efficiency and presentation of EMG signal

decomposition and classification with adaptive ICA algorithm was investigated and presented. Single and

multiple fibers motor unit action potentials (MUAP) patterns were tested and identified. Applying a fixed

point modified ICA method, instead of much decomposition and pattern clustering algorithm localization of

the action-potential source in the vicinity of specific neuromuscular diseases was achieved. ICA has its flex-

ibility for robustly classify and identify the MUAP’s signal stochastic sources with a linear way and localiz-

ing the blind source for bioelectric potential. The utilization of adaptive ICA as an embedded clustering al-

gorithm for separating a blind signal source will assist in construction an automated EMG signal diagnosis

system with aid of new computerized real time signal processing technique. From the proposed system a

stable and robust EMG classifying system based on multiple MUAP’s intensity were developed and tested

through a standardization of clinical EMG signal acquisition and processing.

1 INTRODUCTION

Electromyography (EMG) signals classification and

processing can be used for varieties of clini-

cal/biomedical applications, spectral pattern classifi-

cation of intensity-based analysis, and modern hu-

man computer interaction. EMG signals acquired

from muscles require advanced methods for detec-

tion, decomposition, processing, and classification.

The purpose of this paper is to illustrate the various

ICA algorithms for EMG signal pattern classifica-

tion for identifying the random distribution data

clustering, and to provide efficient and stable me-

thod of understanding the signal and its physiologi-

cal nature. A comparison study was given to show

performance of various EMG signal analysis me-

thods based on different ICA. This paper provides

researchers a good understanding of adaptive ICA

EMG signal decomposition and pattern classification

(Zazula, 1999).

Adaptive signal processing have a principal rule in

defining elementary platform for EMG signals clas-

sification and processing. Biomedical signal pattern

classification can be used for varieties of clini-

cal/biomedical applications, spectral pattern classifi-

cation of intensity-based development, and modern

human computer interaction. As presented above,

the purpose of this work is to illustrate the various

methodologies and algorithms for EMG signal pat-

tern classification based on ICA random distribution

data clustering to provide efficient and robust ways

of understanding the EMG signals and its physiolog-

ical nature.

2 MATERIALS

EMG signal recorded using Delsys

acquisition

system for recording surface EMG (sEMG) and

needle EMG with sensitivity between (0.2-10 uV).

The suspected area of disorder is identified for EMG

recording, for example, the biceps brachii in the

upper arm. The EMG is then triggered to record for

a predetermined time after which the acquired signal

is differentially amplified, band pass filtered, and

digitized. The common feature for classifying intra-

muscular EMG signal is the Euclidean distance

between the MUAP waveforms. For clinical inter-

ests, the main feature of EMG signal is the number

305

K. Abbas A. and Bassam R. (2009).

AUTOMATED EMG-SIGNAL PATTERN CLUSTERING BASED ON ICA DECOMPOSITION .

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 305-310

DOI: 10.5220/0001544503050310

 SciTePress

of active motor unit (MUs), the MUAP waveforms,

and the innervations time statistics. According to

Wellig and Moschytz (Zazula, 1999), the determina-

tion of the MUAP waveform and the No. of active

MUs can be considered as a classification problem

(Wang et al., 1997, Thompson et al., 1996).

Figure 1: Surface EMG electrode and with typical MUAP

action potential unit recorded as single fiber action poten-

tial adopted from Delsys® with permission (Wang et al.,

1997).

The representation of time-triggered and no overlap-

ping MUAPs produce a shimmer. MUAP shimmer

is influenced by the time-offset of the sampled

waveforms, local fluctuation of the baseline and

background noise. Besides background noise and the

effects of signal offset, white noise influences the

classification. The classification with wavelet coef-

ficient needs the wavelet coefficient (Ff[m,n]) of

four frequency bands (m=2, 3, 4, 5) and not below

150 Hz. Classification performance depends also on

distance between the class means, therefore, the best

selection of these four frequency bands depends on

the Fourier transform of the MUAP waveforms

themselves. Boualem (Wang et al., 1997) theorized

that the time frequency representation of wavelets

decompositions (WVD) provided high-resolution

signal characterization in time-frequency space and

good noise rejection performance. This theory is

useful for EMG signal classification. For purpose of

classifying EMG patterns, EMG electrical model is

used in combination with fixed point-ICA decompo-

sition by extracted and compared two types of fea-

tures based on signal processing for the purpose of

classifying EMG patterns. The two features were the

coefficients of EMG source models and the compo-

nents of Fourier frequency spectra. The method

showed better results while describing the EMG

linear envelopes (LE) method (McKeown et al.,

2002).

Figure 2: EMG spectra for single MUAP derived with

continuous wavelet transform (CWT) (McKeown et al.,

2002).

3 EMG PROCESSING METHOD

Surface EMG (sEMG) can be estimated as a mean

absolute value (MAV) or as root mean square

(RMS) of the time varying standard deviation. In

this approach, the signal is first passed through a

noise-rejection filter, uncorrelated, demodulated, and

smoothed (Thompson et al., 1996). Integrated EMG

signal identification can be used in this technique in

order to map the bioaction potentials of correspond-

ing fibers during firing intervals of myofibers. The

area of MUAP waveform, either the entire wave-

form or the rectified version of the waveform. The

characterization of EMG pattern can be based on

the following criteria (Wang et al., 1997, McKeown,

et al., 2002),

(a) Spike Area. Analogous to compute area under the

spikes and calculated only over the spike duration.

(b) Phases. No. of baseline crossings that exceeds a

certain voltage threshold, e.g., 25 μV.

rated by a specified threshold voltage, (e.g. ±25 μV),

(d) Willison Amplitude (WAMP) the No. of counts

for a change of signal in the EMG amplitude above a

predefined threshold (McKeown, 2000, Jung et al.,

2001).

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

306

Figure 3: Surface EMG, decomposition and clustering

(Garcia et al., 2002) consisting of signal processing, signal

decomposition, template matching, and post processing

stages (Thompson et al., 1996).

4 INDEPENDENT COMPONENT

ANALYSIS (ICA)

ICA defines a generative model for the observed

multivariate data, which is typically given as a large

database of samples. In the model, the data variables

are assumed to be linear or nonlinear mixtures of

some unknown latent variables or activation sources,

and the mixing system is also unknown. The latent

variables are assumed nongaussian and mutually

independent and they are called the independent

components of the observed data. These independent

components, also called sources or factors, can be

found by ICA (McKeown, 2000, Jung et al., 2001).

In the basis of spatiotemporal characteristics of

EMG signal, the later behavior can be approximated

as band limited white noise (BLWN) with spatial

frequency of recorded signal fast ICA algorithms for

a hierarchical neural network that extracts the source

signals from their mixtures in a sequential fashion.

In the hierarchical neural network, the output of the

extraction processing unit is described as y

= w

, where w

= [w

1, w

2,…, w

. Contrary to the

cascade neural network, the input vector for each

processing unit of the hierarchical neural network is

the same x

=Q*x vector from the prewhitened EMG

electrode signals. Let us now consider the cost func-

tion and the standard kurtosis for a zero mean signal

y (EMG) (Cichocki and Amari, 2003).

{

}

{

}

[

]

,3))w((),(w J

1411

yEyEyky −==

(1)

Where y

= w

is the output of EMG signal

processing unit.

In order to ﬁnd the optimal value of vector w, we

apply the following iteration rule:

))((

141

⎟

⎠

⎞

⎜

⎝

⎛

∇

lwk

(2)

Where

./)()(

114141

wwkwk

∂

∇

equivalently ap-

plying the following training formula:

));(()1(

1411

lwklw

∇=+

(3)

)1(

(4)

Figure 4 illustrate anatomy of adaptive fixed-point

ICA algorithm used for EMG signal processing and

classification, as illustrated the main module of

algorithm consist of weighting coefficient of inde-

pendent signal decomposition in which the output

pattern indicated the assigned source of EMG poten-

tials (Cichocki and Amari, 2003).

Figure 4: Adaptive fixed Point-ICA algorithm for EMG

signal clustering and processing corresponds of stable

function.

The vector w (l+1) j has unit length in each iteration

step through enforcing minimizing cost function (J).

The gradient of the cost function of EMG ICA Algo-

rithm can be evaluated as:

{} {}{}

)1(

xyEyExyE

−=

∂

=∇

(5)

Thus the ﬁxed-point algorithm in its standard form

can be written as:

AUTOMATED EMG-SIGNAL PATTERN CLUSTERING BASED ON ICA DECOMPOSITION

307

)1(

)( ),(3)1(

11111

=−=+

xlwylwxylw

(6)

To find a stable solution for EMG classification

criteria, an embedded in ICA –learning kernel were

placed in the path of general clinical module. Proto-

typing of an adaptive EMG pattern clustering is one

of the task were achieved, based on a biopotential

acquisition module.

Figure 5 that illustrated the overall EMG acquisition

procedure from site recording to the final output

clustered data in the final clinical decision module.

The capability to improve clinical data transfer pro-

tocols, the physiologist and physicians can use a

wireless transmission module form the main clinical

workstation via a standard protocol of IEEE 802.14a

wireless communication system to visualize and

interact with acquired EMG physiological data. The

automated clinical diagnosis system can be also

integrated as graphical user interface in PDA plat-

form or other mobile communication devices (Bell

and Sejnowski, 1995, Micera et al., 2001).

Figure 5: Block diagram of EMG-ICA decomposition and

pattern clustering with adaptive fixed point-MUSE ICA

algorithm.

5 ADAPTIVE ICA ALGORITHM

As modified ICA, adaptive algorithm was imple-

mented in the path of the main independent compo-

nent estimation method, were the training data and

the main entropy of ICA minimized in order to ob-

tain residual differences in EMG intensity (Cichocki

and Amari, 2003, Merletti and Parker, 2004).

6 ROC ANALYSIS AND EMG

SENSOR ENTROPY

For testing the overall performance of the EMG data

classifier performance, a statistical test for the ro-

bustness and optimality of the data classifier was

applied, to illustrate the results of clustering and then

need to characterize EMG sensor itself for calibra-

tion and buffering purpose. Receiver operating cha-

racteristics (ROC) is analysis of sensor inference

were it is calculated through repeatable EMG re-

cording. Therefore the identification of sensitivity

vs. specificity should consider in this analysis. The

robustness of ICA algorithm was tested in contrast

to the amplifier gain of interface (Bell and Sejnows-

ki, 1995). Surface EMG electrodes which are used in

clinical experiment are disposable and composed of

lined conductive area used for increasing measure-

ment stability and reduction of parasitic noise asso-

ciated with physiological measurement session.

Therefore sensor characteristics should be taken into

account for further classification steps to improve

accuracy of overall clinical assessment. The defini-

tion of sensitivity and selectivity with ROC analysis,

have the following criteria for recursive data cluster-

ing and pattern classification of biomedical and

clinical data (Cichocki and Amari, 2003).

Table1: Stability and performance index for ICA-methods.

EMG signal ICA- algorithms performance comparison

ICA methods EMG-intensity

classified pattern

Entropy

(ψ)

Correlation

index (ξ)

ICA-MUSE

0.1342

2.23

ICA-JADE E

0.1275 2.41

ICA-SOBA

0.1448

2.83

ICA-Adaptive E

,,M

0.2030 2.51

ICA-REA E

0.1942 30

ICE-MJADE E

0.1820 29

ICA-Standard E

0.1872 32

ICA-NLE E

0.1922 36

ICA-Wavelet E

0.2051 35

Entropy index for ICA-decomposition not for EMG classifier.

Correlation index for EMG derived spectral information.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

308

Figure 6: Entropy index of EMG-ICA decomposition algo-

rithm.

As noticed from the performance index of adaptive

ICA observer a good EMG M

, M

classes corres-

ponding to good potential preweighting and separa-

tion, as well as the low indexing criteria for the E

classes, indicate a non-stabilized bioaction potential

pattern that exist in clustered EMG signals (Micera

et al., 2001, Karhunen and Oja, 1997).

Figure 7: 3D Performance index of decomposed EMG signal

based on Adaptive ICA algorithm indicate main components

of E

, E

, E3

, M2, M3 spatial scheme of EMG.

In this work a robust technique for extracting and

classifying MUAPs for EMG signal was developed.

This technique is based on single-channel and short

periods real-time recordings from normal subjects

and artificially generated recordings. This EMG

signal decomposition technique has several distinc-

tive characteristics compared with the former de-

composition methods:

(A) Band pass filtering the EMG signal through

wavelet filter and utilizes threshold estimation calcu-

lated in wavelet transform for noise reduction in

EMG signals, and to detect MUAPs before ampli-

tude single threshold filtering;

(B) Removing the power interference component

from EMG recordings by combining both, (ICA)

and wavelet filtering method as adaptive ICA.

based on the entropy (ψ

EMG

) value normalized with

the sum of median values for EMG signal;

(D) Lastly uses ICA method to subtract all accurate-

ly classified MUAP spikes from original EMG sig-

nals. The technique of our EMG signal decomposi-

tion is fast and robust, which has been evaluated

through synthetic EMG signals and real one, where

the synthetic (artificial) EMG data generated using

an approximated band-limited white noise model

with varying seeding noise developed by Farina

(Merletti and Parker, 2004, Farina and Merletti,

2001).

Table 2: EMG Classification performance index.

EMG-data

subject

p-value Uncorrected

pattern

Corrected

pattern

Percentage

1 0.0023 4 32 88.90%

2 0.0027 3 34 91.90%

3 0.0031 4 34 89.4%

4 0.0032 2 33 94.28%

5 0.00362 3 30 90.09%

6 0.00317 2 29 93.54%

7 0.0047 3 32 91.42%

8 0.00528 4 36 90.0%

9 0.00571 3 35 92.10%

7 RESULTS AND DISCUSSION

Adaptive ICA-decomposition for EMG signal illu-

strates optimality in clustering (approx. 0.0187) for

spontaneous EMG vector classification, but with a

deviation in linearity of MUAP classes due to differ-

ent standard deviation (SD), of each signal, which

can be compensated by increasing the correlation

index or selecting the same order number of finite

impulse response (FIR) filtering module. Testing

additional ICA algorithms to evaluate clustering

efficiency presents with robust hyperplane classifi-

cation based on other criteria such as EMG signal

turns, spike area, integrated area, phases performed

in an adaptive ICA algorithm as concise effective

methods to increase stability of overall clustering

schemes. Euclidean distance have been computed in

background algorithm as state vector mapping

(SVM) matrices for each EMG signal within indi-

vidual channel in contrary this will overload compu-

tation time for reiterative clustering. As we can ob-

served form the Fig.7 where the presentation of 3D

performance matrix seen, there are a convergent

efficiency in ICA clustering method were the posi-

tive value of EMG signal intensity clearly visible

AUTOMATED EMG-SIGNAL PATTERN CLUSTERING BASED ON ICA DECOMPOSITION

309

instead of relatively negative value, which mean an

inhibited EMG intensity due to the transition effect

of depolarization wave through the muscle fibers, in

which this obviously considered as a good marked

triggering for online EMG acquisition and diagnosis.

Figure 8: Histogram of ICA decomposed EMG signals

based on fixed-point adapted ICA algorithm.

8 CONCLUSIONS

Pattern clustering and decomposition based on adap-

tive ICA will improve diagnostic performance of

different neuromuscular pathology patterns. EMG

consisted of activity for six different principal pat-

terns which are identified and extracted. The clus-

tered EMG activities notated as (E1,E2,

E3,M1,M2,M3) , E-related to eccentric myoelectric

activity which dominated in near-surface electrode

and M pattern which related to myofibers compart-

ment, each of which have been classified for a 9

subject group. Source localization and identification

are robustly computer through ICA template algo-

rithm, although some pattern in EMG signal the

algorithm were not detected or classified, this due to

the some leakage in optimization cycle of EMG

signals inside ICA algorithm. The overall behavior

of adaptive ICA classifier predicted a considerable

percentage of poor to moderated classified intensity

modulated EMG signals and this due to invalid clas-

sified or decomposed respective EMG coefficient,

which indeed needs improvement for advanced

research and work on EMG pattern classification

optimization. Future perspective of EMG pattern

clustering was introduced, which may use for devel-

oping a cutting edge electrophysiology module. ICA

technique will assist in developing a robust portable

clinical system, by which acquisition, processing,

and diagnosis for several myopathic and neuropathic

EMG pattern could be achieved.

ACKNOWLEDGEMENTS

We acknowledge Aachen University of Applied

Sciences, RWTH-Aachen University and DAAD

(Deutsche Akademische Ausländische Dienst) for

providing accessibility with financial and scientific

support to put this work on the track of success.

REFERENCES

Zazula, D. (1999). Higher-order statistics used for decomposi-

tion of sEMGs. Proceedings of the 12th IEEE Symposium

on Computer Based Medical System, 72–77.

Wang, R., C. Huang, and B. Li (1997). A neural network-

based surface electromyography motion pattern classifier

for the control of prostheses. Proceedings of the 19

An-

nual International Conference of the IEEE-EMBS 1277.

Thompson, B., P. Picton, and N. Jones (1996). A comparison

of neural network and traditional signal processing tech-

niques in the classification of EMG signals. IEEE Collo-

quium on Artificial Intelligence Methods for Biomedical

Data Processing , pp 312-319.

McKeown, M. J., Torpey, D.C., Gehm W. C.: Non-Invasive

Monitoring of Functionally Distinct Muscle Activations

during Swallowing. Clinical Neurophysiology (2002).

109, 112

McKeown, M. J.: Cortical activation related to arm move-

ment combinations. Muscle Nerve. 9:19-25 (2000). 110,

112

Jung T.P., Makeig S., McKeown M. J., Bell A. J., Lee T.W.,

Sejnowski T. J.: Imaging brain dynamics using indepen-

dent component analysis. Proc. IEEE. 89(7): 1107-22,

(2001). 112

Andrzej Cichocki,

Shun-ichi AMARI, “ Adaptive Blind

Signal and Image Processing: Learning Algorithms and

Applications”, Wiley Press ,2003.

Bell A. J., Sejnowski T. J.: An information-maximization

approach to blind separation and blind deconvolution,

Neural Computation, 7:1129-1159, (1995). 112

Micera S.,Vannozzi G., Sabatini A.M., Dario P.: Improving

Detection of Muscle Activation intervals, IEEE Engineer-

ing in Medicine and Biology, vol. 20 n.6:38-46 (2001).

112, 113

Roberto Merletti , Philip Parker ," Electromyography Physi-

ology, Engineering, and Noninvasive Applications",

IEEE Press Engineering in Medicine and Biology Socie-

ty, 2004, IEEE press, pp 234-256

Farina, D., and R. Merletti, “A novel approach for precise

simulation of the EMG signal detected by surface elec-

trodes,” IEEE Trans BME 48, 637–645 (2001).

Karhunen J., Oja E.: A Class of Neural Networks for Inde-

pendent Component Analysis, IEEE Transactions on

Neural Network, vol. 8 n. 3:486-504, (1997). 112.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

310