ICA for Surface Electromyogram
Ganesh R. Naik
1
, Dinesh K. Kumar
1
, Sridhar P. Arjunan
1
and M. Palaniswami
2
1
School of Electrical and Computer Engineering, RMIT University,
GPO Box 2476V. Melbourne, Victoria-3001, Australia.
2
Department of Electrical and Electronic Engineering, The University of Melbourne,
Parkville, Victoria-3010, Australia.
Abstract. Surface electromyogram (SEMG) is an indicator of the underlying
muscle activity and can be useful for human control interface. One difficulty in
the use of SEMG for identifying complex movements is the mixing of muscle
activity from other muscles, referred to cross-talk. Similarity in frequency and
time domain makes the separation of muscle activity from different muscles ex-
tremely difficult. Independent Component Analysis (ICA) is a useful technique
for blind source separation. This paper reports investigations to test the effec-
tiveness of using ICA for such applications. It determines the impact of different
conditions on the reliability of the separation. The paper reports the evaluation
of issues related to the properties of the signals and number of sources. The pa-
per also tests Zibulevsky’s method of temporal plotting to identify number of
independent sources in SEMG recordings. The results demonstrate that ICA is
suitable for SEMG signals when the numbers of sources are not greater than the
number of recordings. The inability of the system to identify the correct order and
magnitude of the signals is also discussed. It is observed that even when muscle
contraction is minimal, and signal is filtered using wavelets and band pass fil-
ters, Zibulevsky’s sparse decomposition technique does not identify number of
independent sources.
1 Introduction
Blind Source Separation (BSS) by Independent Component analysis (ICA) is emerg-
ing as a new standard in signal processing and data analysis. ICA has received much
attention over the last years in the field of neural computations and bio medical sig-
nal processing due to its potential applications to the array signal processing such as,
separation of Biosignals especially SEMG. Research that isolates MUAP originating
from different muscles and motor units has been reported in 2004 [1], where success
is reported in the isolation of the different MUAP with applications for decomposing
the SEMG at low levels of muscle activation. Recently a denoising method using ICA
and high pass filter is used to suppress the interference of ECG in SEMG recorded
from trunk muscles has been reported [2]. ICA has also been proposed for unsuper-
vised cross talk removal from SEMG recordings of the muscles of the hand [3]. From
literature, ICA appears to be the emerging technology with solutions to most of the
requirements for filtering SEMG.
R. Naik G., K. Kumar D., P. Arjunan S. and Palaniswami M. (2006).
ICA for Surface Electromyogram.
In Proceedings of the 2nd International Workshop on Biosignal Processing and Classification, pages 51-60
DOI: 10.5220/0001223200510060
Copyright
c
SciTePress
Surface electromyography (SEMG) is a kind of non invasive and dynamic mea-
surement of muscular function. It is a result of the superposition of a large number of
transients that have temporal and spatial separation that is semi-random. These tran-
sients are the motor unit action potentials (MUAP). SEMG from different muscles that
need to be separated often have spectral overlaps and this makes the use of spectral fil-
tering not suitable for separating the signals effectively. Wavelets combine the temporal
and spectral properties and are suitable for separating some of the signals that may not
be separable using spectral filtering, but are limited for continuous signals with spectral
overlaps. Further, the above techniques are suitable only when there is prior information
of the signals.
ICA is a multi dimensional signal processing technique to separate signals from
different sources into distinct components. ICA algorithms have been considered to be
information theory based unsupervised learning rules. Given a set of multidimensional
observations, which are assumed to be linear mixtures of unknown independent sources
through an unknown mixing source, an ICA algorithm performs a search of the de mix-
ing matrix by which observations can be linearly translated to form Independent output
components. It is a very convenient technique for source separation as it requires very
little information of the sources or the signals to be separated, and with the availability
of easy to use software packages, is becoming very popular for numerous applications.
ICA can be employed in unsupervised situations and this makes it very attractive for
number of applications. However, the success of using commonly used ICA algorithms
for signal separation is dependent on some properties of the signals and the recordings.
These include the linearity of the mixing medium, small sensor noise, and the inde-
pendence of the underlying sources as well as the equity of the number of sources and
the number of recordings or sensors. Further, most of the ICA techniques available are
based on the assumption that there is no propagation delay. When any one of these is
not met, the output of that separation technique is questionable. The other assumption
that determines the suitability of ICA is that the number of sources need to be less than
or equal to the number of recordings.
When SEMG is recorded, most of the times the number of recording channels cor-
respond to the active muscles being measured, with no spare recording to account for
the artefact. If the artefact was to be removed using ICA, the source of the artefacts
would be another independent source, and in such a situation, the number of sources
would exceed the number of recordings. To overcome the difficulty of separation of
signals when the number of sources exceeds the number of recordings, an alternate to
the entropy based ICA is the use of blind source separation using clustering. Zibul-
vesky et al. [4] showed that during the overcomplete case (number of sources exceeds
number of recordings) audio recordings can be separated by making the data sparse. It
is often the density of the SEMG that carries the information related to the activity of
the specific muscle and making it sparse may alter the information content of the sig-
nal. In applications where the muscle is weakly active and the signal strength is small,
this may provide a solution. This paper uses Zibulvesky’s sparse decomposition tech-
nique for SEMG to tests by making the data sparse whether it is possible to separate the
independent sources.
52
2 Surface Electromyogram
Surface electromyography (SEMG) is the recording of the electrical activity of skeletal
muscle from the skin surface. It is a non-invasive recording of the muscle activity and
finds application in sports training, rehabilitation, machine and computer control, occu-
pational health and safety, and for identifying posture disorders. There is a near linear
relationship between RMS of SEMG and the finger flexion-extension - suggesting the
use of SEMG for bio-control for anthropomorphic tele-operators and Virtual Reality
entertainment [5]. There is useful information of the posture from the muscle activ-
ity of the lumbar muscles. SEMG amplitude and frequency have been investigated as
indicators of localized muscular fatigue. Amplitude and spectral information of EMG
have also been exploited to estimate force of muscle contraction and torque [6]. These
applications require automated analysis and classification of SEMG.
SEMG may be affected by various factors such as the muscle anatomy (number of
active motor units, size of the motor units, the spatial distribution of motor units); mus-
cle physiology (trained or untrained, disorder, fatigue); nerve factors (disorder, neuro-
muscular junction); contraction (level of contraction, speed of contraction, isometric/non-
isometric, force generated); artefacts (crosstalk between muscle, ECG interference),
and recording apparatus factors (recording method, noise, electrode’s properties, record-
ing sites). The anatomical/ physiological processes such as properties and dimensions
of tissues, and force and duration of contraction of the muscle are known to influence
the signal. SEMG is also influenced by onset of muscle fatigue, and contraction of other
muscles in the close vicinity. Each of the factors can be used as a criterion to categorise
the input signal.
One property of the SEMG is that the signal originating from one muscle can gen-
erally be considered to be independent of other bioelectric signals such as electrocar-
diogram (ECG), electro-oculargram (EOG), and signals from neighbouring muscles.
This opens an opportunity of the use of independent component analysis (ICA) for this
application.
2.1 ICA for SEMG Applications
Signals from different sources can get mixed during recording. Often it is required to
separate the original signals, and there is little information available of the original sig-
nals. An example is the cocktail party problem. Even if there is no (limited) information
available of the original signals or the mixing matrix, it is possible to separate the origi-
nal signals using independent component analysis (ICA) under certain conditions. ICA
is an iterative technique that estimates the statistically independent source signals from
a given set of their linear combinations. The process involves determining the mixing
matrix. The independent sources could be audio signals such as speech, voice, music,
or signals such as bioelectric signals.
A number of researchers have reported the use of ICA for separating the desired
SEMG from the artefacts and from SEMG from other muscles. While details differ, the
basic technique is that different channels of SEMG recordings are the input of ICA al-
gorithm.The fundamental principle of ICA is to determine the un-mixing matrix and use
53
that to separate the mixture into the independent components. The independent compo-
nents are computed from the linear combination of the recorded data. The success of
ICA to separate the independent components from the mixture depends on the prop-
erties of the recordings. When examining the various attempted applications of ICA,
two properties of SEMG recordings appear important; (i) number of sources exceeding
number of recordings and (ii) statistical properties. These two properties of SEMG are
examined below.
Number of Sources Exceed Number of Recordings. When SEMG is recorded, most
of the times the number of recording channels correspond to the active muscles being
measured, with no spare recording to account for the artefact. If the artefact was to be
removed using ICA, the source of the artefacts would be another independent source,
and in such a situation, the number of sources would exceed the number of recordings.
It is thus important to determine the conditions under which standard ICA could be used
to remove artefacts from biosignal recordings when the number of sources may exceed
the number of recordings. To analyse this, consider the set of recordings to be a vector
x and the pure signals (unknown) to be a vector s. Then
x = As (1)
where A is an unknown mixing matrix. The output of ICA algorithm is an estimate of
un-mixing matrix W so that
s = Wx
= WAs
(2)
It is evident that WA = I, identity matrix. If the number of recorded data is less than
the number of true independent sources (A is not a square matrix), running standard
ICA on this kind of data will never give truly independent source. The estimated inde-
pendent components will be a mixture of those true independent sources with element
of W as the scale factor. To prove the same, consider two channel recordings x of three
independent sources s and express it as:
x
1
= a
11
s
1
+ a
12
s
2
+ a
13
s
3
(3)
x
2
= a
21
s
1
+ a
22
s
2
+ a
23
s
3
(4)
Consider the estimated un-mixing matrix,
W = [w
11
w
12
;w
21
w
22
] (5)
generated using standard ICA algorithm on that data. The estimated independent com-
ponents can be written as:
es
1
= w
11
x
1
+ w
12
x
2
= w
11
(a
11
s
1
+ a
12
s
2
+ a
13
s
3
)
+ w
12
(a
21
s
1
+ a
22
s
2
+ a
23
s
3
)
(6)
54
es
2
= w
21
x1+ w
22
x2
= w
21
a
11
s
1
+ a
12
s
2
+ a
13
s
3
)
+ w
22
(a
21
s
1
+ a
22
s
2
+ a
23
s
3
)
(7)
If none of the coefficient of mixing matrix A is zero means that all three sources are
present in both mixtures x
1
and x
2
As A is a full rank matrix, then there is no column
or row dependency. Under these conditions, there is no W that will be able to isolate
one source from others. The only possible way that the estimated output would look
very similar to one of the independent sources is when its corresponding magnitude is
higher than others. Since the number of actual independent sources of SEMG signal
recorded from electrode is unknown (and is believed to be many), standard ICA will
not be suitable for applications except when the magnitude of some of the sources is
comparatively much higher.
Statistical Properties of SEMG Recordings. Signals from Gaussian sources cannot be
separated from their mixtures using ICA [7], making such signals unsuitable for ICA
applications. Mathematical manipulation demonstrates that all matrices will transform
this kind of mixtures to another Gaussian data. However, a small deviation of density
function from Gaussian may make it suitable as it will provide some possible max-
imization points on the ICA optimization landscape, making Gaussianity based cost
function suitable for iteration. If one of the sources has density far from Gaussian, ICA
will easily detect this source because it will have a higher measure of non Gaussianity
and the maxima point on the optimization landscape will be higher. If more than one
of the independent sources has non Gaussian distribution, those with higher magnitude
will have the highest maxima point in the optimization landscape. Given a few signals
with distinctive density and significant magnitude difference, the densities of their lin-
ear combinations will tend to follow the ones with higher amplitude. Since ICA uses
density estimation of a signal, the components with dominant density will be found
easier.
Signals such as SEMG have probability densities that are close to Gaussian while
artefacts such as ECG and motion artefacts have non Gaussian distributions. From the
above, it can be suggested that ICA may suitably isolate some of the above signals,
while its efficacy for separating the others maybe questionable. It is difficult to identify
the quality of separation of EMG from one muscle and the neighbouring muscles, or
that of EEG from one channel to the neighbouring recording sites, making it difficult to
confirm or negate the above.
3 Sparse ICA
Sparse representation of signals which is modeled by matrix factorisation has been re-
ceiving great deal of interest in recent years. The research community has researched
many linear transforms that make audio, video and image data sparse, such as the dis-
crete cosine transform (DCT), the Fourier transform, the wavelet transform and their
derivatives [8]. Chen et al. [9] discussed sparse representations of signals by using large
scale linear programming under given over complete basis (e.g., wavelets). Olshausen
et al. [10] represented sparse coding of images based on maximum posterior approach
55
but it was Zibulvesky et al. who noticed that in the case of sparse sources, their lin-
ear mixtures can be easily separated using very simple ”geometric” algorithms. Sparse
representations can be used in blind source separation. When the sources are sparse,
smaller coefficients are more likely and thus for a given data point t. if one of the
sources is significantly larger, the remaining ones are likely to be close to zero. Thus
the density of data in the mixture space, besides decreasing with the distance from the
origin shows a clear tendency to cluster along the directions of the basis vectors. Spar-
sity is good in ICA for two reasons. First the statistical accuracy with which the mixing
matrix A can be estimated is a function of how non-Gaussian the source distributions
are. Thus, roughly speaking the sparser the sources are the less data is needed to esti-
mate A. Secondly the quality of the source estimates given A, is also better for sparser
sources.
A signal is considered sparse when values of most of the samples of the signal do
not differ significantly from zero. These are from sources that are minimally active.
Zibulevsky et al. have demonstrated that when the signals are sparse, and the sources
of these are independent, these may be separated even when the number of sources
exceeds the number of recordings [4]. The over-complete limitation suffered by nor-
mal ICA is no longer a limiting factor for signals that are very sparse. Zibulevsky also
demonstrated that when the signals are sparse, it is possible to determine the number of
independent sources in a mixture of unknown signal numbers. One application where
the use of blind source separation for SEMG is required is when the signal strength is
very small, and the sources are minimally active, such as during maintained posture.
This leads to the argument of the use of Zibulevsky’s ICA technique to separate muscle
activity originating from muscles that are minimally active. It also provides the basis
for identifying the number of active independent sources in the mixture to validate the
use of ICA for SEMG application.
3.1 Identification of Sources using Plotting of Sparse Data
Zibulvesky et al. developed a simple probabilistic method for over determined ICA
source separation. For a more general case they used maximum aposteriori approach
which includes the situation of over complete dictionary and more sources than sen-
sors. They have also demonstrated the combination of clustering and shortest path de-
composition technique to be faster and more robust. This required the estimation of
the mixing matrix before hand by clustering and then reconstruction of the sources by
shortest path decomposition. They demonstrated the separation of up to six different
audio sounds mixed into two mixtures (recordings). During maintained posture of the
unloaded hand, muscles are minimally active and the SEMG signal strength is very
small. Hence SEMG is expected to be sparse in these conditions. The plotting of the
recording against each other would be expected to demonstrate the number of indepen-
dent sources. As the first stage, it is necessary to suitably linearly transform and filter the
signals to ensure the signals are sufficiently sparse. There are number possible methods
that are available. Most common one is spectral filtering of the data for the signal to be
sparse in the time domain. The filter properties such as frequency and order needs to be
selected according to the frequency content of the signal. Typically, while maintaining
the properties of the original signal, and desiring to make the signal sparse, filtering is
56
performed where approximately 1
σ
or 12% of the signal is removed and 85% to 90%
of the energy of the original data is kept. The signal may also be filtered in time domain
by applying a threshold function to the signal data.
4 Methodology
4.1 Experiment
The experiment was conducted where Zibulevsky’s technique was applied to SEMG of
minimally active muscles to determine the number of distinct independent sources in
the mixture, thus establishing whether this test could be used for isolating muscle activ-
ity from different muscles.The experiments were approved by the Human Experiments
Ethics Committee of the University. A male subject participated in the experiment.
The experiment used 2 channel EMG configurations as per the recommended recording
guidelines. A two channel, continuous recording BIO PAC equipment was used for this
purpose. Raw signal sampled at 2000 samples/ second was recorded. The target sites
were shaved to remove hairs and cleaned with alcohol wet swabs. Ag/AgCl electrodes
(AMBU Blue sensors from MEDICOTEST, Denmark) were mounted on appropriate
locations close to the selected muscles in the right forearm. The SEMG was recorded
from muscles of the right arm while performing simple finger posture (gesture), where
the muscles were minimally activated at approximately 5% maximum voluntary con-
traction (MVC). The aim of the experiment was to determine the effectiveness of using
Zibulevsky’s technique where the signals (SEMG recordings) in time domain are plot-
ted against each other to identify number of independent sources in the mixture that is
the muscle activity.
4.2 Analysis
The aim of this experiment was to justify the underlying theory of the use of ICA
for separation of the EMG signals. This will determine if it is appropriate to assume
that the sources of MUAPs can be considered as independent. For this purpose, the
SEMG recordings were first made sparse. The recorded signals were analysed using
MATLAB software. The aim was to make the data sparse. The Signals were initially
filtered with Butterworth filter of order four so that the energy of the signal after filtering
was maintained at 90%. Corresponding histograms were plotted and compared with the
original histograms to make sure that they maintain the gaussianity as shown in Figure
(1) and Figure (2). Scatter plotting was done by the resultant sparse signals. These were
visually observed to identify the number of sources.
5 Results and Observations
The SEMG data was made sparse by band pass filter. Figure (1) and Figure (2) shows
the histograms for both original and sparse data. Figure (3) shows the scatter plot of the
sparse recorded SEMG signals. From the scatter plot it can be visualised that there are
no distinguishable lines in the directions of the basis vectors, which shows that even
57
0 500 1000
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0 0.01 0.02 0.03 0.04
0
5
10
15
20
25
30
35
40
Fig.1. Example of one Channel EMG Recording of finger movements and the Histograms.
0 500 1000
−0.015
−0.01
−0.005
0
0.005
0.01
−0.05 0 0.05
0
20
40
60
0 500 1000
−0.01
−0.005
0
0.005
0.01
−0.05 0 0.05
0
10
20
30
40
Fig.2. Two Channel Sparse EMG Recordings and the respective Histograms.
58
when SEMG is recorded from minimal contraction and filtered, the data is not suitable
to determine the number of independent sources in SEMG recordings.
The sparseness of the SEMG recordings is observed from the histogram plot. While
the original data was modestly sparse, the signal was made more sparse after filtering,
where nearly 12% of the energy was removed (based on 1 sigma). The results demon-
strate that sparse decomposition technique is not able of identifying the independent
sources in SEMG recordings. This could suggest that either the signal was not sparse
enough even after the filtering, or the sources are not independent, or the number of
sources was very large.
−0.015 −0.01 −0.005 0 0.005 0.01
−8
−6
−4
−2
0
2
4
6
8
x 10
−3
Fig.3. Scatter plot of Sparse data using Zibulevsky’s Sparse Decomposition Technique.
6 Discussions and Conclusion
The results of the experiments demonstrate that using Zibulvesky’s sparse decompo-
sition technique, it is not possible to determine the number of independent sources in
SEMG recordings. The reason for this could be either because there are very large
numbers of independent sources, or that SEMG signal, even at extremely low levels of
contraction and after filtering, is not sparse enough. From the above, it is concluded
that Zibulvesky’s Sparse Decomposition technique cannot be used for the separation of
SEMG signals.
References
1. Hideo, Nakamura., Masaki, Yoshida., Manabu, Kotani., Kenzo, Akazawa., Toshio, Mori-
tani. : The application of independent component analysis to the multi-channel surface elec-
tromyographic signals for separation of motor unit action potential trains, Vol. 14. Journal of
Electromyography and Kinesiology, (2004) 423 - 432
59
2. Yong, Hu., Li, X. H., Xie, X. B., Pang, L. Y., Yuzhen, Cao., Luk, K. D. K. : Applying Inde-
pendent Component Analysis on ECG Cancellation Technique for the Surface Recording of
Trunk Electromyography, IEEE Engineering in Medicine and Biology 27th Annual Confer-
ence, Shanghai (2005)
3. Greco, A., Costantino, D., Morabito, F. C., Versaci, M. A. : A Morlet wavelet classification
technique for ICA filtered SEMG experimental data, Vol. 1. Neural Networks Proceedings
of the International Joint Conference, (2003) 66 - 71
4. Zibulevsky, M., Pearlmutter, B. A., Bofill, P., Kisilev, P. : Blind source separation by sparse
decomposition in a signal dictionary, In: Roberts, S. J. and Everson, R. M.(eds): Independent
Components Analysis: Principles and Practice, Cambridge University Press (2000)
5. Gupta, V., Reddy, N. P. : Surface electromyogram for the control of anthropomorphic tele-
operator fingers, Vol. 29. Student Health Technology Information, (1996) 482 - 487
6. Moritani, T., Muro, M. : Motor unit activity and EMG power spectrum during increasing
force contraction, Vol. 56. Eur. J. Appl. Physio. Occup. (1987) 260 - 265
7. Hyvarinen, A., Karhunen, J., Oja, E. :Independent Component Analysis, John Wiley, New
York (2001)
8. Mallat, S. :A wavelet tour of signal processing, Cambridge University Press (2000)
9. Chen, S., Donoho, D.L. :Atomic decomposition by basis pursuit,Vol. 20. C, SIAM J. Sci.
Comput. (1999) 33 - 61
10. Olshausen, B.A., Millman, K.J. :Learning sparse codes with a mixture-of-Gaussians
prior,vol. 12. Advances in neural information processing systems, MIT Press (2000) 841
84
60