2006 by the author(s). This paper is Open Access and is published in Biological Procedures Online under license from the auth or(s). Copying, [609750]

© 2006 by the author(s). This paper is Open Access and is published in Biological Procedures Online under license from the auth or(s). Copying,
printing, redistribution and storage permitted. Journal © 1997- 2006 Biological Procedures Onli ne – www.biologicalprocedures.co m Biol. Proced. Online 2006; 8(1): 11-35.
doi:10.1251/bpo115 March 23, 2006
Techniques of EMG signal analysis: dete ction, processing, classification and
applications

M. B. I. Reaz,1* M. S. Hussain1 and F. Mohd-Yasin1

1Faculty of Engineering, Multimedia Univer sity, 63100 Cyberjaya, Selangor, Malaysia.

*Corresponding Author: M.B.I. Raez, Faculty of Engineering, Mu ltimedia University, 63100 Cyberjaya, Selangor, Malaysia. Email:
[anonimizat]
Submitted: October 4, 2005; Revised: Janu ary 9, 2006; Accepted: January 18, 2006.
Indexing terms: Electromyography; Fourier Analysis; Muscles; Nervous System.

ABSTRACT

Electromyography (EMG) signals can be used for clinical /biomedical applications, Evolvable Hardware Chip (EHW)
development, and modern human computer interaction. EMG signals acquired from muscles require advanced
methods for detection, decomposition, processing, and classi fication. The purpose of this paper is to illustrate the
various methodologies and algorithms for EMG signal analysis to provide efficient and effective ways of
understanding the signal and its nature. We further poin t up some of the hardware implementations using EMG
focusing on applications related to prosthetic hand cont rol, grasp recognition, and human computer interaction. A
comparison study is also given to show performance of va rious EMG signal analysis methods. This paper provides
researchers a good understanding of EMG signal and its anal ysis procedures. This knowledge will help them develop
more powerful, flexible, and efficient applications.

INTRODUCTION

Biomedical signal means a collective electrical signal
acquired from any organ that represents a physical
variable of interest. This signal is normally a function of time and is describable in terms of its amplitude,
frequency and phase. The EMG signal is a biomedical
signal that measures electrical currents generated in
muscles during its cont raction representing
neuromuscular activities. Th e nervous system always
controls the muscle activity (contraction/relaxation). Hence, the EMG signal is a complicated signal, which is
controlled by the nervous system and is dependent on
the anatomical and physiological properties of muscles.
EMG signal acquires nois e while traveling through
different tissues. Moreover, the EMG detector, particularly if it is at the surface of the skin, collects signals from different motor units at a time which may
generate interaction of diffe rent signals. Detection of
EMG signals with powerful and advance methodologies

is becoming a very important requirement in biomedical engineering. The main reason for the interest in EMG
signal analysis is in clinical diagnosis and biomedical applications. The field of management and rehabilitation
of motor disability is identified as one of the important
application areas. The shapes and firing rates of Motor
Unit Action Potentials (MUAPs) in EMG signals provide
an important source of information for the diagnosis of
neuromuscular disorders. Once appropriate algorithms and methods for EMG signal analysis are readily
available, the nature and characteristics of the signal can
be properly understood and hardware implementations
can be made for various EMG signal related applications.
So far, research and extensiv e efforts have been made in
the area, developing better algorithms, upgrading
existing methodologies, improving detection techniques
to reduce noise, and to acquire accurate EMG signals.
Few hardware implementations have been done for
prosthetic hand control, grasp recognition, and human-
machine interaction. It is quite important to carry out an

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 12
investigation to classify the actual problems of EMG
signals analysis and justify the accepted measures.
The technology of EMG recording is relatively new.
There are still limitations in detection and
characterization of existing nonlinearities in the surface
electromyography (sEMG, a special technique for studying muscle signals) signal, estimation of the phase,
acquiring exact information due to derivation from
normality (1, 2) Traditional system reconstruction algorithms have various limitations and considerable
computational complexity and many show high variance
(1). Recent advances in techno logies of signal processing
and mathematical models have made it practical to
develop advanced EMG detection and analysis
techniques. Various mathematical techniques and Artificial Intelligence (AI) have received extensive
attraction. Mathematical models include wavelet
transform, time-frequency approaches, Fourier transform, Wigner-Ville Distribution (WVD), statistical
measures, and higher-order statistics. AI approaches
towards signal recognition include Artificial Neural
Networks (ANN), dynamic recurrent neural networks
(DRNN), and fuzzy logic system. Genetic Algorithm
(GA) has also been applied in evolvable hardware chip
for the mapping of EMG inputs to desired hand actions.

Wavelet transform is well suited to non-stationary
signals like EMG. Time-frequency approach using WVD in hardware could allow for a real-time instrument that
can be used for specific motor unit training in
biofeedback situations. High er-order statistical (HOS)
methods may be used for an alyzing the EMG signal due
to the unique properties of HOS applied to random time
series. The bispectrum or third-order spectrum has the advantage of suppressing Gaussian noise.

This paper firstly gives a brief explanation about EMG signal and a short historical background of EMG signal
analysis. This is followed by highlighting the up-to-date
detection, decomposition, processing, and classification methods of EMG signal along with a comparison study.
Finally, some hardware implementations and
applications of EMG have been discussed. MATERIALS AND METHODS

EMG: anatomical and physiological background

EMG stands for electromyography. It is the study of
muscle electrical signals. EMG is sometimes referred to
as myoelectric activity. Muscle tissue conducts electrical
potentials similar to the way nerves do and the name
given to these electrical signals is the muscle action potential. Surface EMG is a method of recording the
information present in these muscle action potentials.
When detecting and recording the EMG signal, there are
two main issues of concern that influence the fidelity of
the signal. The first is the si gnal-to-noise ratio. That is,
the ratio of the energy in the EMG signals to the energy
in the noise signal. In general, noise is defined as
electrical signals that are not part of the desired EMG
signal. The other issue is th e distortion of the signal,
meaning that the relative contribution of any frequency component in the EMG signal should not be altered. Two
types of electrodes have been used to acquire muscle
signal: invasive electrode and non-invasive electrode. When EMG is acquired from electrodes mounted directly
on the skin, the signal is a composite of all the muscle
fiber action potentials occurring in the muscles
underlying the skin. These action potentials occur at
random intervals. So at any one moment, the EMG signal may be either positive or ne gative voltage. Individual
muscle fiber action potentials are sometimes acquired
using wire or needle electrodes placed directly in the
muscle. The combination of the muscle fiber action
potentials from all the muscle fibers of a single motor unit is the motor unit action potential (MUAP) which can
be detected by a skin surface electrode (non-invasive)
located near this field, or by a needle electrode (invasive)
inserted in the muscle (3). Equation 1 shows a simple
model of the EMG signal:

∑−
=+− =1
0)( ) ()( )(N
rnwrnerh nx (1)

where x(n) , modeled EMG signal, e(n), point processed,
represents the firing impulse, h(r), represents the MUAP,
w(n) , zero mean addictive white Gaussian noise and N is
the number of motor unit firings.

The signal is picked up at the electrode and amplified.
Typically, a differential amplifier is used as a first stage amplifier. Additional amplification stages may follow. Before being displayed or stored, the signal can be

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 13
processed to eliminate low-frequency or high-frequency
noise, or other possible artifacts. Frequently, the user is interested in the amplitude of the signal. Consequently, the signal is frequently rectified and averaged in some format to indicate EMG amplitude. The nervous system is both the controlling and
communications system of the body. This system
consists of a large number of excitable connected cells
called neurons that communicate with different parts of
the body by means of electrical signals, which are rapid and specific. The nervous system consists of three main parts: the brain, the spinal cord and the peripheral nerves. The neurons are the basic structural unit of the
nervous system and vary considerably in size and shape.
Neurons are highly specialized cells that conduct messages in the form of nerve impulses from one part of
the body to another. A muscle is composed of bundles of specialized cells
capable of contraction and relaxation. The primary
function of these specialized cells is to generate forces,
movements and the ability to communicate such as
speech or writing or other mo des of expression. Muscle
tissue has extensibility and elasticity. It has the ability to
receive and respond to stimuli and can be shortened or
contracted. Muscle tissue has four key functions:
producing motion, moving substance within the body, providing stabilization, and generating heat. Three types
of muscle tissue can be id entified on the basis of
structure, contractile properties, and control
mechanisms: (i) skeletal muscle, (ii) smooth muscle, and
(iii) cardiac muscle. The EMG is applied to the study of
skeletal muscle. The skeletal muscle tissue is attached to
the bone and its contraction is responsible for supporting
and moving the skeleton. The contraction of skeletal
muscle is initiated by impulses in the neurons to the
muscle and is usually under voluntary control. Skeletal
muscle fibers are well-supplied with neurons for its
contraction. This particular type of neuron is called a
“motor neuron” and it approaches close to muscle tissue,
but is not actually connected to it. One motor neuron
usually supplies stimulation to many muscle fibers.

The human body as a whole is electrically neutral; it has
the same number of positive and negative charges. But in the resting state, the nerve cell membrane is polarized due to differences in the concentrations and ionic composition across the plasma membrane. A potential
difference exists between the intra-cellular and extra-cellular fluids of the cell. In response to a stimulus from the neuron, a muscle fiber depolarizes as the signal propagates along its surface and the fiber twitches. This depolarization, accompanied by a movement of ions, generates an electric field near each muscle fiber. An
EMG signal is the train of Motor Unit Action Potential
(MUAP) showing the muscle response to neural stimulation. The EMG signal appears random in nature and is generally modeled as a filtered impulse process where the MUAP is the filter and the impulse process stands for the neuron pulses, often modeled as a Poisson process (3). Figure 1 shows the process of acquiring EMG
signal and the decomposition to achieve the MUAPs.

Fig. 1: EMG signal and decomposition of MUAPs.

EMG: the history

The development of EMG started with Francesco Redi’s
documentation in 1666. The document informs that highly specialized muscle of the electric ray fish generates electricity (3). By 1773, Walsh had been able to
demonstrate that Eel fish’s muscle tissue could generate
a spark of electricity. In 1792, a publication entitled “De Viribus Electricitatis in Motu Musculari Commentarius” appeared, written by A. Galvani, where the author showed that electricity could initiate muscle contractions (4). Six decades later, in 1849, Dubios-Raymond discovered that it was also possible to record electrical
activity during a voluntary muscle contraction. The first
recording of this activity was made by Marey in 1890, who also introduced the term electromyography (5). In
1922, Gasser and Erlanger used an oscilloscope to show

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 14
the electrical signals from muscles. Because of the
stochastic nature of the myoelectric signal, only rough information could be obtained from its observation. The capability of detecting el ectromyographic signals
improved steadily from the 1930s through the 1950s and researchers began to use improved electrodes more widely for the study of muscles (1). Clinical use of
surface EMG for the treatment of more specific disorders
began in the 1960s. Hardyck and his researchers were the first (1966) practitioners to use sEMG (5). In the early 1980s, Cram and Steger introduced a clinical method for scanning a variety of muscles using an EMG sensing device (5).
It is not until the middle of the 1980s that integration
techniques in electrodes ha d sufficiently advanced to
allow batch production of the required small and
lightweight instrumentation an d amplifiers. At present a
number of suitable amplifiers are commercially available. In the early 1980s, cables became available which produce artifacts in th e desired microvolt range.
During the past 15 years, research has resulted in a better
understanding of the properties of surface EMG recording. In recent years, surface electromyography is increasingly used for recording from superficial muscles in clinical protocols, where intramuscular electrodes are used for deep muscle only (2, 4).
There are many applications for the use of EMG. EMG is
used clinically for the diagnosis of neurological and neuromuscular problems. It is used diagnostically by gait laboratories and by clinicians trained in the use of biofeedback or ergonomic assessment. EMG is also used in many types of research laboratories, including those involved in biomechanics, motor control, neuromuscular
physiology, movement disorders, postural control, and
physical therapy.

Electrical noise and fact ors affecting EMG signal

The amplitude range of EMG signal is 0-10 mV (+5 to -5) prior to amplification. EMG signals acquire noise while traveling through different tissue. It is important to understand the characteristics of the electrical noise. Electrical noise, which will affect EMG signals, can be categorized into the following types:
1. Inherent noise in electronics equipment : All electronics
equipment generate noise. This noise cannot be eliminated; using high quality electronic components
can only reduce it.
2. Ambient noise: Electromagnetic radiation is the source
of this kind of noise. The surfaces of our bodies are constantly inundated with electric-magnetic radiation and it is virtually impossible to avoid exposure to it on the surface of earth. The ambient
noise may have amplitude that is one to three orders
of magnitude greater than the EMG signal.
3. Motion artifact: When motion artifact is introduced to
the system, the information is skewed. Motion artifact causes irregularities in the data. There are two main sources for motion artifact: 1) electrode interface and 2) electrode cable. Motion artifact can
be reduced by proper design of the electronics
circuitry and set-up.
4. Inherent instability of signal : The amplitude of EMG is
random in nature. EMG signal is affected by the firing rate of the motor units, which, in most conditions, fire in the frequency region of 0 to 20 Hz. This kind of noise is considered as unwanted and the
removal of the noise is important.

The factors that mainly affect the EMG signal can also be
classified. This kind of classi fication is set so that EMG
signal analysis algorithms can be optimized and equipments can be designed in a consistent manner. Factors affecting EMG signal falls into three basic categories: 1. Causative Factors: This is the direct affec t on signals.
Causative factors can be divided into two classes:
i. Extrinsic : This is due to electrode structure and
placement. Factors like area of the detection surface, shape of electrode, distance between electrode detection surface, location of electrode with respect to the motor points in the muscle, location of the muscle electrode on the muscle
surface with respect to the lateral edge of the
muscle, orientation of the detection surfaces with respect to the muscle fibers mainly have an effect on EMG signal.
ii. Intrinsic : Physiological, anatomical, biochemical
factors take place due to number of active motor units, fiber type composition, blood flow, fiber
diameter, depth and location of active fibers and
amount of tissue between surface of the muscle and the electrode.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 15
2. Intermediate Factors: Intermediate factors are physical
and physiological phenomena influenced by one or more causative factors. Reasons behind this can be the band-pass filtering aspects of the electrode alone with its detection volume, superposition of action potentials in the detected EMG signal, conduction velocity of the action potential that propagate along
the muscle fiber membrane. Even crosstalk from
nearby muscle can cause Intermediate Factors.
3. Deterministic Factors : These are influenced by
Intermediate Factors. The number of active motor units, motor firing rate, and mechanical interaction b e t w e e n m u s c l e f i b e r s h a v e a d i r e c t b e a r i n g o n t h e information in the EMG signal and the recorded
force. Amplitude, duration, and shape of the motor
unit action potential can also be responsible.
The maximization of the qual ity of EMG signal can be
done by the following ways:

1. The signal-to-noise ratio should contain the highest
amount of information from EMG signal as possible
and minimum amount of noise contamination.
2. The distortion of EMG signal must be as minimal as
possible with no unnecessary filtering and distortion
of signal peaks and notch filters are not recommended.

During the EMG signal proce ssing, only positive values
are analyzed. When half-wave rectification is performed,
all negative data is discarded and positive data is kept. The absolute value of each data point is used during full-
wave rectification. Usually for rectification, full-wave rectification is preferred.

EMG signal detection

Precise detection of discrete events in the sEMG (like the
phase change in the activity pattern associated with the initiation of the rapid motor response) is an important issue in the analysis of the motor system. Several
methods have been proposed for detecting the on and off
timing of the muscle.

The most common method for resolving motor-related
events from EMG signals consists of visual inspection by
trained observers. The “single-threshold method,” which
compares the EMG signal with a fixed threshold, is the most intuitive and common computer-based method of time-locating the onset of muscle contraction activity (6).
This technique is based on the comparison of the rectified raw signals and an amplitude threshold whose value depends on the mean power of the background noise (7). The method can be useful in overcoming some
of the problems related to visual inspection. However, this kind of approach is generally not satisfactory, since
measured results depend strongly on the choice of
threshold. This kind of method often rely on criteria that
are too heuristic and does not allow the user to set independently the detection and false alarm probabilities (8). In “single-threshold method,” the relationship between the probability of detection P
dk and the
probability Pγ that a noise sample is above the threshold
γ is given by equation 2.





+=
10
101) ln(expSNR dkPPγ (2)

In 1984, Winter (9) observed that this approach is generally unsatisfactory, since it strongly depends on the
choice of the threshold. To overcome the “single-
threshold” problems, Bornato et al . (8) introduced
“double-threshold detection” method in 1998. Double-threshold detectors are superior to single-threshold because they yield higher detection probability. Double-threshold detectors allow the user to adopt the link
between false alarm and detection probability with a
higher degree of freedom than single-threshold. The user can tune the detector according to different optimal criteria, thus, adapting its performances to the characteristics of each specific signal and application (8). The sEMG signal recorded during voluntary dynamic
contractions may be considered as a zero-mean Gaussian
process s(t)
∈N(0,σs) modulated by the muscle activity
and corrupted by an independent zero-mean Gaussian additive noise n(t)
∈N(0,σn). If the probability of detection
is Pd then the double-threshold method is given by
equation 3.
()km
dkk
dkm
rkd P PkmP
o−
=− ∑


= 1 (3)

The behavior of the double-threshold detector is fixed by
the parameters: the threshold ro, and the length of the

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 16
observation window, m. Their values are selected to
minimize the value of the false-alarm probability and maximize P
d for each specific signal-to-noise ratio (SNR)
(8). In 2004, Lanyi and Adler (10) found that the double-threshold method proposed by Bornato is complex and computationally expensive, requiring a whitening of the signal. It is also not very sensitive. Lanyi and Andy
proposed a new algorithm based on the double-
threshold method that is more sensitive, stable, and efficient with decreased computation cost. For specific applications, besides the accuracy in the detection, the speed of the algorithm can be an important consideration. Algorithms with high computation time are unsuitable for online detection. One specific
drawback to the method of Bornato et al . (8) is the detection probability to be maximum when P
fa is fixed,
the second threshold has to be chosen as equal to “1.” The second threshold is fixed during detection, which implies that the double-threshold detector actually becomes single-threshold detector. This method does not require the signal-whitening step, which is needed previously. The whitening process takes a lot of
computation time. Moreover, the whitening process
reduces probability of the signal. This feature will cause the detection to miss a part of activation interval. The methods proposed by Lanyi and Adler (10) provides a
fast and more reliable muscle on-off detection. Table 1 shows the comparison of the different detection methods at a glance based on research works by Merlo and Farina
(11) in 2003.

Table 1: Comparison of 3 main EMG detection methods.
SNR(db)
2 4 6 8
Method Bias Std Bias Std Bias Std Bias Std Remark
Improved method (11) -39 26 -22 25 -12 22 -3 17 Best
Double threshold (8) 41 68 21 69 12 47 0 53 Good
Single threshold (11) 55 154 67 147 62 135 72 139 Worse

EMG signal decomposition

EMG signals are the superposition of activities of
multiple motor units. It is necessary to decompose the EMG signal to reveal the mechanisms pertaining to muscle and nerve control. Various techniques have been devised with regards to EMG decomposition. Decomposition of EMG signal has been done by wavelet
spectrum matching and principle component analysis of
wavelet coefficients. According to Jianjung et al . (12),
more than one single motor unit (SMU) potential will be registered at same time overlapping with each other,
especially during a strong mu scle contraction. In 1997,
they developed a technique using wavelet transform to classify SMU potentials and to decompose EMG signals
into their constituent SMU potentials. The distinction of
this technique is that it measures waveform similarity of SMU potentials from wavelet domain, which is very
advantageous. This technique was based on
spectrummatching in wavelet domain. Spectrum matching technique is someti mes considered to be more
effective than waveform matching techniques, especially
when the interference is induced by low frequency
baseline drift or by high frequency noise. The technique developed for multi-unit EMG signal decomposition consists of four separate pr ocedures: signal de-noising
procedure, spike detection procedure, spike classification
procedure, and spike separation procedure. According to Daniel et al . (13), only wavelet coefficients of lower
frequency bands are more important in the differentiation of action potential (AP) characterization
than higher bands. This concept is a subjective one which
was designed empirically. Experimental results of Rie Yamada et al. (14) in 2003 showed that high frequency
information, which were not considered, are also
important in the classification of MUAP. To overcome
the subjective criterion for feature selection, they proposed another method using principle components analysis (PAC) for wavelet coefficients. The decomposition algorithm consists of four processing
stages: segmentation, wavelet transform, PCA, and
clustering. The advantage of this method is that it does not require manual selection of coefficients, and takes all frequency information in account.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 17
EMG signal decomposition using non-linear least mean
square (LMS) optimization of higher-order cumulants
h a s b e e n p r o p o s e d b y E r i c a n d D a m j a n ( 1 5 ) i n 2 0 0 2 . Their decomposition is based on the third-order cumulants whose values en ter as coefficients of
nonlinear system of equations. The system is solved by
nonlinear LMS optimization. For this technique a multiple-input multiple-output model was used as it can describe several MUAP impositions of EMG signal.

EMG signal processing

Raw EMG offers us valuable information in a
particularly useless form. This information is useful only
if it can be quantified. Various signal-processing methods are applied on raw EMG to achieve the accurate and actual EMG signal. This section gives a review on EMG signal processing us ing the various methods.

Wavelet analysis

Both the time and frequency domain approaches have
been attempted in the past. The wavelet transform (WT) is an efficient mathematical tool for local analysis of non-stationary and fast transient signals. One of the main
properties of WT is that it can be implemented by means
of a discrete time filter bank. The Fourier transforms of the wavelets are referred as WT filters. The WT represents a very suitable method for the classification of EMG signals.

Guglielminotti and Merletti (16) theorized that if the
wavelet analysis is chosen so as to match the shape of the
MUAP, the resulting WT yields the best possible energy
localization in the time-scale plane (16). In 1997, Laterza and Olmo (17) found out that WT is an alternative to other time frequency representations with the advantage of being linear, yielding a multiresolution representation and not being affected by crossterms; this is particularly
relevant when dealing with multicomponent signals.
Under certain conditions, the EMG signal can be considered as the sum of scaled delayed versions of a single prototype. Based on Guglielminotti’s theory, Laterza and Olmo (17) have used wavelet analysis to match the shape of the MUAP. For a unipolar recorded
signal and under certain hypotheses presented by Gabor
in 1946 (18), the typical MUAP shape can be
approximated as the second-order derivative of a Gaussian distribution. The result suggested using the
well-known Mexican hat wavelet, which is indeed the second-order derivative of a Gaussian distribution. The comparison between Mexican hat wavelet and typical unipolar MUAP shape is shown in Figure 2. Based on the
research, Laterza and Olmo concluded that the WT is particularly useful for MUAP detection in the presence of
additive white noise. In this situation, the noise
contributions are spread over the entire time scale plane, independently of the wavelet used. The disadvantage of this proposal (17) was that the Mexican hat wavelet is not perfectly matched to the MUAP shape. Therefore, the obtained results are likely to be subject to further improvement if a perfect matchi ng is performed. In 1998,
I s m a i l a n d A s f o u r ( 1 9 ) c a m e w i t h a t h e o r y s a y i n g t h a t ,
the most common method used to determine the frequency spectrum of EMG are the fast and short term Fourier transforms (FFT and SFT). But they also concluded that the major drawback of these transformation methods is that they assume that the signal is stationary. However, EMG signals are non-
stationary.

Fig. 2: Comparison between Mexican hat wavelet and typical unipolar
MUAP shape.

In 1999, Pattichis and Pattichis (20) discovered that the
WT can also be used to analyze signals at different resolution levels. According to the theory, the process of analyzing signals at different resolution level is known as multiresolution analysis. They analyzed the relationship between wavelet coefficients and the time-frequency
plane. The WT algorithm consists of the decomposition
phase and reconstruction phases. Pattichis and Pattichis briefly outlines how coefficients from each stage of the WT can be used to construct functional approximation to the original signal. Given signal samples x
0, x1, x2….., the
corresponding continuous time signal is given by equation 4:

()( )∑− =
kkoktx tf φ (4)

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 18
where ()kt−φ is called a scaling function. This assumes
that the signal samples are weighted averages of the
continuous signal.
Again in 2003, Kumar et al. (21) came with a similar kind
of proposal saying that the WT decomposes signal into
several multiresolution components according to a basis function called “wavelet function” (WF). The WF is both dilated and translated in the time undertaking a two-dimensional cross correlation with the time domain
sEMG signal. This method can be seen as a mathematical microscope that provides a tool to detect and characterize a short time component within a nonstationary signal. It is the technique that provides information related to the time-frequency variation of the signal. Kumar et al. also concluded that the Short Fourier
Transform (SFT) with the relatively short time windows can attempt to track spectral variation with time, but does not adopt an optimal time or frequency resolution for the nonstationary signal. In (21), sEMG has been decomposed using WT with various WF and the output of the power transform domain is calculated and used as the deciding parameter in choosing the WF that provides the best contrast between sEMG cases. As a result of their
research activity, it can be said that using sEMG and wavelet transforms, it is possible to determine the muscle fatigue (muscle failure) status simply by determining the Sym4 or Sym5 wavelet decomposition of the signal at level 8 and 9 (out of 10 levels). Figure 3 shows the experimental procedure.

Fig. 3: Block diagram of the experiment procedure.

Time-frequency approach

Attempts to gain quantitative information from EMG
recordings have been extensively investigated when signal is represented as function of time (time domain). Cohen class transformation, Wigner-Ville distribution
(WVD), and Choi-Williams distribution are some of the
time-frequency approaches used for EMG signal
processing. Piper (22) showed at the beginning of this century (1912)
that during a sustained muscle contraction the spectral components of the surface myoelectric signal are compressed towards the lower frequencies. The mechanisms that regulate this phenomenon have only been clarified during the last two decades. When sEMG is recorded under dynamic contractions, the assumption
of stationary does not hold because frequency contents of
the signal continuously changes over time. Nonstationaries of the surface myoelectric signal can be classified as slow or fast. Sl ow nonstationaries are mostly
due to the accumulation of metabolites that causes the electrical manifestations of muscle fatigue. Fast
nonstationaries are mainly related to the biomechanics of
the task. Variations in muscle force cause a modification
of the frequency content of the signal. Cohen class transformation proposed by Cohen during
1995 (23) has received considerable attention, particularly in biomedical signal processing. The class time-frequency representation is particularly suitable to analyze surface
myoelectric signals recorded during dynamic
contractions, which can be modeled as realizations of nonstationary stochastic pr ocess. Previous works by
Martin and Flandrin (24), Amin (25) and Syeed and Jones (26) demonstrated that any Cohen class time-frequency spectrum S(t,f) may be written as equation 5:

() ()() { }()()τθ τθτ ττπ πθdtdd e e g tx txE ftSfj tt j / 2 2 /* //,2 2,∫∫∫∞
∞−∞
∞−∞
∞−−−−− + = (5)

where E{} is expectation operator, x(t) is realization of the
stochastic process under consideration, x*(t) is the
complex conjugation, g(θ,τ) is the kernel of the
transformation.
Cohen in 1995 also inferred that when one chooses g(θ,
τ)=1, the resulting distribution is referred as the Wigner-
Ville distribution (23). WVD is optimal to analyze signals constituted by a single comp onent. However, it is not
well-suited for application to multicomponent signals, since the bilinearity of the transform induces the presence of interference terms. Syeed and Jones (26) also
demonstrated that the formulation presented in equation
5 might also be utilized when a single realization of the analyzed stochastic process is available, as is the case when processing surface myoelectric signals recorded dynamic contractions.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 19
The WVD is a time-frequency that can display the
frequency as a function of time, thus utilizing all available information contained in the EMG signal. Although the EMG signal can often be considered as quasi-stationary there is still important information that is transited and may only be distinguished by WVD. Ricamato et al. (27) in 1992 discovered that WVD could
be used to display the freq uency ranges of the motor
unit. It is possible to show the recruitment patterns as the muscle performs difficult tasks. The Wigner-Ville distribution is given in equation 6:

() τττωωτd e tx tx tWj
x−∞
∞−− ∫+=2 2,* (6)

where x(t) and x’(t) are the signal and its complex
conjugate respectively. Implementing the WVD with digital computer requires a
discrete form. This allows the use of fast Fourier
transform (FFT), which produces a discrete-time, discrete-frequency representation. The common type of time frequency distribution is the Short-time Fourier Transform (STFT). According to Davies and Reisman (28)
(1994), the major difficulty with the STFT distribution is that it does not satisfy four important properties that are
desired for time-frequency distributions. The two
properties are the time and frequency marginals and the other two are time and frequency support. They also inform that, the joint density spectrum produced by WVD is very noisy but displays very good localization properties and it is generally concentrated around the instantaneous frequency of th e signal. The Choi-Williams
method proposed in 1993 is an example of a reduced
interference distribution (2 9). Davies and Reisman (28)
discovered that, although the Choi-Williams distribution does not satisfy all the desired properties for a time frequency distribution but it does satisfy an important one, reducing interference. The STFT does not satisfy the
marginal properties. This fact or implies that when a time
slice of the STFT distribution is taken, it does not equal the power density spectrum at that point in time. The same is true for a frequency slice of the distribution. The time support property is not satisfied because the
distribution is not necessarily zero before the signal begins or after it ends. Time frequency techniques
require a very clean signal. There are many other time-frequency distributions. Davi es and Reisman (28) chose
the STFT and Wigner-Ville distributions because they
have been used widely in the past. According to their research, The STFT appears to most clearly show the compression of the spectrum as the muscle fatigue. The WVD has cross-terms and therefore is not a precise representation of the changing of the frequency
components with fatigue. When walls appear in the
Choi-William distribution, there is a spike in the original signal. It will decide if the walls contain any significant information for the study of muscle fatigue.
Autoregressive model

The autoregressive (AR) time series model has been used
to study EMG signal. A surface electrode will pick up
EMG activity from all the active muscles in its vicinity, while the intramuscular EMG is highly sensitive, with only minimal crosstalk from adjacent muscles. Thus, to combine convenience and accuracy there is a great need to develop a technique for estimating intramuscular EMG and their spectral properties from surface
measurement. Researchers have represented sEMG
signal as an AR model with the delayed intramuscular EMG as the input. In 1975, Graupe and Cline (30) first introduced the autoregressive moving average (ARMA) model to represent EMG signals. The empirical result of Graupe
and Cline shows that the EMG could be considered
stationary over sufficient short time intervals. Sherif (31) replaced the model in 1980 because the electrical behavior of the medical deltoi d was nonstationary. Sherif
in his dissertation has emph asized the non-stationary
nature of the EMG and used an AR, integrated moving
average (ARIMA) representation. He characterized the
non-stationary nature of the EMG during different phase
of muscle activity. Doerschuk et al . in 1983 (32) have
approached a problem similar to Graupe and Cline, namely control of prosthetic devices from EMG signals,
by AR models of multiple EMG signals. In 1986, Zhou et
al. (33) represented the surface EMG as an AR model
with the delayed intramuscular EMG signal as the input.
The model, referred to as “tissue filter,” relate the
intramuscular EMG signal waveform to the surface EMG. Assuming that protot ypes of intramuscular and
surface EMG signals are available, the parameters of the time series model that tr ansforms the intramuscular

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 20
signals to the surface signals are identified. The
identified model is then used in estimating the intramuscular signal from the surface signal. This model is illustrated using real EMG waveforms. Hefftner et al.
in 1988 (34) evaluated the previous models and selected an AR model for EMG signature discrimination because of its computational speed. Bernatos et al . in 1986 (35)
employed a static nonlinear element with time-varying
autoregressive moving average (ARMA) model and Moser and Graupe in 1989 (36) proposed a nonstationary
identifier of time-varying AR parameters. In 1992, Tohru (37) considered that the more precise model such as ARMA or ARIMA was not necessary for dynamic muscle movements. The computation cost of ARIMA model is
high, and the determination of the model order is
complex and sometimes difficult. AR model was chosen by Tohru (37) mainly because of its computational cost which is a problem in the simulation. Their investigation was based on AR model parameters computed by quasi-stationary processing.

Artificial intelligence

Some Artificial Intelligence techniques mainly based on
Neural Networks have been proposed for processing
E M G s i g n a l . T h i s k i n d o f t e chnique is very useful for
real-time application like EMG signal recording and analysis. A real-time application of artificial neural network that can accurately recognize the myoelectric signal (MES) is proposed by Del and Park (38) in 1994. According to
their research, MES features are first extracted through
Fourier analysis and clustered using fuzzy c-means algorithm. Fuzzy c-means (FCM) is a method of clustering which allows data to belong to two or more
clusters. The neural network output represents a degree of desired muscle stimulation over a synergic, but enervated muscle. Real time operation is achieved by
taking advantage of hardware multipliers present in
Digital signal processing (D SP) processors to perform
Fast Fourier Transform for feature extraction and neurode input integration for featured classification. Adaptive interfaces are a natural and important class application for artificial neural network (ANN). Error-back propagation method is used as a learning procedure
for multilayred, feedforward neural network. By means
of this procedure, the network can learn to map a set of inputs to a set of outputs. The network topology chosen
was the feedforward variety with one input layer containing 64 input neurodes , one hidden layer with two
neurodes and one output neurode (38). The model using ANN is not only an advance on MES signal recognition in real-time but also, it curtails subjects training to a minimum. Neural network ar chitectures provide a two-
fold solution: a fast way of system customization to the
patient and a better patient adoption to the system, improving the low rate of acceptance of the devices. The method proposed by Del and Park can solve problems (acceptable cost and performance criteria) that conventional statistical methods cannot.
Another ANN based approach is made in 1996 by
Cheron et al . (39) with the objectives of developing an
alternative approach based on artificial dynamic recurrent neural networks (DRNN) to identify the relationship between the muscle EMG activity and the arm kinematics. His objective was to prove that this DRNN identification is bio-mechanically plausible. The
neural network consists of fully interconnected neuron-
like units with two types of adaptive parameters: classical weights between the units and the time constants associated with each neuron. Specifically, this
network identifies some of the complex relationships between the muscle activity EMG and the upper-limb kinematics during complex movements. According to the
method proposed by Pearlmu tter in 1989, the artificial
neural network is a fully connected 20-neuron network. The method is used by Cheron et al . (39), which is
governed by equation 7:
()i i iii I xFydtdyT ++−= (7)

where yi is the state activation level of unit I, F(α) is the
squashing function 1) 1()(−−+=αα e F and xi is given
by equation 8.

∑=
jjji i yw x (8)

The main feature of the proposed DRNN is that its
simulated movements are the result of the interaction
between raw EMG signals without any theoretical
assumptions concerning the type of control. The

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 21
suitability of the DRNN is mainly due to the adaptive
time constants associated to each neuron-like unit. Fuzzy logic systems are advantageous in biomedical
signal processing and classi fication. Biomedical signals
such as EMG signals are not always strictly repeatable
and may sometimes even be contradictory. According to
Chan et al . (40), one of the most useful properties of
fuzzy logic systems is that co ntradictions in the data can
be tolerated. Furthermore, using trainable fuzzy systems,
it is possible to discover patterns in data which are not e a s i l y d e t e c t e d b y o t h e r m e t h o d s , a s c a n a l s o b e d o n e
with neural network. Finally, the experience of medical
experts can be incorporated. It is possible to integrate this
incomplete but valuable knowledge into the fuzzy logic
system, due to the system’s reasoning style, which is
similar to that of a human being. This is a significant
advantage over the artificial neural network (ANN).
Fuzzy logic systems emulate human decision-making
more closely then the ANN. Th e kernel of a fuzzy system
is the fuzzy inference engine. The knowledge of an expert or well-classified examples are expressed as or
transferred to a set of “fuzzy production rules” in the
form of IF-THEN, leading to algorithms describing what
action or selection should be taken based on the
currently observed information (40).

The blind source separation (BSS) method proposed in 2001 by Belouchrani et al. (41) is a neural network based
method that separates a linear mixture of stationary
independent sources received by different sensors by the
use of higher-order statisti cal moments in the learning
algorithm. In 2004, Farina et al. (42) discovered that the
EMG signals generated by different muscles may overlap
in the time and frequency domain, thus classic linear
filtering approaches cannot be applied for the purpose of
source separation. She inform ed that previous studies
aimed at applying BSS approaches to sEMG signals did not provide any validation of the performance and did
not discuss the assumptions and the limitations of the
BSS method to sEMG signal analysis. To overcome the
problems, an approach based on spatial time-frequency
distributions were applied to separate both simulated
and experimental nonstationary sEMG signals (42). Table
2 shows the diagnosis performance of Time domain, Frequency domain and wavelet coefficients using
Artificial Neural Networks. Table 2: Diagnosis performance of time domain,
frequency domain and wavelet coefficients using Artificial Neural Networks.
Feature Set Average %
Time domain 78.3 Frequency domain 62.5 Wavelet DAU4 66.2
Wavelet DAU20 59.6
Wavelet CH 63.3 Wavelet BL 65.8

Higher-order statistics

Higher-order statistics (HOS) is a technique for analyzing and interpreting the characteristics and nature of a random process. The subject of HOS is based on the theory of expectation (probability theory). Due to the limitations of:

i. The detection and characterization existing
nonlinearities in the sEMG signal;
ii. Estimate the phase; and
iii. Exact information due to derivation from normality.
HOS have been introduced in the 1960s and applied in the 1970s.

A statistical method to estimate the amplitude and the
number of newly MUAPs has been proposed by Kanosue et al. in 1974 (43). The method uses the second-
and fourth-order moments with parametric model of the
elementary MUAP waveforms. Low-order models are obtained using second-order statistics (SOS) and provide
parsimonious description of real data. Recently, there has
been an increasing interest towards employing higher-order statistics (HOS). Higher-order Statistics (HOS) is a technique for analyzing and interpreting the characteristics and nature of a random process. The subject of HOS is based on the theory of expectation (probability theory) (1). In 1991, Giannakis and Tsatsanis
(44) used HOS for EMG signal analysis. According to
Giannakis and Tsatsanis, SOS is phase-blind, but has low-variance estimators and when limited to linear-Gaussian processes, they yield computationally and statistically efficient models. In 1995, Yana et al. (45) has
generalized the method to estimate MUAP waveforms and their occurring frequency without any assumption
for the MUAP waveforms. The method was utilized as a
noninvasive method to analyze the forth production

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 22
mechanism of the muscle. According to his theory, H(w)
and λ respectively donate a single MUAP waveform and
its occurring frequency is given by equation 9 and 10.
() () ( )
()2
212 1 2 1
,ωωωωωωλ
BP P P += (9)

()()
λωωPH= (10)

() ( ) ( ) ( )2 1 2 1 21, arg ωωϕωϕωϕωω +−+= B (11)

According to equation 11, () () ()ωϕω ωϕ . argH= is
estimated from the bispectrum using the phase
estimation algorithm. MUAP waveforms can be found
using equation (9) through (11). Their research shows
that MUAP amplitude increases when load weight is increased.
From 1987 to 1993, HOS based signal analysis techniques have been developed by researchers like Nikias, Mendal, Raghuveer, and Petropulu for deterministic and non-
deterministic phase signals, testing of Gaussianity and
linearity, coherence and coupling of the signal, and more. During the 1990s, Nikias et al. (2, 46, 47) had discovered
that the main advantage of HOS over SOS is that HOS can suppress Gaussian noise in detection, parameter estimation, and classification. Nikias informs that HOS is blind to any kind of Gaussian process; a non-zero HOS
measurement can provide a test of the extent of non-
Gaussianity in a signal. Another feature of HOS is that the HOS spectrum of the sum of two or more statistically independent random processes equal the sum of their individual HOS spectra, therefore, HOS can extract information due to derivation from Gaussianity and it provides suitable measurement of the extent of statistical
dependence in time series. Fu rther, the bispectrum, first
member of HOS spectra, carries magnitude and phase information that allows one to recover both the Fourier magnitude and phase value of the system impulse response with the expectation of a linear phase term. In 2000, Kaplanis et al. (48) have given their theory of sEMG
signal analysis using HOS. According to their theory, to
quantify the non-Gaussianity of a random process, the
normalized bispectrum, or bicoherence is estimated according to equation 12: ()()
() () ( )2 1 2 12121,,
ωωωωωωωωπ+=
P P PBB (12)

where P(.) is the power spectrum.

The test of Gaussianity is based on the mean bicoherence
power defined in equation 13 with the summation performed over the non-redundant region.
()∑=21,ωωn g B S (13)

The bicoherence index was used for characterizing the
Gaussianity of the signal. Re sults indicate that sEMG
signal distribution is highly non-Gaussian at low and
high levels of force whereas the distribution has
maximum Gaussianity at mid level of maximum
voluntary contraction level (MVC). A measure of
linearity of the signal, based on deciding weather or not
the estimated bicoherence is constant, follows the reverse
pattern with the measure of Gaussianity. The power
spectrum’s median frequency decreases with the increase
of force.
In 2004, Shahjahan Shahid (1) applied HOS for EMG
signal analysis and characterization due to its advantages of HOS over SOS. He has proposed the “Bispectrum of
Linear System.” Modeling the bispectrum of a time series
signal as the output of a linear system allows an application of useful techni ques for identification and
characterization of the system, which produces the
system output signal. Let e(n) be a zero mean, stationary
random signal applied to a LTI system according to
Figure 4, whose frequency response is H(k) (where the
time domain system response h(n), is causal and stable).
Assume that w(n) is an independent identically
distributed random Gaussian white noise that represents
the system noise and x(n) is the system output.
According to the convolution theorem for a LTI system
output x(n) can be written as equation 14.

∑ +− =∞
=0)( ) ()( )(
knwknekh nx (14)

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 23

Fig. 4: A model of the LTI system.

By using the bicepstrum, a system can be reconstructed
from its output signal upon reconstructing the phase components of a system. Since an LTI system output
signal carries all the information about a system plus
noise, upon considering the system output signals as non-Gaussain noise, it is po ssible to estimate the system
impulse response by using the system reconstruction
algorithm. Algorithms can be developed for system’s
input impulse characterization so that the actual EMG signal from the muscle can be acquired. According to Shahjahan Shahid, traditional system reconstruction algorithms have various limitations and considerable
computational complexity and many show high
variance. The most common bispectrum based system reconstruction algorithm has been improved by separating out the skewness parameter. Also, the
Cepstrum of Bispectrum – a new branch of cepstrum – has
been developed and applied by Shahjahan Shahid to
perform system impulse function reconstruction in a computationally simple manner. The algorithm developed shows better performance than traditional
algorithms. The cepstrum of bispectrum is also used to
develop an algorithm for reconstruction of system input impulse sequence from a LTI system output signal. The results showed that resting muscle’s EMG contains a
train of impulse-like MUAPs whose peaks are oriented to
both sides of zero level. This means that there is no involvement of motor unit in the resting muscle. On the other hand MUAPs tend to be oriented to one side of the zero level when the muscle is contracting. Figure 5 shows
a sample raw EMG signal and its bispectrum curve.

Fig. 5: Sample EMG signal and its bispectrum curve.
Other methods

There are some other models proposed by various
researchers for the purpose of EMG signal processing.
Some of these models are briefly explained here.
In 1969, Rosenfalck (49) mathematically formulated
90 96)(3− =−zex zg based on the experimental works
of Ludin on the intercostals muscle. Nandedkar and
Stalberg (50) modified the expression in 1983 from g(z) to
e(z) = g(2z) in order to match better experimental data,
leading to 90 768)(23− =−zez ze . This is taken as the
default intracellular formulation for the single fiber
action potential modeling.
Nanderdar and Barkhaus (51) has a model proposed in
1992 based on a simple prin ciple of vector summation.
According to Slawnych, Laszlo, and Hershler theory (1990), Nandedkar model assumes that the MUAP amplitude adds algebraically to generate the compound
action muscle protential (CAMP) amplitude. Since
MUAP waveforms do not occur synchronously, this assumption is not valid. If two MUAPs of amplitude A
1
and A2 are summated, then the amplitude of the resulted
waveform is not equal to A1+A 2. In other words, a MUAP
contributes less than its amp litude to CMAP amplitude,
this phenomena is called phase cancellation. According to (51) the amplitude of their sum denoted as A
12 lower
then A1+A2. It is expressed in equation 15.

α. 2212
22
12
12 AA A A A −+= (15)

In 1994 , Englehart and Parker (52) considered two types
of interpulse interval (IPI) probability density function (pdf) models. The discharge sequence as a series of IPI, estimation of the IPI mean, variance, and probability density function (pdf) have been used as descriptors of
motor neuron activity. The Gaussian density function is
expressed by equation 16:
()()



−=22
2exp
21
xx
xxxxf
σµ
πσ (16)

where xµ is the mean and 2
xσ is the variance.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 24
The Gamma density function is expressed by equation
17:
()()αβα
βα
ρβρ
〈


−−


−
Γ=−
xx xxfx ; exp11
(17)

where α is the location parameter, β is the scale
parameter, ρ is the shape parameter and Γ( ) is the
gamma function.

According to the model, the estimates of the moments
and the pdf of the neural discharge sequence are susceptible to bias if the data are nonstationary. Some
factors that may affect the degree of stationary of
experimentally IPI data are du ration of contraction, the
means of force production, and the level of contraction.
An analytical expression for the myoelectric signal was
derived using the integralpulse frequency and amplitude modulation (IPFAM) model by Zhang et al. (53) in 1995.
The model has three main elements: The pulse amplitude
modulation (PAM), the pulse frequency modulation
(PFM) and the linear system. The PAM describes the association of the EMG amplitude with variations in muscular force, the PFM describes the variations in the
EMG signal caused by changes in the nerve firing rates
and the linear system, p(t), represents the compound
motor unit action potential including effects of
propagation dispersion and tissue filtering. In this
model, the potential rises until a pre-determined threshold is reached, which causes an action potential or event to occur. Thus, the IPFAM model includes the most
important features associated with the generation of real
EMG signals. A real-time system for EMG signal analysis was done by
Karlsson and Nystrom in 1995 (54). The aim was to
develop a system for clinical use with the characteristics of graphics feedback, flexible parameter selection,
standard method and flexible addition processing. To
produce a time-frequency representation of a signal, the short-time Fourier transform was proposed to be used. A major drawback of this method was that stationary
signal was assumed. Even when there is no voluntary
change of muscle state, myoelectric signals are non stationary simply due to the inherent physiology of the organs. A model of EMG is proposed by Duchene and Hogrel
(55) in 2000. According to Duchene and Hogrel any new processing algorithm needs to be optimized by comparing its result to the original parameter values to get an optimized criterion. This optimization can only be done if all actual values are known. Only a comprehensive simulation model can help fulfill this
requirement. The extracellular action potential is
calculated after the intracellular action potential for modeling the single fiber action potential. According to the initial work of de Lorente (56), the potential at an observation point [z
0, y 0] can be expressed by equation
18:
()() () ()




∫∫ ∫ ∫∂∂−
∂∂+∂∂=∞+
∞− 12221 1 1,
SS SooE dSrzzedzrzzedS dSrzzeK yzV (18)

where z and y are the axial and radial directions,
respectively, S1 and S2 are the fiber sections at the fiber
ends and r is the distance between the surface elements
dS and the observation point.
Hamilton and Stashuk (57) proposed the latest
simulation of clinical EMG signals so far in 2005.
According to the proposal, the first requirement for EMG signal simulation is the creation of a model of the structure of a muscle. This is performed in the following stages: 1. Muscle and motor unit territory diameter
calculation;
2. MY territory center location;
3. Fiber layout and assignment;
4. Updating MU territory centers;
5. Calculating actual MU territory centers;
6. Assignment of fiber diameters; and
7. Assignment of neuromuscu lar junction locations.

This model is unique because it incorporates the followings:

1. The spatial relationship between muscle fibers, the
MUs they constitute, and the macro level muscle
morphology;
2. MUP calculations combining the clinical
measurement of needle tip and cannula detected voltages contributed by physiologically positioned and activated individual fibers;

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 25
3. Variability of detected MUAPs due to NMJ
transmission delay variability;
4. A new mechanism for MU recruitment based
entirely on muscle morphology; and
5. Clinically realistic needle placement.

EMG signal classification

The common feature for classifying intramuscular EMG
signal is the Euclidean distance between the MUAP waveforms. For clinical interests, the main feature of the EMG signal is the number of active motor unit (MUs), the MUAP waveforms, and the innervations time statistics. According to Wellig and Moschytz (58), the determination of the MUAP waveform and the number
of active MUs can be consid ered as a classification
problem.

The representation of time-triggered and non-overlapping MUAPs produces a shimmer. MUAP shimmer is influenced by the time-offset of the sampled waveforms, local fluctuation of the baseline and background noise. MUAP shimmer can also be influenced by all noises that are different from both background noise and noise ca used by offsets. Besides
background noise and the effects of the offset, white
noise influences the classificati on. If the classification is
to be performed in the wavelet domain, wavelet coefficients which refer to frequency bands lying below 150 Hz should be avoided. The classification with wavelet coefficient needs the wavelet coefficient ( F
f[m,n] )
of four frequency bands ( m=2, 3, 4, 5 ). The classification
performance not only depends on the MUAP shimmer
on the variance within a class but also on the 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 and Peter (59) theorized that the time frequency representation of WVD provided high-resolution signal
characterization in time-frequency space and good noise
rejection performance. This th eory can be very useful for
EMG signal classification pu rposes. For the purpose of
classifying EMG patterns, AR parametric model is used. In 1991, Zhang et al . (60) extracted and compared two
types of features based on signal processing for the purpose of classifying EMG patterns. The two features
were the coefficients of AR parametric models and the
components of Fourier frequency spectra. The method showed better results while describing the EMG linear
envelopes (LE). In 1995, Christodoulou and Pattichis (61) proposed that the classification procedure using ANN is implemented in three phases:

i. In the first phase unsupervised learning is applied
based on one dimensional self-organizing feature map and competitive learning.
ii. In the second phase, in order to improve
classification performance, a self-supervised learning technique, the learning vector quantization is applied.
iii. In the third phase, the actual classification takes
place.

Classification of real EMG data into their constituent Motor unit action Potential is often a difficult task
because of MUAPs waveform variability, jitter to single
fiber potentials and MUPAs superposition. According to Christodoulou and Pattichis ANN appears attractive for the solution of such problem because of their ability to adopt and to create complex classification boundaries.
Figure 6 shows EMG classification strategy using ANN
approach.

Fig. 6: EMG classification strategy using ANN approach.

The DRNN proposed by Chan et al. (40) is much more
adaptive to temporal treatment than the classical feedforward network which is more dedicated to classification tasks. Their resu lt shows that it is successful
in identifying the complex mapping between full-wave
rectified EMG signals and upper-limb trajectory. The training process and classification results of the fuzzy-logic method by Cheron et al. (39) are superior to those of
Neural Network based approaches; primarily in that the
fuzzy system gives more consis tent classification results
and is insensitive to over-training. Typical EMG Classification accuracy rate is given in Table 3.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 26
Table 3: Typical EMG classification accuracy rate.
Method Accuracy rate
Coefficients of AR 99%
Neural Networks 84% Fuzzy System 85%

Motor unit number estimation (MUNE)

Accurate estimation of MUAP templates in the presence of background EMG activity and instrumentation noise is an important requirement of quantitative clinical EMG analysis, especially if EMG signal decomposition is
utilized. MUNE is a procedure used to evaluate the
number of motor axons connected to a muscle. All MUNE techniques rely on assumptions that must be fulfilled to produce a valid estimate. In 1971, McComas proposed a simple neurophysiological technique for estimating the number of motor units in a
muscle (62). A maximal bioelectric response of the
muscle was recorded using sEMG following a supramaximal electrical stimulation of the muscle’s nerve. The maximal EMG response was then divided by an estimate of the average single motor unit response. The result was an estimate of the number of single motor
unit responses that made up the maximal EMG response.
According to Stashuk et al . (63), the number of motor
u n i t s i n a m u s c l e c a n b e e s t i m a t e d b y d i v i d i n g s i z e -related parameter values measured from a maximal M-Wave by corresponding pa rameter values measured
from an average surface-detected motor unit action
potential (S-MUAP). The accuracy of the estimate is
dependent on how representative the average S-MUAP is of the population of S-MUAPs which contributed to the maximal M-Wave. F-Wave responses have been shown to represent the full range of S-MUAP sizes. An automated system was been developed to obtain a maximal M-Wave, to extract a sample of F-Wave
responses, to compute an average S-MUAP and to
estimate the number of MUs in a muscle. In 1998, Zhengquan Xu and Shaojun Xiao (64) presented a method for estimating the mean and standard deviation of inter-pulse intervals (IPIs) of individual MUAP trains. Through a weighted matching between
the observed IPI probability density function and the
modeled function, the firing parameters are estimated. The weighted function is used to approximate the
validity of IPI data so that all valid information provided by IPI data are utilized as far as possible. For this reason, the method can provide reliable estimations even if the MUAP trains are extracted with significant errors. Thus, this method is very useful for estimating the firing statistics of surface EMG where the individual MUAP
trains are difficult to be accurately identified.

Given that the MUAP originates at some distance below
a standard sEMG electrode, the basic shapes of surface
MUAPs can ideally be represented by only a very small number of waveforms or wa velet functions. Based on
this determination, Ping and Rymer (65) in 2003 evaluated ways to estimate the number of MUAPs present in standard surface EMG records, using wavelet based matching techniques to identify MUAP occurrences. The reason for this approach is that
estimates of the numbers of MUAPs are likely to be a
more accurate reflection of the neural command to the muscle. The wavelet matching m e t h o d s , u s i n g a n e v e n
more selective surface electrode, can correctly estimate the number of MUAPs in the surface EMG signals at higher force levels. However, the maximum MUAP number correctly estimated in the surface EMG cannot be significantly increased.

Recently in 2005, Major and Jones (66) used the model to
simulate four MUNE techniques (Incremental
Stimulation, Revised Incremental, Multiple Point Stimulation, and Spike-Triggered Averaging) and have compared the reliability of each. They had also compared the relative utility of using EMG versus force as the output measurement from the muscle. The use of models allows a detailed testing of methodological assumptions
in different MUNE techniques which will lead to a more
accurate and reliable method of performing MUNE. This
will translate into earlier diagnosis and improved treatment assessment of pati ents with neuromuscular
disease. The basic principle underlying the four MUNE
techniques they simulated is the division of the total muscle response by an estimated mean single motor unit
response (SMUP). Muscle responses can be measured
using EMG or force. The surface EMG response of multiple motor units to an electrical stimulus applied to a nerve is known as the compound muscle action potential (CMAP). Thus, the estimated number of

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 27
functional motor units in a muscle (or group of muscles)
is given by equation 19.

) () (max
SMUP meanCMAPN= (19)

Hardware models

Due to the advanced development of the biomedical
science, the application of biomedical instruments
becomes essential in daily life. Design of application
specific integrated circuit for the biomedical instrument has become quite important recently. Various hardware has been implemented to develop prosthetic hands for disabled people. Hardware chips have also been designed to filter EMG signal to achieve the accurate
signal for the prosthetic arm control and other
applications like grasp recognition and human computer interactions.

The microprocessor system for myoelectric signal identification proposed by Graupe et al. (67) is based on
an 8080 Intel Corporation microprocessor which is an 8-bit parallel central processing unit. It is fabricated on a
single Large Scale Integration (LSI) chip using N-channel silicon gates and is furnished in a 40-pin dual in-line
ceramic package, having a 2 µs instruction time. The
microprocessor is then interfaced with its input-output ports and with a 4K-bytes semiconductor memory. Furthermore, to increase speed, the microprocessor is interfaced with a hardware multiplier unit based on Fairchild 9344 4×2 bit multiplier modules where multiplication time is 350 ns versus 1 µs in the
microprocessor itself.
Analog processor chip can be designed to handle the physiological signals. Since EMG signal has the characteristics of very low voltage amplitude and carries some low-frequency common-mode noise, Yen et al. (68)
integrated the instrumentation amplifier, gain control
stage, and filters into the chip for processing the EMG
signal into the adequate amplitude and limited bandwidth. It is divided into three parts: analog signal processing unit, wireless data transmission unit, and digital processing unit. Their research focused on the transmission system design. By the design concept of the
system on a chip, the chip ha s achieved goals of low cost,
low power consumption and minimizing layout area. To enhance the lives of people who has lost a hand,
prosthetic hands have existed for a long time. Evolvable hardware (EHW) chip has been implemented for myoelectric prosthetic hand application. The EHW chip
for an autonomous mobile robot and a myoelectric
artificial hand was also developed in April 1998 to serve as an off-the-shelf device for gate-level hardware evaluation. The chip consists of three components: 1) a
PLA; 2) the GA hardware with a 2K word chromosome
memory and a 2K word training pattern memory; and 3) a 16-bit 33 MHz CPU core (NEC V30; 8086 compatible). Arbitrary logic circuits can be reconfigured dynamically
on the PLA component according to the chromosomes
obtained by the GA hardware. The CPU core interfaces with the chip’s environment and supports fitness
calculations when necessary. The size of the GA
hardware, excluding memories , is about 16K gates. In
terms of gate size, this is almost one-tenth of a 32-bit CPU core (e.g., NEC V830). However, genetic operations
carried out by this chip are 62 times faster than on a Sun
Ultra2 (200 MHz). The chip implemented by Kajitani et
al. in 1999 (69) consists of GA (genetic algorithm)
hardware, reconfigurable hardware logic, a chromosome
memory, a training data memory, and a 16-bit CPU core
(NEC V30). Myoelectric prosthetic hands are operated by signals generated in muscular movement. The proposed
EHW chip consists of seven functional blocks, GA unit,
PLA Unit (2 array), CPU, Re gister File, Random Number
Generator, Chromosome Memory and Training Data Memory. The workflow of the EHW chip can be divided
in two phases. The first phase is to make the two children
and evaluate phase and the second is the “select two chromosome” phase. The GA adaptively implements the circuit on the PLA in the EHW controller.

In 2001, Torresen described a two-step incremental evaluation of a prosthetic hand controller that requires a
floating point CPU or a neural network chip (70). Using gate level EHW, a much more compact implementation can be provided making it more feasible to be installed
inside a prosthetic hand. Such a complex controller could
probably only be designed by adapting the controller to
each dedicated user. It consists of AND gates succeeded b y O R g a t e s . O n e o f t h e m a i n p r o b l e m s i n e v o l v i n g hardware system is that there seems to be limitation in the chromosome string length. A long string is normally
required for representing a complex system. A large
number of generations are required by genetic

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 28
algorithms (GA) as the st ring increases. The main
advantage of the method is that evolution is not performed in one operation on the complete evolvable hardware unit; instead it is performed in a bottom-up way. The digital gate based architecture of the prosthetic hand controller is illustrated in Figure 7 . I t c o n s i s t s o f
one subsystem for each of the six prosthetic motions. In
each subsystem, the binary inputs x
0 . . . x 15 are processed
by a number of deferent units, starting by the AND-OR unit. This is a layer of AND gates followed by a layer of OR gates. Each gate has the same number of inputs, and
the number can be selected to be two, three or four. The outputs of the OR gates are routed to the Selector. This unit selects which of these outputs those are to be
counted by the succeeding counter. That is, for each new
input, the Counter is counting the number of selected
outputs being “1” from the corresponding AND-OR unit. Finally, the Max Detector outputs which counter corresponding to one specific motion having the largest
value. Each output from the Max Detector is connected to the corresponding motor in the prosthesis. If the
Counter having the largest value corresponds to the
correct hand motion, the input has been correctly classified.

Fig. 7: The digital gate based architecture of the prosthetic hand
controller.

Two types of artifacts usually exist in the EMG signal
from an electrically-stimulated muscle: stimulation artifacts and M-wave. In 2000, Peasgood and his researchers (71) assumed that the M-wave is stationary and therefore used a fixed comb filter. But the M-wave is
clearly a non-stationary signal in a statistical sense,
mainly due to the fact that its temporal variation depends on many factors, such as stimulation intensity, fatigue, the contraction level of the muscle, etc. An adaptive prediction error filter (PEF) based on the Gram-Schmidt (GS) algorithm is presented in 2004 by Yeom et
al. (72) for the suppression of the M-waves. The presented filter is implemen ted on a field programmable
gate array (FPGA). Implementation is done using a 6
th
order GS PEF using Xilinx XC2S200pq208-6 FPGA chip.
The design was synthesized using Xilinx ISE 5.2i and
verified using ModelSim XE 5.6a. One major advantage of separating the correlation computation and filtering process in hardware is that the filter system is not
involved with a complicated state machine. Figure 8
shows the schematic of the core processing unit implemented on FPGA. M-waves must be removed in order to use voluntary EMG from electrically stimulated muscle. The proposed M-wave cancellation system based
on the GS PEF is not only efficient to eliminate periodic signals like M-waves, but also suitable to FPGA
implementations than the conventional linear PEF (72).

Fig. 8: Schematics of the core processing unil implemented on FPGA.

Applications of EMG

EMG signals can be used for variety of applications like
clinical/biomedical applications, EHW chip development, human machine interaction, etc. Clinical applications of EMG as a diagnostics tool can include neuromuscular diseases, low back pain assessment, kinesiology and disorders of motor control. EMG signals can be used to develop EHW chip for prosthetic hand
control. Grasp recognition (73) is an advanced
application of the prosthetic hand control.

EMG can be used to sense is ometric muscular activity
(type of muscular activity that does not translate into
movement). This feature makes it possible to define a class of subtle motionless gestures to control interface without being noticed and without disrupting the surrounding environment. The device for this purpose includes a high input impedance amplifier connected to

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 29
electrodes, an anti-aliasing filter, a microcontroller to
sample and process the EMG signal, and a Bluetooth communication module to transmit the processing results. When activation is detected, the controller sends a signal wirelessly to the main wearable processing unit, such as a mobile phone or PDA. Using EMG, the user can react to the cues in a su btle way, without disrupting
their environment and without using their hands on the
interface. The EMG controller does not occupy the user’s hands, and does not require them to operate it; hence it is
“hands free ” (74).

Interactive computer gaming offers another interesting
application of bio-signal based interfaces. The game system would have access to heart rate, galvanic skin
response, and eye movement signals, so the game could
respond to a player’s emotiona l state or guess his or her
level of situation awareness by monitoring eye
movements. An interactive game character could
respond to a user who stares or one who looks around, depending on the circumstances. This use of eye tracking
is easier than using the eyes as a precision pointing
device, which is difficult because the eyes constantly explore the environment and do not offer a stable
reference for a screen pointer. To provide more fun and
strategies, there are usually tw o styles of attack possible
in fighting games. One is the weak attack and the other is
the strong attack. Common input devices for fighting
action games are the joypad and joystick . These use a stick
to move the character and a button to make a certain
type of attack, for example, a punch or kick. To make a
strong attack the user has to input a complex key
sequence that makes that motion difficult to invoke,
thereby achieving a balance between two types of attack. Though those devices are cheap and easy to use, they
have disadvantages. These interfaces are not intuitive for
human fighting movement control, and the user has much to memorize, such as the meaning of the button
and the input sequence for a strong attack motion. A
human-computer interface device designed for a fighting action game, “Muscleman,” has been developed by D. G.
Park and H. C. Kim in Korea. The game characters are
usually depicted as making an isometric contraction of their arms as an expression of power concentration to
make a strong attack like a fireball (75).

Fig. 9: System block diagram of “Muscleman.”

To measure the force of the isometric muscle contraction,
a surface EMG was used. Moreover, to obtain more precise information about the user ʹs forearm movement,
the gaming system is installed with an accelerometer. By analyzing acceleration data record obtained from the
accelerometer, it is possible to know which direction the
forearm is moving. Furthermor e, the classification of
attack movement in cases such as whether the motion was a straight punch motion or an upper cut motion is possible. Wireless transmission is adopted so as not to disturb the user’s motion. By adopting wireless transmission, the stage of a game can be extended
virtually with no limits in space. Figure 9 shows the
system block diagram of “Muscleman.”
At the NASA Arms Research Center at Moffett Field, California, the extension of the Human Senses Group
uses bio-control systems interfaces. They have used EMG/EEG signal in their research program on human
interfaces to flight systems. The group seeks to advance
man-machine interfaces by directly connecting a person to a computer via the human electrical nervous system. Based on EMG and EEG signals, this research applies pattern recognition system to interpret these signals as computer control commands. These NASA researchers have used EMG signal to substitute for mechanical
joysticks and keyboards. As an example, they developed
a method for flying a high-fidelity flight simulator of a transport aircraft using EMG based joystick. Figure 4 shows the flight control using EMG technology. The virtual joystick was actuated through an armband implanted with eight electrodes connected to sensors as
the pilot gestures to land the aircraft. The pilot could also
make emergency landings of a simulated aircraft that had been damaged. Charles Jo rgensen, head of NASA’s
Ames neuroengineering lab, states that this is a fundamentally new way to co mmunicate with machines.
His research group is moving away from the idea of

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 30
controlling machines with levers and knobs. Instead,
they plan to have machines respond directly to human gestures. In addition to aircraft control, the technology might also help astronauts in bulky space suits to control power tools used to work outside the space vehicle, such as in repair activities or construction. A more ambitious idea for reconfigurable airplanes and other
transportation machinery is a virtual wearable cockpit or
command center. The US Air Force and other military branches increasingly use unmanned vehicles for surveillance missions. One way to control these systems from the field is a wearable cockpit. One could use a wearable computer with a wireless link and display goggles, and then employs EMG-based gestures to
manipulate the switches and control sticks necessary for
flight. Noncontact EMG sensors sewn into the field uniform could then sense movements as the acting pilot pretended to manipulate control inputs. A space-based application could let astronauts’ type into a computer despite being restricted by a spacesuit. If a depressurization accident occurred on a long-term space
mission and astronauts needed to access onboard
computers, they could use EMG electrodes in their spacesuits to replicate a computer interface (76). Unvoiced speech recognition – Mime Speech Recognition –
recognizes speech by observing the muscle associated with speech. It is not based on voice signals but EMG. It
will realize unvoiced communication, which is a new
communication style. Because voice signals are not used, it can be applied in noisy environments; it can support people without vocal cords and aphasics (77).
Communication with a computer by certain muscle contractions would make it possible to perform all sorts
of computer-control lable actions using EMG. The muscle
contractions can be detected in a robust way, almost insensitive to any kind of noise, so an interface device b a s e d o n m u s c l e t o n e c o u l d a l s o b e u s e d t o c o n t r o l moving objects, such as mobile robots or an electrical wheel chair which can be great help for persons with
disabilities. Of course, this might offer an alternative for
able-bodied persons as well for controlling home entertainment appliances. The constant stream of EMG signals associated with any arbitrary muscle of the wheelchair driver is monitored and reduced to a stream of contraction events. The reduced stream affects an internal program state which is translated into appropriate commands understood by the wheelchair
electronics. The standard way of steering an electrical wheelchair involves the use of one hand to operate some sort of two dimensional joystick.

DISCUSSION

The study shows that double-threshold detectors are
better than single-threshold detectors because of their higher detection probability. They also allow the user to
adopt the link between false alarm and detection
probability with a higher de gree of freedom than single-
threshold ones. Decomposition of EMG signal by wavelet spectrum matching shows that the technique is accurate, reliable, and fast. Th e technique is very useful
in the study of motor control mechanisms at the SMU level. On the other hand, th e nonlinear LMS optimization
decomposition method based on HOS is also reliable in a
noiseless case. Testing in different levels of additive Gaussian noise found that the well-known HOS robustness leads to satisfactory results also in noisy environments. For EMG signal processing, the WT is an alternative to other time frequency representations. WT has the advantage of being linear, yielding a
multiresolution representation. Crossterms do not affect
WT when dealing with multicomponent signals. We see that a major drawback of SFT is that stationary signal is assumed. The joint density spectrum produced by Wigner-Ville distribution displays very good localization properties and it is generally concentrated around the
instantaneous frequency of the signal. The disadvantage
of WVD is that it is very noisy. Although the Choi-Williams reduces the interferen ce but it does not satisfy
all the other desired properties for a time frequency distribution. While reviewing the properties of fuzzy logic systems, we find that co ntradictions in the data can
be tolerated, which is an advantage. It is also clear that
using trainable fuzzy systems, it is possible to discover
patterns in data which are not easily detected by other methods, as can also be done with neural network. As a
result, Fuzzy logic system s emulate human decision-
making more closely then the ANN. Higher-order statistical (HOS) methods are used for analyzing the EMG signal. This is possible due to the unique properties
of HOS that can be applied to random time series. The
study shows that Gaussian noise can be suppressed using bispectrum or third-order spectrum. Moreover, it carries both the magnitude and phase information,

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 31
which can be used to recover the system impulse
function and input impulse sequence from the linear time-invariant (LTI) system output signal. The main advantage of HOS over SOS is that HOS can suppress Gaussian noise in detection, parameter estimate and classification problem. As HOS is blind to any kind of
Gaussian process, a non-zero HOS measurement can provide a test of the extent of non-Gaussianity in a signal. A summary of the ma jor methods is given in
Table 4.
Table 4: Summary of major methods.
Method Advantage/Disadvantage
Double-threshold detection • Double-threshold detectors are better than sing le-threshold ones because of their higher
detection probability.
• Allow the user to adopt the link between fa lse alarm and detection probability with a
higher degree of freedom th an single-threshold ones.
Wavelet Transform • An alternative to other time frequency representations.
• WT is linear, yielding a multiresolution representation.
• Crossterms do not affect WT when de aling with multicomponent signals.
• A major drawback of SFT is that stationary signal is assumed.
Wigner-Ville distribution • The joint density spectrum produced by WV distribution displays very good
localization properties.
• It is generally concentrated around th e instantaneous frequency of the signal.
• The disadvantage is that it is very noisy.
Choi-Williams method • Reduces interference.
• Does not satisfy all the desired properti es for a time frequency distribution.
Artificial Neural Networks
(ANN) • The network can learn to map a set of inputs to a set of outputs. It is possible to
discover patterns in data which are not easily detected by other methods.
• ANN is not only an advance on MES signal re cognition in real-time but also, it curtails
subjects training to a minimum.
Fuzzy Logic • Contradictions in the data can be tolerated.
• It is possible to discover patterns in data which are not easily detected by other
methods.
• Fuzzy logic systems emulate human decision -making more closely then the ANN.
Higher-order Statistics (HOS) • (HOS) methods may be used for analyzing the EMG signal due to its unique properties
applied to random time series.
• The bispectrum or third order spectrum has the advantage of suppressing Gaussian
noise.
• It carries both the magnitude and phase info rmation, which can be used to recover the
system impulse function and input impulse sequence from the linear time-invariant
(LTI) system output signal.
• HOS is blind to any kind of Gaussian process, a non-zero HOS measurement can
provide a test of the extent of non-Gaussianity in a signal.

If a quantitative relationship between the EMG signal
and force is required, then the contraction must be
isometric. However, even under this constraint the relationship between force and EMG signal remains problematic. It is generally agreed that when the EMG signal is sufficiently smoothed, the relationship is monotonic, but the linearity appears to differ amongst muscles (assuming that there are no technical and other confounding factors such as crosstalk). However,
because the amplitude of the surface EMG signal is a random variable, the instantaneous value of the amplitude is not monotonic with respect to the force value. Furthermore, the estimate of the signal amplitude will vary as a function of force due to intrinsic anatomical and physiological factors. Figure 10 shows the force / EMG signal relationship.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 32

Fig. 10: Normalized Force / EMG signal relationship for three different
muscles. The data have been greatly smoothed, with a window width of 2
s. Note the difference in the linearity of the relationship among the muscles (78).

Physiologists have become accustomed in using the force
output of a muscle as the index of muscle fatigue. In particular, the point at which a contraction can no longer be maintained (the failure point) has been generally designated as the point at which the muscle is said to fatigue. This approach implies that fatigue occurs at a specific point in time; a notion that is inconsistent with
the concept of fatigue accepted by engineers and
physical scientists. Figure 11 shows the EMG signal as a fatigue index.

Fig. 11: A diagrammatic explanation of the spectral modification which
occurs in the EMG signal du ring sustained contractions. The muscle
fatigue index is represented by the median frequency of the spectrum
(78).

While reviewing the hardware implementations we understand that, although reconfigurable hardware
devices, such as FPGA and PLD are spreading rapidly and the usefulness of reconfigurable hardware is being
more widely recognized, reconfiguration in FPGA’s is not autonomous and requires human intervention. Thus, EHW indicates a new direction in reconfigurable hardware beyond FPGA’s.

CONCLUSION

EMG signal carries valuable information regarding the
nerve system. So the aim of this paper was to give brief
information about EMG and reveal the various
methodologies to analyze th e signal. Techniques for
EMG signal detection, decomposition, process and classification were discussed along with their advantages and disadvantages. Disc overy of a problem or
disadvantage in one method leads to other improved methods. This study clearly points up the various types
of EMG signal analysis techniques so that right methods
can be applied during any clinical diagnosis, biomedical research, hardware implementations and end user applications.

REFERENCES

1. Shahid S. Higher Order Statistics Techniques
Applied to EMG Signal Analysis and
Characterization. Ph.D. thesis, University of
Limerick; Ireland, 2004.
2. Nikias CL, Raghuveer MR. Bispectrum estimation: A
digital signal processing framework. IEEE
Proceedings on Communications and Radar 1987;
75(7):869-891.
3. Basmajian JV, de Luca CJ. Muscles Alive – The
Functions Revealed by Electromyography. The
Williams & Wilkins Company; Baltimore, 1985.
4. Kleissen RFM, Buurke JH, Harlaar J, Zilvold G.
Electromyography in the biomechanical analysis of human movement and its clinical application. Gait
Posture 1998; 8(2):143-158.
5. Cram JR, Kasman GS, Holtz J. Introduction to
Surface Electromyography. Aspen Publishers Inc.;
Gaithersburg, Maryland, 1998.
6. Micera S, Vannozzi G, Sabatini AM, Dario P.
Improving detection of muscle activation intervals. IEEE Engineering in Medicine and Biology Magazine
2001; 20(6):38-46.
7. Thexton AJ. A randomization method for
discriminating between signal and noise in

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 33
recordings of rhythmic electromyographic activity. J
Neurosci Meth 1996; 66:93-98.
8. Bornato P, de Alessio T, Knaflitz M. A statistical
method for the measurement of the muscle activation intervals from surface myoelectric signal gait. IEEE Trans Biomed Eng 1998; 45:287-299.
9. Winter DA. Pathologic gait diagnosis with computer-
averaged electromyographic profiles. Arch Phys Med
Rehab 1984; 65:393-398.
10. Lanyi X, Adler A. An improved method for muscle
activation detection during gait. Canadian Conference
of Electrical and Computer Engineering 2004;
1:357-360.
11. Merlo A, Farina D. A Fast and Reliable Technique for
Muscle Activity Detection from Surface EMG
Signals. IEEE Trans Biomed Eng 2003; 50(3): 316-323.
12. Fang J, Agarwal GC, Shahani BT. Decomposition of
EMG signals by wavelet spectrum matching. Procedures of the 19th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society 1997; Chicago, IL, USA. pp. 1253-1256.
13. Zennaro D, Welling P, Koch VM, Moschytz GS,
Laubli T. A Software Package for the Decomposition of Long-Term Multichannel EMG Signal Using Wavelet Coefficients. IEEE Trans Biomed Eng 2003 ;
50(1):58-69.
14. Y a m a d a R , U s h i b a J , T o m i t a Y , M a s a k a d o Y .
Decomposition of Electromyographic Signal by
Principal Component Analysis of Wavelet
Coefficient. IEEE EMBS Asian-Pacific Conference on
Biomedical Engineering 2003; Keihanna, Japan. pp.
118-119.
15. Plevin E, Zazula D. Decomposition of surface EMG
signals using non-linear LM S optimisation of higher-
order cumulants. Proceedi ngs of the 15th IEEE
Symposium on Computer-Based Medical System
2002; pp. 149-154.
16. Guglielminotti P, Merletti R. Effect of electrode
location on surface myoelectric signal variables: a s i m u l a t i o n s t u d y . 9 t h I n t . C o n g r e s s o f I S E K 1 9 9 2 ; Florence, Italy.
17. L a t e r z a F , O l m o G . A n a l y s i s o f E M G s i g n a l s b y
means of the matched wavelet transform. Electronics
Letters 1997; 33(5):357-359.
18. Gabor D. Theory of communication. J Inst Elect Eng
1946; 93:429-457.
19. Ismail AR, Asfour SS. Continuous wavelet transform
application to EMG signals during human gait. Thirty-Second Asilomar Conference on Signals,
Systems & Computers 1998; 1:325-329.
20. Pattichis CS, Pattichis MS. Time-scale analysis of
motor unit action potentials. IEEE Trans Biomed Eng
1999; 46(11):1320-1329.
21. Kumar DK, Pah ND, Bradley A. Wavelet analysis of
surface electromyography to determine muscle
fatigue. IEEE Trans Neural Syst Rehabil Eng 2003;
11(4):400-406.
22. Piper H. Electrophysillogie Muschliche Muskeln.
Basel, Switzerland: Verlag von Julius; 1912.
23. Cohen L. Time-frequency analysis. Englewood Cliffs,
Prentice-Hall; New Jersey, USA, 1995.
24. Martin W, Flandrin P. Wign er–Ville spectral analysis
of nonstationary processes. IEEE Trans Acoust Speech
Signal Processing 1985; 33:1461-1470.
25. Amin MG. Time-frequency spectrum analysis and
estimation for nonstationary random processes. Time-Frequency Signal Analysis Methods and Applications, Ed: B. Boashash, Longman Chesire 1992; Melbourne, Australia, pp. 208-232.
26. Syeed AM, Jones DL. Optimal kernel for
nonstationary spectral estimation. IEEE Trans Signal
Processing 1995; 43:478-491.
27. Ricamato AL, Absher RG, Moffroid MT, Tranowski
JP. A time-frequency approach to evaluate electromyographic recordings. Proceedings of Fifth Annual IEEE Symposium on Computer-Based
Medical Systems 1992; pp. 520-527.
28. Davies MR, Reisman SS. Time frequency analysis of
the electromyogram during fatigue. Proceedings of the 20th Annual Northeast Bioengineering Conference 1994; pp. 93-95.
29. Amin M, Cohen L, Williams WJ. Methods and
Applications for Time Frequency Analysis.
Conference Notes, University of Michigan; 1993.
30. Graupe D, Cline WK. Functional Separation of EMG
signals via ARMA identification. IEEE Trans Syst
Man Cybern 1975; 5:252-259.
31. Sherif MH. Stochastic Model of Myoelectric Signals
for Movement Pattern Recognition in Upper Limb
Prostheses. Ph.D. thesis, School of Engineering and
Applied Sciences, University of California at Los Angeles, 1980.
32. Doerschuk PC, Gustafson W. Upper Extremity Limb
Function Discrimination Using EMG Signal Analysis. IEEE Trans Biomed Eng 1983; 30:18-38.

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 34
33. Zhou Y, Chellappa R, Be key G. Estimation of
intramuscular EMG signals from surface EMG signal
analysis. IEEE International Conference on Acoustics,
Speech, and Signal Processing 1986; 11:1805-1808.
34. Hefftner G, Zucchini W, Jaros G. The
electromyogram (EMG) as a control signal for functional neuro-muscular stimulation part 1:
Autoregressive modeling as a means of EMG
signature discrimination. IEEE Trans Biomed Eng
1988; 35:230-237.
35. Bernatos L, Crago P, Chizeck H. A discrete-time
model of electricity stimulated muscle. IEEE Trans
Biomed Eng 1986; 33:829-838.
36. Moser A, Graupe D. Identification of nonstationary
models with application to myoelectric signals for
controlling electrical stimulation of paraplegics. IEEE
Trans Acoust Speech Signal Process 1989; 37:713-719.
37. Tohru K. Investigation of Parametric analysis of
dynamic EMG signals by a muscle-structured stimulation study. IEEE Trans Biomed Eng 1992;
39(3):280-288.
38. del Boca A, Park DC. Myoelectric signal recognition
using fuzzy clustering and artificial neural networks in real time. IEEE International Conference on Neural
Networks and IEEE World Congress on Computational Intelligence 1994; 5:3098-3103.
39. Cheron G, Draye J-P, Bourgeios M. A Dynamic
Neural Network Identification of electromyography
and trajectory relationship during complex
movements. IEEE Trans Biomed Eng 1996; 43(5):552-
558.
40. Chan FHY, Yang Y-S, Lam FK, Zhang Y-T, Parker
PA. Fuzzy EMG classification for prosthesis control. IEEE Transactions Rehabilitation Engineering 2000;
8(3):305-311.
41. Belouchrani A, Abed-Meraim K, Amin MG, Zoubir
A. Joint-antidiagonalization for blind source separation. Procedures in ICASSP 2001; pp. 2789-2792.
42. F a r i n a D , F e v o t t e C , D o n c a r l i C , M e r l e t t i R . B l i n d
separation of linear instantaneous mixtures of
nonstationary surface myoelectric signals. IEEE Trans
Biomed Eng 2004; 51(9):1555-1567.
43. Kanosue K, Yoshida M, Akazawa K, Fujii K. The
number of active motor units and their firing rates in voluntary contraction of human brachialiis muscle.
Japan J Physiol 1979; 29(4):427-443. 44. Giannakis GB, Tsatsanis MK. HOS or SOS for
parametric modeling? Procedures in IEEE Int Conf
Acoustics Speech Signal Process 1991; 5:3097-3100.
45. Y a n a K , M i z u t a H , K a j i y a m a R . S u r f a c e
electromyogram recruitment analysis using higher order spectrum. IEEE 17th Annual Conference on Engineering in Medicine and Biology Society 1995;
2:1345-1346.
46. Nikias CL, Petropulu AP. Higher-Spectral Analysis:
A Nonlinear Signal Processing Framework. Prentice Hall; New Jersey, 1993.
47. Nikias CL, Mendel JM. Signal Processing with
Higher-Order Spectra. IEEE Signal Processing
Magazine 1993; 10:10-37.
48. Kaplanis PA, Pattichis CS, Hadjileontiadis LJ, Panas
SM. Bispectral analysis of surface EMG. 10th Mediterranean Electrotechnical Conference 2000; 2:770-773.
49. Rosenfalck P. Intra- and ex tracellular potential fields
of active nerve and muscle fibers. Acta Physiol Scand
1969; 321(1):1-169.
50. Nandedkar SD, Stålberg E. Simulation of single fiber
action potentials. Med Biol Eng Comput 1983; 21:158-
165.
51. Nandedkar SD, Barkhaus PE. Phase interaction in the
compound muscle action potential: application to motor unit estimates. IEEE Colloquium on Neurological
Signal Processing 1992; pp. 4/1-4/5.
52. Englehart KB, Parker PA. Single motor unit
myoelectric signal analysis with nonstationary data. IEEE Trans Biomed Eng 1994; 42(2):168-180.
53. Zhang YT, Herzog W, Liu MM. A mathematical
model of myoelectric signals obtained during locomotion. IEEE 17th Annual Conference on
Engineering in Medicine and Biology Society 1995,
2:1403-1404.
54. Karlsson S, Nystrom L. Re al-time system for EMG
signal analysis of static and dynamic contractions. IEEE 17th Annual Conference on Engineering in Medicine and Biology Society 1995; 2:1347-1348.
55. Duchene J, Hogrel J. A Model of EMG Generation.
IEEE Transactions on Biomedical Engineering 2000;
47(2):192-201.
56. de Lorente NR. Analysis of the distribution of action
currents of nerve in volume conductors. Stud
Rockfeller Inst Med Res 1947; 132:384-477.
57. Hamilton-Wright A, Stashuk DW. Physiologically
based simulation of clinical EMG signals. IEEE

Raez et al . – Techniques of EMG signal anal ysis: Detection, processing, classification and applications
www.biologicalprocedures.com 35
Transactions on Biomedical Engineering 2005; 52(2):171-
183.
58. Wellig P, Moschytz GS. Analysis of wavelet features
for myoelectric sign al classification. IEEE
International Conference on Electronics, Circuits and Systems 1998; 3:109-112.
59. Boualem R, Peter O. A Methodology for Detection
and Classification of Some Underwater Acoustic
Signals Using Time-Frequen cy Analysis Techniques.
IEEE Trans Acoust Speech Signal Processing 1990;
38(11):1829-1841.
60. Zhang LQ, Shiavi R, Hunt MA, Chen J. Clustering
analysis and pattern discrimination of EMG linear envelopes. IEEE Trans Biomed Eng 1991; 38(8):777-
784.
61. Christodoulou CI, Pattichis CS. A new technique for
the classification and deco mposition of EMG signals.
Proceedings in IEEE International Conference on Neural Networks 1995; 5:2303-2308.
62. McComas AJ, Fawcett PR, Campbell MJ, Sica RE.
Electrophysiological estima tion of the number of
motor units within a human muscle. J Neurol
Neurosurg Psychiatry 1971; 34(2):121-131.
63. Stashuk DW, Kassam A, Doherty TJ, Brown WF.
Motor Unit Estimates Based on the Automated Analysis of F-Waves. Proceedings of the Annual
International Conference on Engineering in Medicine and Biology Society 1992; 14:1452-1453.
64. Zhengquan X, Shaojun X. Estimation of motor unit
firing statistics from surface EMG. Proceedings of the
20th Annual International Conference on Engineering in Medicine and Biology Society 1998; 5:2639-2642.
65. Ping Z, Rymer WZ. Motor unit action potential
number estimation in the surface electromyogram: wavelet matching method and its performance
boundary. Proceedings of the 1st International IEEE
EMBS Conference on Neural Engineering 2003; pp. 336-
339.
66. Major LA, Jones KE. Simulations of motor unit
number estimation techniques. Journal of Neural
Engineering 2005; 2(2):17-34.
67. G r a u p e D , M a g n u s s e n J , B e e x A . A m i c r o p r o c e s s o r
system for multifunctional control of upper-limb prostheses via myoelectric signal identification. IEEE
Transactions on Automatic Control 1978; 23(4):538-544. 68. Yen C-J, Chung W-Y, Lin K-P, Tsai C-L, Lee S-H,
Chen T-S. Analog integrated circuit design for the wireless bio-signal transmission system. The First
IEEE Asia Pacific Conference on ASICs 1999; pp. 345-
346.
69. Kajitani I, Murakawa M, Nishikawa D, Yokoi H,
Kajihara N, Iwata M, Ke ymeulen D, Sakanashi H,
Higuchi T. An evolvable hardware chip for
prosthetic hand controller. Proceedings of the Seventh
International Conference on Microelectronics for Neural, Fuzzy and Bio-Inspired Systems 1999; pp. 179-186.
70. Torresen J. Two-Step Incremental Evolution of a
Prosthetic Hand Controller Based on Digital Logic Gates. ICES 2001; pp. 1–13.
71. Peasgood W, Whitlock T, Bateman A. EMG-
controlled closed loop electrical stimulation using a digital signal processor. Electronics Letters 2000;
36(22):1832-1833.
72. Yeom HJ, Park YC, Yoon YR, Shin TM, Yoon HR. An
adaptive M-wave canceller for the EMG controlled functional electrical stimulator and its FPGA
implementation. Conference Proceedings in 26th
Annual International Conference of Engineering in Medicine and Biology Society 2004; 6:4122-4125.
73. Ferguson S, Dunlop G. Grasp Recognition From
Myoelectric Signals. Procedures Australasian
Conference Robotics and Automation 2002; pp. 78-83.
74. Stanford V. Biosignals offer potential for direct
interfaces and health monitoring. Pervasive
Computing, IEEE 2004; 3(1):99-103.
75. Park DG, Kim HC. Muscleman: Wireless input
device for a fighting action game based on the EMG signal and acceleration of the human forearm. [http://melab.snu.ac.kr/Research/melab/doc/HCI/muscleman_paper.pdf].
76. Wheeler KR, Jorgensen CC. Gestures as input:
neuroelectric joysticks and keyboards. Pervasive
Computing, IEEE 2003; 2(2):56-61.
77. Manabe H, Hiraiwa A, Sugimura T. Unvoiced
Speech Recognition using EMG-Mime Speech Recognition. Conference on Human Factors in
Computing Systems 2003; pp. 794-795.
78. de Luca CJ. The use of surface electromyography in
biomechanics. J Appl Biomech 1997; 13:135-163.