Critical Reviews™ in Biomedical Engineering, 38(5):467–485 (2010)

A Review of Clinical Quantitative Electromyography

Charles Farkas,1 Daniel W. Stashuk,1* Andrew Hamilton-Wright2, & Hossein Parsaei1
1
Department of Systems Design Engineering, University of Waterloo, Ontario, Canada; 2Department of Math and
Computer Science, Mount Allison University, New Brunswick, Canada

*Address all correspondence to Dr. Daniel W. Stashuk, Department of Systems Design Engineering, University of Waterloo, 200 University Avenue West, Waterloo,
Ontario, Canada, N2L 3G1. Tel.: 519-888-4567, ext. 32982; Fax: 519-746-4791; e-mail: stashuk@uwaterloo.ca.

ABSTRACT: Information regarding the morphology of motor unit potentials (MUPs) and motor unit firing pat-
terns can be used to help diagnose, treat, and manage neuromuscular disorders. In a conventional electromyographic
(EMG) examination, a clinician manually assesses the characteristics of needle-detected EMG signals across a
number of distinct needle positions and forms an overall impression of the condition of the muscle. Such a subjec-
tive assessment is highly dependent on the skills and level of experience of the clinician, and is prone to a high error
rate and operator bias. Quantitative methods have been developed to characterize MUP waveforms using statistical
and probabilistic techniques that allow for greater objectivity and reproducibility in supporting the diagnostic pro-
cess. In this review, quantitive EMG (QEMG) techniques ranging from simple reporting of numeric MUP values
to interpreted muscle characterizations are presented and reviewed in terms of their clinical potential to improve
status quo methods. QEMG techniques are also evaluated in terms of their suitability for use in a clinical decision
support system based on previously established criteria. Aspects of prototype clinical decision support systems are
then presented to illustrate some of the concepts of QEMG-based decision making.

KEY WORDS: EMG signal; motor unit potentials; motor unit potential trains; motor unit firing patterns; quan-
titative EMG; clinical decision support

I. CLINICAL QUANTITATIVE ELECTROMYOGRAPHY

Neuromuscular disorders change the morphology and activation patterns of the motor units of the af-
fected muscles. As depicted in Figure 1, myopathic disorders in general maintain the number of motor
units in a muscle but reduce the number of muscle fibers in motor units due to fiber necrosis, atrophying
some fibers while hypertrophying others and thus increasing the range of fiber diameters and densities
in the motor units. To compensate for fiber loss and diameter changes, more motor units are required to
be recruited and to fire more often in myopathic muscle than in healthy muscle to create a specific level
of contraction. In contrast, neurogenic disorders in general cause a loss of motoneurons and thus motor
units. However, denervated fibers are reinnervated by axonal sprouts of surviving motor units, resulting
in motor units with large territories and numbers of fibers, and uneven motor unit fiber densities. These
motor unit fiber changes are also depicted in Figure 1. As a result of motor unit loss and reinnervation,
fewer motor units are required to be recruited in neurogenic muscle than in healthy muscle to create a
specific level of contraction.
Characteristics of suitably detected electromyographic (EMG) signals can reflect the morphology
and activation patterns of the motor units of a muscle, and can therefore be used to infer the degree and
type of disorder that may have affected the muscle. Therefore, EMG signals suitably detected, processed,

ABBREVIATIONS

EMG, electromyographic; LDA, linear discriminant analysis; MUP, motor unit potential; MUPT, motor unit po-
tential train; QEMG, quantitative electromyography

0278-940X/10/$35.00 © 2010 by Begell House, Inc. 467

468 Farkas et al

FIGURE 1. Schematic representation of the effects of the various categories of neuromuscular disor-
ders (Reprinted with permission from Stalberg and Falck1). A, Normal muscle; B, neurogenic muscle,
and C, myopathic muscle.

and analyzed can be used to detect neuromuscular
disorders. In this regard, the differences between
needle- and surface-detected EMG signals need
to be considered. Although surface-detected EMG
signals have many areas of application and have
received much recent attention, in general, they do
not provide the sensitivity needed for clinical use.
Typical surface electrodes cannot be positioned
sufficiently close to motor unit territories to reflect
subtle morphological changes, and the respective
changes in the EMG signal characteristics become
obscured and confounded. In addition, surface
electrodes typically cannot be used to statistically
sample motor units throughout a muscle, and do
not provide enough discrimination between the
motor unit potentials (MUPs) of near and distant
motor units. One exception to this is the use of
high-density selective surface EMG techniques
capable of resolving the spatial and temporal ar-
rangements of the MUPs of concurrently active
motor units.2-4 In principle, such techniques could
be used clinically. However, their use is currently
impeded by instrumentation complexities and their
clinical utility is only beginning to be evaluated.
This review therefore focuses solely on the clinical
use of needle-detected EMG signals.
Current clinical practice is to qualitatively
interpret insertional, spontaneous, and voluntary

Critical Reviews™ in Biomedical Engineering

A Review of Clinical QEMG
needle-detected EMG activity to determine the
degree and type of disorder that has affected a
muscle under examination. Although such qualita-
tive assessments allow an experienced physician to
infer a host of underlying conditions and diseases,
it is subjective and related to the skill and experi-
ence of the physician. Moreover, it is lacking the
precision of quantitative analysis.
Quantitative EMG (QEMG) is the process of
extracting quantitative information related to the
morphology and activation of motor units from
EMG signals.1,5 Such information can be used for
exploration of the underlying structure and opera-
tion of the peripheral neuromuscular control system,
and for the clinical characterization of muscle tissue
with respect to detection of disease or abnormality.
Several QEMG techniques have been developed to
quantize important aspects of needle EMG exami-
nations, resulting in a more precise numerical and
statistical representation of their results.5-10 These
techniques provide a wide range of outputs, from
a simple reporting of numeric MUP values to in-
terpreted muscle characterizations. QEMG-based
muscle characterizations, and quantitative methods
in general, can have a significant impact by present-
ing quantitative information in a way that supports
the clinical decision-making process. QEMG tools
have the advantage of higher accuracy and greater
precision, but must be tailored to a specific purpose
and thus are not intended to replace current clini-
cal examinations but to augment certain aspects by
improving their accuracy and consistency.
QEMG techniques have been demonstrated to
be effective in a number of research studies using
human data.11-20 They have also been applied in vet-
erinary practice, wherein QEMG-based techniques
were shown to be more effective than the analysis
of muscle enzyme activity for monitoring training
adaptation in racehorses.21 Nonetheless, despite
their potential for improving subjective analysis,
QEMG techniques are not widely understood or
used by clinicians. One of the main impediments
to the widespread acceptance of QEMG tech-
niques is the inherent difficulty of presenting vast
amounts of quantified data in a way that can be
easily interpreted. It is difficult for human decision
Volume 38, Number 5, 2010
469
makers to interpret and mentally compare numeric
information without the necessary context or level
of explanation. Such shortcomings of quantified
data highlight a specific characteristic that deci-
sion support and QEMG methods must possess:
transparency. Transparency is especially important
when dealing with complex data structures, such as
multivariate distributions of MUP feature values
or other statistical parameters. A system that can
explain its conclusions in a manner that is consis-
tent with, and analogous to, the cognitive processes
employed by a human decision maker can thus
improve upon the quality and accuracy of the in-
ferences made. Therefore, the degree to which the
findings and supporting evidence can be presented
and explained to support the decision-making pro-
cess is an important factor to consider when devel-
oping a clinical decision-support system. How the
findings and supporting evidence are presented and
explained to support the decision-making process
is also very important.
QEMG techniques attempt to reproducibly
extract useful muscle information from sets of fea-
ture values used to represent important EMG signal
characteristics. Features can represent characteris-
tics of a composite EMG signal (i.e., interference
pattern) or of individual MUP trains (MUPTs) ex-
tracted from a composite EMG signal using either
simple triggering techniques or more sophisticated
decomposition techniques. With respect to the
latter, time domain features related to MUP size,
shape, and stability or features of transformed or
modeled MUP waveforms have been used. QEMG
methods use rigorous statistical and/or probabilis-
tic inference in which feature values from EMG
signals detected in the muscle under examination
(i.e., test EMG signals) are compared with respect
to distributions of corresponding values obtained
from exemplary EMG data. By describing the sta-
tistical similarities or differences between test val-
ues and exemplary data, it is possible to character-
ize EMG signals or constituent MUP waveforms
acquired from a muscle under examination. This
allows a clinical decision maker to use to the fullest
the information present in the test EMG signals
and, in addition, to leverage information extracted
470
from previously acquired (exemplary) signals. Such
exemplary data must be acquired and stratified by
an expert clinician. In addition, it must be specifi-
cally acquired for each muscle or groups of muscles
studied, and should be stratified by age or any other
relevant stratification criteria (e.g., gender) and in-
corporated into a repository (database).
QEMG analysis has the advantage of provid-
ing greater objectivity and consistency, and is use-
ful for equivocal cases to increase the certainty of
a diagnosis.22 Furthermore, the precision obtained
with more sophisticated quantitative techniques
can provide continuous measures that also relate to
the level of involvement of a disorder.23 Compared
with qualitative methods, QEMG methods can be
more robust and reproducible. They can provide
analytic confidence (i.e., provide measures of un-
certainty or error) and have the ability to generalize
across sets of EMG signals.
The reproducibility and robustness of EMG
data assessment are two important criteria used to
evaluate QEMG methods. Being able to discuss
the degree to which similar feature values can be
obtained from similar EMG signals, coupled with
discussions related to the sources and types of con-
founding errors present, makes it possible to consider
the entire workflow of obtaining decision-making
data from an EMG signal in terms of reproducibil-
ity, robustness, and the general quality of the out-
come measures provided by EMG-based analysis.
Whether EMG data are used for the exploration
of muscle structure and function or to detect and
characterize disease, it is clear that metrics for the
quality of the data are invaluable, and by measur-
ing and improving data quality, any inference made
based on this data will also be improved.
Regardless of whether a qualitative or quan-
titative assessment is used, a similar hierarchical
process is followed. First, in order to account for
the large variability in motor unit size and MUP
shape throughout a muscle, MUP or/and other sig-
nal features are assessed at several needle positions
within a muscle. The data from these various needle
positions are then characterized based on whether
they possess attributes consistent with certain dis-
ease processes. The characterized sampled data are
Farkas et al
then combined to arrive at an overall impression
or characterization of the muscle. Finally, a rule or
heuristic is applied to categorize the muscle based
on the characterization measures obtained.
I.A. Needle EMG Examination
A recent review by Daube and Rubin22 describes
in detail how to properly complete a clinical nee-
dle EMG examination. Briefly, a needle electrode
(monopolar or concentric) is inserted into the su-
perficial layers of a muscle and the resultant detect-
ed EMG signals are assessed based on three types
of detected activity: insertional, spontaneous, and
voluntary. In all cases, the electrode must be posi-
tioned to sample several regions of the muscle to
obtain a sufficient statistical sampling of the mo-
tor units. When assessing voluntary activity, EMG
signals are assessed at a low level of contraction,
with only a few active motor units contributing to
the signal. However, if decomposition techniques
are used, EMG signals with five to seven motor
unit contributions can be assessed.22 Either way, it
is important to position the needle to detect MUPs
with a rapid rise time to ensure that the detection
surface is sufficiently close to the fibers of the ac-
tive motor units. It is also important to ensure that
the level of each contraction is consistent, so that
abnormally large or small MUPs may be attrib-
uted to morphological changes in the muscle and
thus representative of disease and not be related
to different sizes of active motor units due to dif-
fering levels of muscle activation. QEMG tech-
niques have not as of yet been applied to study
EMG signals related to insertional or spontaneous
muscle activity. Therefore, the following discussion
focuses on EMG signals detected during volun-
tary muscle activity.
I.B. Qualitative Characterization
Traditionally, needle EMG signals detected dur-
ing voluntary activation are qualitatively charac-
terized based on visual and auditory assessment of
the morphology, stability, and times of occurrence
of their MUPs. In addition, the intensity of com-
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
posite EMG signals (i.e., interference patterns) are
described using terms ranging from full to sparse
in an effort to assess motor unit activation. These
characterizations are usually manually quantized
and charted. While such qualitative assessment is
prone to a high error rate, a skilled practitioner can
use it to detect not just the broad category of dis-
ease, but also certain processes that are symptomatic
of a specific disease or group of disease processes.
MUP analysis is useful in determining the type
of disorder (myopathic or neurogenic), as well as
the time course and severity of the disease. How-
ever, qualitative analysis is limited to the MUPs of
only a few active motor units at a time. Motor unit
morphology is inferred based on the assessment of
the duration, amplitude, and number of phases of
detected MUPs. During qualitative assessment, a
physician keeps track of abnormally small or large
MUPs and polyphasic MUPs and assigns rankings
based on severity.
MUP stability refers to consistency of MUP
morphology across the MUPs of a MUPT and can
indicate impaired transmission across the neuro-
muscular junctions of a motor unit. Although MUP
instability is typically a sign of disease of the neuro-
muscular junctions, such as myasthenia gravis, any
disorder associated with denervation/reinnervation
may cause unstable MUPs.
Motor unit activation patterns are assessed
in terms of the level of motor unit activation and
recruitment. Low activation represents a central
process such as a central nervous system disorder
or manifestation of pain. Compromised motor
unit recruitment is found in neurogenic diseases,
and sometimes in end-stage myopathy, whereas
early recruitment (i.e., increased activation) is typi-
cally a sign of myopathy. According to Preston and
Shapiro,23 motor unit activation analysis is one of
the most difficult tasks for an electromyographer.
Qualitative motor unit activation analysis is limited
to very low levels of contraction, in which only a
small number of motor units are active; otherwise,
the interference pattern makes motor unit activa-
tion become qualitatively indiscernible. Daube and
Shapiro describe the normal and abnormal varia-
tions of motor unit activation characteristics, and
Volume 38, Number 5, 2010
471
provide further references to standardized values
for manual visual and auditory assessment.22,24
II. QUANTITATIVE CHARACTERIZATION
This review focuses on QEMG techniques using
features extracted from individual MUPTs. Quan-
titative characterization of extracted MUPTs can
be accomplished in several ways and with vary-
ing degrees of complexity. MUPTs can be isolated
semi-automatically using level or window trigger-
ing, or automatically using signal decomposition
methods. In principle, any of these isolation tech-
niques can be used as long as the MUPTs used to
train a system are isolated in the same way as those
being characterized. Standardization across EMG
clinics requires the use of consistent needle elec-
trodes, filter settings, and detection protocols. Fur-
thermore, training data must be grouped based on
the specific muscle, age, and possibly gender. Oth-
erwise, even though numerical data are extracted,
the final interpretation may be biased and have a
large margin for error.
Properly performed, the most immediate ben-
efit of quantitative characterization is the provision
of more precise continuous values for features like
duration, amplitude, and number of phases/turns.
Continuous-value features make it possible to
derive relationships among features such as area,
thickness, and size index,25 as well as measures of ir-
regularity, such as the “irregularity coefficient.”26,27
Decomposition-based quantitative analysis5,6,7
has the advantage that EMG signals detected dur-
ing higher levels of contraction can be analyzed
(usually five to seven MUPs), and features related
to MUP stability, such as jitter,28 and to motor
unit activation patterns, such as the mean and
standard deviation of the inter-discharge intervals
of a MUPT, can be provided. The automation of
decomposition methods also provides a consider-
able time-saving advantage, especially with the
recent availability of cost-effective computational
resources.29-32
The first QEMG methods developed charac-
terize MUPs based on morphological features of
the shape and size of a template or average MUP,
472
because this is similar to what is done qualitatively
and it is easy to define and measure these features.
Quantitative MUP characterizations based on
transformations of the MUP template have also
been developed, but suffer from less transparency
and greater complexity. Quantitative analogs for
characterization based on measures of MUP sta-
bility and motor unit activation have not yet been
developed. While such measures have been quanti-
fied, their values are currently reported solely for
qualitative interpretation.
II.A. MUP Characterization
MUP characterization and classification was in-
troduced by Pattichis et al. and further developed
by several groups of investigators.33-38 Based on a
training set of features extracted from exemplary
MUP templates, MUP characterization calculates
a set of scores that indicate the likelihood that the
muscle from which a MUP template was calculat-
ed is affected by a specific category of disease based
on the characteristics of the MUP template. In a
two-category case, the likelihood of a muscle be-
ing normal or abnormal conditioned on the char-
acteristics of the MUP template is estimated. In a
three-category case, which is more common, the
likelihood of a muscle being normal, myopathic,
or neurogenic conditioned on the characteristics
of the MUP template is estimated. Features of the
MUP template used for training and testing can
be morphological, extracted from a time domain
representation, or spectral, related to a frequency
domain representation.
1. Conventional Classifiers
Several MUP characterization techniques based
on classifiers that use continuous valued features
have been proposed. Linear discriminant analysis39
was used by Pfeiffer and Kunze, and later by Pino,
Stashuk, and Podnar as a candidate for a muscle
characterization methodology.11,12,23 Linear dis-
criminant analysis statistically transforms data in an
attempt to minimize within-class scatter (variance)
while simultaneously maximizing between-class
Farkas et al
scatter to achieve an optimal decision function for
each muscle category. Pino et al.14,40 evaluated the
performance of linear discriminant analysis deci-
sion trees, and a standard naive Bayes classifier,39
and found that MUP categorization performance
was comparable across these methods. The advan-
tage of conventional classifier techniques is that
their simplistic nature allows for some degree of
transparency, provided that a clinician has a mod-
erate statistical background, although their perfor-
mance is subject to the data upholding certain as-
sumptions, such as obeying a Gaussian distribution
and independence across features.
2. Advanced Pattern Recognition Tech-
niques
When MUP categorization accuracy is considered
as the evaluation criteria, conventional classification
techniques do not provide the best performance.
Therefore, advanced pattern recognition techniques
have been used to improve performance. In these
studies, the true categories of MUP templates were
manually determined by an expert neurophysiolo-
gist to create a database of MUP templates. In their
early work, Pattichis et al.33,41 were the first to de-
velop a decision-making strategy based on artificial
neural networks combined with Kohonen self-or-
ganizing feature maps and a learning vector quanti-
zation technique to perform MUP characterization
based on morphological MUP features. Katsis et
al. 37 used artificial neural networks with radial ba-
sis functions and probabilistic neural networks in
a two-stage classification approach. The first stage
used an artificial neural network or probabilistic
neural network to discriminate between normal
and abnormal MUPs, and the second stage used
a C4.5 decision tree42 to determine whether the
abnormality was myopathic or neurogenic. Several
techniques using alternate MUP representations
have also been suggested. Pattichis and others have
used autoregressive and cepstral analysis to rep-
resent MUPs,35 as well as the wavelet transform
method,34 while Katsis et al.36 used MUP time
domain samples and a support vector machine to
perform raw EMG-based classification of decom-
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
posed MUPs. Although these advanced pattern
recognition-based MUP characterization tech-
niques are robust and accurate, several drawbacks
have been noted. For example, artificial neural net-
work methods have a tendency toward over-fitting,
which results in a difficulty of generalizing to new
data.33,41 In addition, while most of the advanced
pattern recognition techniques, including artificial
neural networks, support vector machines and oth-
ers offer greater sensitivity and specificity, they do
not provide good transparency. It is difficult to in-
tuitively appreciate how they make their decisions.
Poor transparency leads to poor acceptance of the
results and the techniques by clinicians.
3. Transparent Rule-based Techniques
The issue of transparency can be addressed by em-
ploying rule-based methods. The two-stage meth-
od developed by Katsis et al.37 is able to provide a
level of transparency and interpretability by using
a decision tree to discriminate between myopathic
and neurogenic categories. Pino14,40 used the pat-
tern discovery technique developed by Wang and
Wong43 to characterize MUPs. Pattern discovery
quantizes all continuous feature values into dis-
crete events. In a training set, patterns of events
across the features used to represent a MUP and
including a specific muscle category, which oc-
cur more often than expected under the assump-
tion of independent features and categories, are
selected by pattern discovery as rules. The entire
set of discovered rules form a knowledge-based
system that can be used for the categorization of
test MUPs. If a pattern of events created by the
quantized feature values of a test MUP matches
a rule associated with a pattern discovered in the
training data that occurred more often than ex-
pected, it is used as positive evidence and supports
the category contained in the rule. Alternatively,
if the matched rule is associated with a pattern
discovered in the training data that occurred less
often than expected, it is used as negative evidence
and refutes the category contained in the rule. The
degree of support or refutation is measured by the
weight of evidence statistic.43
Volume 38, Number 5, 2010
473
While pattern discovery was been shown to
have accuracy comparable to other classification
techniques,13,40 it is not considered to be as robust
as some of the other classification methods because
of the need for discretizing the continuous feature
values. The number of intervals (“bins”) used for
discretization must be carefully selected to manage
the trade-off between the level of granularity (and
accuracy) and the quality of evidence. For example,
if the number of intervals increases, more training
data are needed to determine if specific patterns are
significant. Decreasing the number of intervals re-
sults in larger discretization ranges, which decreases
the precision of the MUP representation.
In other areas of medicine, fuzzy-logic-based
inference systems offer an attractive alternative
to conventional characterization techniques. The
general idea is to represent hard numeric values
by linguistic expressions, in which a membership
function determines the degree to which a variable
falls within a particular decision boundary. The
benefit of such a system is that it allows quanti-
ties to be represented by linguistic qualifiers like
somewhat high, high, very high, and expressions such
as hot, cold, warm. Hamilton-Wright38 proposed
a fuzzy inference system that augmented pattern
discovery with fuzzy logic theory to yield a hybrid
system. Discretization error is greatly reduced be-
cause values within a bin are assigned memberships
based on their continuous value. The “fuzziness” of
membership values allows multiple rules with the
same feature value to be considered simultaneously,
selecting the rule that yields the best discrimina-
tion. Information presented in this way is also more
in line with the way humans interpret and reason
with data they are presented with, making it an ex-
cellent candidate for use in decision support.
4. Evaluation of MUP Characterization Tech-
niques
To be clinically useful, a classifier must meet cer-
tain requirements in addition to being accurate. A
more complete description of these requirements
is presented in section III.A. One of the most im-
portant of these requirements is transparency (i.e.,
474
the ability for a clinician to understand the crite-
ria used to make a specific classification decision).
Conventional classifier methods are simplistic
enough that they may provide some level of trans-
parency to clinicians familiar with the concepts
of differential diagnosis. However, these methods
have several drawbacks, such as unmet assumptions
regarding feature data distributions that can cause
the accuracy of these methods for MUP categori-
zation to be quite poor. Advanced pattern recogni-
tion techniques, such as artificial neural networks
and support vector machines, are quite robust and
offer better performance. However, artificial neural
networks are essentially “black-box” classifiers that
do not permit an explanation of their output. Even
if such an explanation were possible, it would not
be in a form that is easily understood by a human
decision maker. Support vector machines also suf-
fer from complexity and require a deeper apprecia-
tion of their theoretical mechanisms. While pro-
viding accuracy comparable to other methods, the
pattern discovery and fuzzy inference techniques
introduced by Pino and Hamilton-Wright offer a
completely transparent classification scheme. Pat-
tern discovery is unique in that it combines infor-
mation theoretic principles with linguistic inter-
pretation. At the heart of the method lies a robust
statistical and probabilistic analysis, but the rules
derived from patterns in the data are in a form that
can be intuitively understood and easily displayed.
In addition, pattern discovery and fuzzy inference
are capable of handling continuous and discrete
feature values as well as missing data. Further, it is
believed that the rule-based nature of pattern dis-
covery allows it to generalize its conclusions to new
data better than other classification techniques.
II.B. Muscle Characterization
Given the wide variability of MUPs detected in a
muscle dependent on the specific needle positions
during signal detection and the varying degrees to
which a disorder may affect specific motor units,
individual MUP characterization scores are not
robust indicators of the actual category or state of
a muscle (i.e., myopathic, normal, or neurogenic)
Farkas et al
and cannot be used in isolation for accurate diag-
nosis. A more robust indicator can be achieved if
several MUP characterizations across a muscle are
aggregated. During qualitative analysis, a clinician
also aggregates, based on experience and training,
the information extracted from the MUPs and in-
terference patterns examined to create an overall
clinical impression of the muscle under examina-
tion. Unfortunately, the consistency and accuracy
of the clinical impression obtained is dependent on
the experience of the clinician.
1. Statistical Muscle Characterization
The most common method used to address the
shortcomings of subjective qualitative interpre-
tation is to use summary statistics of quantified
MUP features. Statistical techniques go beyond
simply summarizing quantitative data, and attempt
to characterize a muscle based on the distributions
of their sampled MUP parameters. Such predictive
analysis techniques form the basis of the early deci-
sion support techniques.13,14,18,19,44,45
Common practice has been to calculate mean
values for morphological MUP features, such as
duration, amplitude, and number of phases across
sets of MUPs, and to compare them with normative
standards reported in the literature to get a muscle-
level impression of the state of disease involve-
ment. For example, the early works of Buchthal46
are still used as standards for formal quantization,
and Podnar15-17,47-49 has published predictive values
for limb, genioglossus, and anal sphincter muscles.
Researchers have cited constraints on the mini-
mum number of motor units sampled in a muscle
(i.e., the preferred number of MUPs sampled for
robust diagnosis), and the general consensus is that
a minimum of 20 MUPs should be sampled.18,50
However, Podnar showed that sampling additional
MUPs can increase sensitivity and specificity.51
The major drawback with the statistical methods
is that for each MUP feature, suitable threshold
values for defining normality/abnormality must be
established, as well as rules regarding the number
of abnormal features required to declare a muscle
as abnormal.
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
To implement a statistical method, a set of train-
ing data consisting of exemplary “normal” MUPs
is used to generate distributions for each feature.
Mean44 and/or outlier45 threshold values are then
defined, and feature values of a muscle under test are
classified based on their locations relative to these
limits. For example, Stalberg used a mean norma-
tive range of ±2 standard deviations and considered
three different outlier criteria: 95% confidence lim-
its; 95% confidence limits for the third highest and
third smallest values; and extreme upper and lower
outlier limits. No appreciable difference was found
across these methods. A further strategy based on
this approach applies rules and heuristics13,19 to in-
terpret the numbers of outliers and deviations from
normal mean limits. Several works by Podnar52-54
have performed comparisons of different outlier
criteria and have established standardized norma-
tive limits for diagnostic criteria.
The intention of QEMG analysis is to increase
both sensitivity (number of correct positive deci-
sions) and specificity (number of correct negative
decisions). For every decision, the objective is to
maximize sensitivity and specificity. However, sen-
sitivity and specificity cannot be simultaneously
maximized. Therefore, with regard to the deci-
sion-making process, whether it is better to be less
specific in order to be more sensitive must first be
decided. Best practice is to determine the minimum
number of overall errors by balancing sensitivity and
specificity unless the cost of poor sensitivity relative
to specificity or vice versa can be determined. In this
regard, statistical muscle characterization methods
have two main implementation issues. First, it is dif-
ficult to suitably define thresholds of normality/ab-
normality. Second, much confusion exists in deter-
mining suitable categorization rules. For example, a
rule might state that a minimum number of outliers
must be present to declare abnormality, but results
may vary depending on which features happen to
have outliers. On the other hand, a single feature
might have sufficient evidence due to the number
of outlying MUPs. Such a feature may or may not
fail the “mean” test depending on the distribution of
its MUP feature values, but based on outlier criteria
may clearly demonstrate abnormality.
Volume 38, Number 5, 2010
475
2. Classifier-Based Muscle Characterization
As an analog to MUP characterization, classifica-
tion techniques have been applied to features that
represent the entire muscle. For example, Xie et
al.55 developed a multi-domain fusion strategy that
incorporates mean MUP morphological, frequen-
cy domain, and wavelet transform-based features.
Information from these domains is combined us-
ing fuzzy integral theory. Komur and Dobrowolski
applied Fourier spectral analysis56,57 and classified
wavelet coefficients from MUP scalograms.58 The
former has the advantage of providing a single diag-
nostic parameter derived from the power spectrum
of the signal and does not require the onset and end
locations of the MUPs to be accurately determined.
The latter method uses linear discriminant analysis
to reduce the number of wavelet based features to
two, transforming the decision space into a linear
classification problem.
3. Probabilistic Muscle Characterization
Statistical muscle characterizations have two main
weaknesses: i) they are unable to provide a mea-
sure of the quality of or confidence in a particular
inference, and ii) they use thresholds defined us-
ing averages and deviations across sets of MUPs
(20 or more), making it difficult to investigate the
rationale for a particular result at the level of the
individual MUPs. Both of these weaknesses reduce
the transparency of decisions made. A probabilis-
tic approach addresses these weaknesses by pro-
viding information at the muscle level as well as
the MUP level, thus allowing for the provision of
a detailed explanation of the underlying decision-
making process. Recent work has applied pattern
recognition techniques to the probabilistic analysis
of MUP features, and in some cases to features of
composite EMG signals.11, 13,19,20,23,34-37,40,58,59 The
term “characterization” in this sense takes on a more
complex definition, but with the same end result as
described for statistical characterization. A proba-
bilistic characterization is one that assigns a score
or likelihood measure to each muscle category un-
der consideration. Ideally, the score represents the
476
probability of the examined muscle being affected
by a disease of a particular category conditioned by
the specific characteristics of the set of its sampled
MUPs. A characterization is a set of n scores, where
n is number of muscle categories under consider-
ation (typically 2 or 3).
a. Aggregating MUP Characterizations
Despite the fact that MUPs can be characterized
with high accuracy, there is a high degree of vari-
ability across MUP characterizations, even for
MUPs detected in the same normal muscle, there-
by limiting their diagnostic potential. However,
a robust muscle characterization can be obtained
by aggregating information across a set of MUPs
detected from a muscle under test. An overall
muscle characterization is achieved by aggregating
the characterizations from individual MUPs. Just
as in the MUP case, a score is produced for each
category under consideration. The muscle is then
categorized as being from the category that has
the highest conditional probability score. A muscle
conditional probability can also be thought of as
the confidence in a particular characterization.
Pfeiffer and Kunze11,12 first introduced the idea
of probabilistic characterization by using Bayes’ rule
for multiple pieces of evidence to aggregate MUPs
that were characterized using Fisher’s linear dis-
criminant analysis. Hamilton-Wright and Stashuk,
using MUP characterizations created by a pattern
discovery-based method modified to include fuzzy
inference, aggregated MUP characterizations using
a fuzzy logic-based centroid calculation weighted
by the confidence of each MUP characterization.38
Pino14,19 considered several aggregation metrics,
including arithmetic mean, Bayes’ rule, and the z-
transform.60 These metrics were used to aggregate
scores from several classifier methods (naive Bayes,
linear discriminant analysis, pattern discovery, deci-
sion trees). All methods seem to provide a high level
of accuracy, with Bayes’ rule-based muscle charac-
terization performing best. The muscle character-
ization scores generated using Bayes’ rule tend to
saturate to 0 or 1 as more evidence (higher number
of diseased MUPs) is presented, which is desirable
in most circumstances. However, when the classifier
Farkas et al
is uncertain of the outcome, this should be clearly
reflected in the characterization score, rather than by
saturation in an incorrect category. The probabilistic
methodology has been compared to the statistical
techniques of the previous section on several occa-
sions13,19,20,23 and has shown much promise.
b. Measures of Confidence and Involvement
A muscle characterization score should reflect the
probability that the muscle from which the MUPs
were detected is actually of the given category, condi-
tioned on the evidence provided by the set of MUP
characterizations. In this way, a muscle character-
ization measure can be thought of as the confidence
in making a particular categorization based on the
available evidence. A well-calibrated muscle confi-
dence score of 80% for a given category means that,
out of all the muscles that are assigned that score,
80% are truly of that category. It is therefore useful
to calibrate muscle scores to reflect true conditional
probabilities. One such calibration technique13 uses
Monte Carlo sampling39 to generate many variations
of the training data. These instances are then used
to calculate the actual probabilities associated with
each score value. Curves fitted to these points can
be used to transform subsequent test muscle scores
into well-calibrated conditional probabilities.
When the arithmetic mean is used to aggregate
MUP conditional probabilities, the conditional
probabilities representing a muscle characterization
correlate well with the level of involvement of a dis-
ease. In particular, the arithmetic mean aggregation
method had a correlation of 0.9 using the Spearman’s
rank test.23 Although correlation is high, much dif-
ficulty lies in predicting the level of involvement due
to the fact that at lower levels of disease involvement,
the confidence in making a correct characterization
is also low, resulting in greater variability and lower
accuracy in the measurement.
II.C. Accuracy of Characterization Methods
The metrics used to evaluate MUP or muscle char-
acterization performance are greatly dependent on
the specific objectives and desired outcome of the
techniques in question. This makes it difficult to
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
consistently compare the individual techniques. For
example, Katsis has done a considerable amount
of work to reduce the error in detecting diseased
MUPs, and achieved overall accuracy values of 86%
to 88% and a disease-specific sensitivity of 85% to
99%.36,37 However, the MUPs used in these stud-
ies were characterized and annotated by an expert,
thus focusing mainly on pattern recognition aspects
of the problem.
In contrast to this approach, probabilistic
methods use the true label of the muscle that the
MUPs were extracted from as the class label for
classifier training and performance evaluation. The
difficulty of this approach is that there is a great
deal of ambiguity in the interpretation of MUPs
detected from a muscle, because MUPs represent-
ing each of the clinical disease states (normal,
myopathic, neurogenic) can be detected from any
muscle regardless of its condition.11 This ambiguity
is due to several factors such as the variability of
needle position and muscle structure or composi-
tion. As such, single MUP categorization accuracy
is rather poor by comparison and in the range of
60% to 70%40; therefore, discussing the accuracy of
MUP categorization in this manner will produce
poor results compared with the methods men-
tioned above, but MUP categorization is not the
main goal. Rather, it is the distribution of MUPs
from each of the clinical states being studied that
provides valuable information, in a probabilistic
sense, about the likelihood of a disease, its level
of involvement, and the confidence in this assess-
ment. Muscle characterization methods must con-
sider the confusion created by MUPs that have
a high likelihood of belonging to more than one
disease state, and it is this information, in contrast
to highly confident MUP characterizations, that
makes it possible to properly categorize a muscle.
Evaluating the performance of muscle character-
ization techniques must therefore be done at the
muscle level after combining the individual MUP
characterizations, and their performance must be
compared against other techniques designed to
achieve a similar goal (i.e., statistical muscle char-
acterization, EMG signal characterization, etc.).
Numerous studies have shown that the accuracy
Volume 38, Number 5, 2010
477
of muscle categorization when using probabilistic
methods is significantly better than the accuracy of
competing strategies. For example, the probabilistic
method achieved accuracies of 80% to 95% com-
pared with 76% to 86% using statistical methods.19
Other studies have demonstrated accuracies of 59%
(37% improvement over conventional) using linear
discriminant analysis and Bayes’ rule aggregation,59
and 95% to 98% using advanced pattern recogni-
tion techniques,33 while in one case only a slight
improvement was observed (60% vs. 66%).13
Studies by Pino and others13,19,23,40 have shown
that several commonly used classification tech-
niques perform comparably with regard to MUP
categorization and do not impact the overall muscle
categorization outcome significantly. However, they
do affect the conditional probability estimates that
make up the muscle characterization vector. The low
accuracies obtained are mainly a result of the over-
lap of MUP feature value distributions mentioned
previously, rather than of poor classifier design or
feature selection. By aggregating MUP character-
izations, an overall decision can be made by taking
into account evidence that is representative of the
entire muscle. Depending on the aggregation strat-
egy used, such methods can either be conservative,
focusing on the level of agreement between MUP
categorizations, or extremist by weighting outlier
MUPs (MUPs with a high likelihood for a given
disease category) more heavily.
In either case, a decision based on multiple
pieces of evidence is shown to be superior in terms
of muscle characterization performance, despite
its low classifier performance at the MUP level.
Therefore, it is not so much the predicted category
(categorization) assigned to an MUP that is rel-
evant, but rather the MUP characterization vec-
tor that describes the probabilities of a particular
disease conditioned on the features of the detected
MUP. The aggregation of several of these MUP
characterizations provides a robust representation
of the clinical state of the muscle. Furthermore,
the method of MUP aggregation used to obtain
a muscle characterization, combined with a trans-
parent MUP characterization technique, forms the
basis for clinical decision support.
478
III. CLINICAL DECISION SUPPORT
Numerous methods of analyzing and interpreting
data to provide decision support have been report-
ed, particularly in the medical field, in which large
amounts of disparate data are analyzed and amal-
gamated to make critical and time-sensitive deci-
sions. One of the most important aspects of a clini-
cal decision support system is its ability to combine
two important (and sometimes conflicting) capa-
bilities: the ability to characterize consistently and
accurately make a prediction and the ability to
explain that prediction in a manner that is easily
understood by a decision maker. The statistical and
probabilistic characterization techniques discussed
in the previous section can be used in a decision
support framework to present the vast amounts of
information obtained from QEMG analysis in a
clinically useful and relevant fashion.
The prevalence of clinical decision support
systems in medicine is growing exponentially,
and systems have been proposed both in the field
of neuromuscular disease18,19,61 and in diagnostic
medicine in general.61-65 For example, the MUNIN
system66 uses Bayesian networks to infer a disease
based on findings from clinical tests, examinations,
and patient reports. A major limitation of this
system is that human medical expertise is required
to “train” the system by defining conditional prob-
ability tables. A machine that can “learn” without
a great deal of user intervention would be an asset.
The KANDID system61 uses logical operators and
known rules to build a knowledge base. The system
is used by making queries that are answered using
first-order logic. The statistical muscle characteriza-
tion techniques presented earlier18 can be used in a
decision support framework to indicate the pres-
ence or absence of disease. By incorporating a set
of known rules, the disease category (neurogenic or
myopathic) can also be specified.
In most clinical decision support systems, data
from multiple sources, such as clinical evaluations,
imaging, and other medical tests, are used to achieve
a more robust inference. For example, miniTUBA67
is a medical inference system based on dynamic
Bayesian networks. Dynamic Bayesian networks
Farkas et al
are able to interpret heterogeneous, fluctuating
data and are able to capture time-varying clinical
parameters as well as predict the course of disease
progression. Revett et al.62 developed a system for
breast cancer diagnosis by combining “rough sets”
and neural networks, and Gevaert et al.63 developed
a breast cancer prediction scheme based on Bayes-
ian networks that combines high-dimensional
microarray data with clinical findings. In the field
of pulmonary diseases, Economou et al.64 initially
used artificial neural networks in a decision support
system and then developed a knowledge base mod-
eled after the methodology used by physicians for
clinical differential diagnosis to provide decision
support.65
III.A. Requirements for Clinical Decision
Support
Attempts have been made to formalize the re-
quirements of a clinical decision support system.
Such requirements are intended to guide the de-
velopment and evaluation of future systems. The
works of Konokenko and Sprogar68,69 and a later
summary by Pino14 have generated a list of require-
ments for clinical decision support systems to be
clinically useful and safe. These are: i) transparency,
ii) accuracy, iii) a measure of confidence in deci-
sion support provided, iv) a numeric or continuous
characterization value or measure, v) the ability to
handle mixed-mode and multivariate feature data;
vi) the ability to generalize to new data, and vii) the
ability to handle missing features or data. In addi-
tion, it is desirable to have a reasoning strategy and
flow of information that, even in an abstract way, is
analogous to the cognitive processes and data as-
similation strategies employed by a human decision
maker. This helps to establish trust in the system,
and provides a “consultative” approach rather than
a simple statement of the facts.
III.B. Suitable Methods for QEMG-Based
Clinical Decision Support
The evaluation by Pino showed that the statisti-
cal methods (i.e., the means and outlier analysis of
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
Stewart and Stalberg19,44,45) perform as well as the
classifier-based and probabilistic methods.13,18 The
statistical methods are also quite intuitive in terms
of explaining the decision support provided, mak-
ing them useful for presentation purposes. How-
ever, the statistical methods are not well suited for
clinical decision support. They are not tractable to
larger feature sets and are not able to provide a con-
tinuous measure of confidence or degree of involve-
ment. Similar arguments can also be made against
most of the continuous classifier methods as well.
They may be quite robust, but they lack the needed
transparency. In contrast, the probabilistic methods
meet the requirements for clinical decision support
presented above, particularly when implemented
using pattern discovery and/or the fuzzy inference
frameworks.
Regardless of the specific inference techniques
used, a fundamental component of decision sup-
port systems is the knowledge base or repository.
The quality of inference made by the system greatly
depends on the nature of the data being defined
as exemplary. Therefore, exemplary data must first
be acquired and stratified by an expert clinician.
Specifically, it must be acquired for each muscle or
groups of muscles studied, and should be stratified
by age or any other relevant stratification criteria
(e.g., gender). The physical architecture of a deci-
sion support system with respect to its knowledge
base can take on several forms. For example, a
stand-alone system with its own repository of
knowledge can provide a self-consistent defini-
tion of clinical disease states based on acquisition
protocols and studies performed at a particular
clinic. In this case, the knowledge base would be
built up by the users in that particular clinic, and
its use would be confined to that particular area
of expertise. Conversely, client-server architecture
could be used to centralize the knowledge base
and promote consistency of gold standard data, but
also to accommodate different detection protocols
from laboratory to laboratory. This type of frame-
work would support collaboration of experts across
clinics, would be useful for training students and
researchers, and could even support the diagnosis
of patients in remote areas.
Volume 38, Number 5, 2010
479
III.C. Examples of Clinical Decision Support
One advantage of the probabilistic muscle charac-
terization frameworks proposed by Pinot14 is the
ability to appreciate how each MUP template con-
tributes to a muscle characterization and to explore
the rationale for this contribution. These methods
allow a decision maker to “drill down” from high
levels of abstraction to detailed explanations.
Figure 2 is a visualization displaying MUP
characterizations obtained using a combined pat-
tern discovery/fuzzy inference methodology38
having several decision support components. At
the top left of the display is the suggested muscle
categorization (myopathic), and the confidence in
and conflict with this suggestion are presented at
the top right. Below this are the individual MUP
characterizations calculated from the EMG sig-
nals detected in the muscle under study. Each
component of each MUP characterization is plot-
ted (point with tail) in a 2D plot representing the
assertion and confidence of the MUP character-
ization for its muscle category (myopathic, normal,
and neurogenic). Points with multiple tails repre-
sent multiple MUP characterization components
with the same assertion and confidence values. The
muscle characterization scores are calculated using
a centroid calculation of MUP assertions weighted
by their confidence measures, and are shown as a
colored circle in each respective muscle category
plot. The suggested muscle category corresponds
to the largest muscle characterization score. In
the bottom-left corner, the MUP template and its
respective features of a selected MUP character-
ization are presented. Next to this, the assertion
and confidence of the selected MUP characteriza-
tion are presented, along with the main patterns of
feature events used for their calculation. By select-
ing different MUP characterizations, a decision
maker is able to “drill down” into the details of the
presented muscle characterizations. In addition,
due to the fact that all of the MUP characteriza-
tions are available, it is possible to show how the
data associated with each MUPT relate to all of
the other MUPTs sampled from the muscle be-
ing examined. The collective display of the set of
480
FIGURE 2. Fuzzy inference-based muscle characterization.
MUP characterizations helps a decision maker to
visualize the resultant muscle characterizations
as well as its component MUP characterizations.
This in turn allows a decision maker to appreciate
the basis of, and confidence in, a suggested muscle
categorization.
Figures 3 and 4 provide visualizations related
to MUP characterizations obtained using pattern
discovery-based estimates of conditional probabil-
ity. Figure 3 presents individual MUP character-
izations as small pie charts, each of which is shown
to contribute to the overall muscle characterization
and level of involvement shown on the right. The
muscle characterization scores and involvement
percentages (using Bayes’ rule and autoregressive,
respectively) for each category are represented by
large pie charts. Figure 4 shows the feature events
that underlie a single MUP characterization. The
plots on the bottom left provide a visual interpre-
tation of the rules generated by pattern discovery.
Each shape represents a discrete feature value, low
(L), medium (M) or high (H), and shapes stacked
along the same vertical line represent feature val-
Farkas et al
ues that occurred together as an nth-order pattern.
Patterns for or against each category are displayed
with their x-position reflecting the level of asser-
tion or refutation provided by the respective rule.
The solid red bars represent the combined asser-
tion across all rules for a category. In this case, the
evidence suggests that the muscle from which the
MUPT was detected is likely myopathic, but also
possesses normal and neurogenic characteristics.
This confusion is driven by the presence of rules
that provide evidence both for and against the
myopathic and neurogenic categories. Generally,
higher order patterns provide a higher weight of
evidence and thus form the basis of the decision, in
this case showing strong support for the myopathic
category and strong refutation of the neurogenic
category. A second-order rule is also present in
both the myopathic and neurogenic cases, provid-
ing contrary evidence to the higher-order rule, but
to a much lesser degree. The amount of evidence
is less convincing for the normal category (shown
by lower assertion values), but the presence of two
separate rules bring the overall assertion farther to
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG 481

FIGURE 3. Probabilistic muscle characterization. The small pie charts on the left represent individual
MUP characterizations, the large pie chart at the top right represents muscle characterization, and the
large pie chart at the bottom right represents the estimated level of involvement.

the left. The overall characterization is still myo-
pathic, but not with a very high level of confidence
given the conflicting evidence. In Figure 4, in the
right column are plots of the distributions of the
values for each feature. The three bins used for
quantization (L, M, H) are demarcated, and green
and red bars represent the normal and abnormal
ranges for each feature value. The white bar and
plot represent the feature value and distribu-
tion of feature values for the MUP template and
MUP template set, respectively. Figures 3 and 4
again demonstrate the ability of the probabilistic
methods to provide a quantitative overall muscle
characterization complete with a measure of con-
fidence and the ability to see the distribution of
MUP characterizations, as well as the ability to
“drill down” into the details of the individual MUP
characterizations. These examples demonstrate the
advantages of using transparent decision support
systems for QEMG.
IV. SUMMARY
The concepts and methods applied to quantitatively
use clinically detected needle EMG signals to pro-
vide decision support related to the diagnosis, treat-

Volume 38, Number 5, 2010

482
FIGURE 4. Probabilistic muscle characterization. Details of a specific MUP characterization with support
for the myopathic category are shown.
ment, and management of neuromuscular disorders
have been reviewed. The fundamental basis for the
clinically relevant information present in needle
EMG signals has been described. Methods for the
quantitative analysis (characterization) of individu-
al MUPs with the objective of providing evidence
regarding the clinical state of the muscle under
examination have been described and compared.
Methods for aggregating information from a set of
MUPs detected in the muscle under examination
to obtain a characterization of the clinical state of
the muscle have also been described and compared.
Finally, the requirements of a clinical decision sup-
port system have been presented and the suitability
of the various muscle characterization methods for
clinical decision support has been compared and
examples of decision support methods shown. This
review demonstrates the progress made toward the
Farkas et al
development of more clinically useful methods of
quantitative analysis of needle EMG signals.
In summary, QEMG assessment in general
provides several benefits in terms of analytical
strength and reproducibility. Many of the benefits
of QEMG methods stem from the underlying fact
that once a reproducible system of measurement is
constructed, integration of new data into that sys-
tem is straightforward.
QEMG methods provide a starting point for
a discussion of which features individually may be
used to identify and characterize EMG signals, as
well as the utility of feature combinations.
The variety of tools available to support infer-
ence based on QEMG data is large and growing.
The subtlety of inference that QEMG analysis cur-
rently supports is already quite impressive, and can
be expected to improve in the future.
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
REFERENCES
1. Stalberg EV, Falck B. The role of electromyog-
raphy in neurology. Electroencephalogr Clin
Neurophysiol. 1997;103(6):579–98.
2. Gazzoni M, Farina D, Merletti R. A new
method for the extraction and classification
of single motor unit action potentials from
surface EMG signals. J Neurosci Methods.
2004;136(2):165–77.
3. Merletti R, Farina D. Analysis of intramuscu-
lar electromyogram signals. Philos Transact A
Math Phys Eng Sci. 2009;367(1887):357–68.
4. Holobar A, Minetto MA, Botter A, Negro F,
Farina D. Experimental analysis of accuracy
in the identification of motor unit spike trains
from high-density surface EMG. IEEE Trans
Neural Syst Rehabil Eng. 2010;18(3):221–9.
5. Stashuk D, Brown W. Quantitative electro-
myography. In: Brown WF, Bolteon C, Ami-
noff M, editors. Neuromuscular function and
disease: basic, clinical, and electrodiagnostic
aspects. Philadelphia: WB Saunders; 2002. p.
311–48.
6. Stashuk DW. Decomposition and quantitative
analysis of clinical electromyographic signals.
Med Eng Phys. 1999; 21(6-7):389–404.
7. Stashuk D. EMG signal decomposition: how
can it be accomplished and used? J Electro-
myogr Kinesiol. 2001;11(3):151–73.
8. Doherty TJ, Stashuk DW. Decomposition-
based quantitative electromyography: Meth-
ods and initial normative data in five muscles.
Muscle Nerve. 2003;28(2):204–11.
9. McGill KC, Lau K, Dorfman LJ. A compari-
son of turns analysis and motor unit analysis
in electromyography. Electroencephalogr Clin
Neurophysiol. 1991;81(1):8–17.
10. McGill KC, Lateva ZC, Marateb HR.
EMGLAB: An interactive EMG decom-
position program. J Neurosci Methods
2005;149(2):121–33.
11. Pfeiffer G, Kunze K. Discriminant classifi-
cation of motor unit potentials (MUPs) suc-
cessfully separates neurogenic and myopathic
conditions. A comparison of multi- and uni-
variate diagnostical algorithms for MUP anal-
ysis. Electroencephalogr Clin Neurophysiol.
1995; 97(5):191–207.
Volume 38, Number 5, 2010
483
12. Pfeiffer G. The diagnostic power of motor unit
potential analysis: An objective Bayesian ap-
proach. Muscle Nerve. 1999;22(5):584–91.
13. Pino L, Stashuk D, Podnar S. Bayesian char-
acterization of external anal sphincter mus-
cles using quantitative electromyography. Clin
Neurophysiol. 2008;119(10):2266–73.
14. Pino L. Neuromuscular clinical decision sup-
port using motor unit potentials characterized
by ‘pattern discovery’ [dissertation]. Waterloo,
Ontario, Canada: University of Waterloo; 2009.
15. Podnar S. Predictive values of motor unit po-
tential analysis in limb muscles. Clin Neuro-
physiol. 2009;120(5):937–40.
16. Podnar S, Mrkaić M. Predictive power of mo-
tor unit potential parameters in anal sphincter
electromyography. Muscle Nerve. 2002;26(3):
389–94.
17. Podnar S. Predictive values of the anal sphinc-
ter electromyography. Neurourol Urodyn.
2009;28(8):1034–5.
18. Stålberg E, Bischoff C, Falck B. Outliers, a way
to detect abnormality in quantitative EMG.
Muscle Nerve. 1994;17(4):392–9.
19. Pino L, Stashuk D, Boe S, Doherty T. Proba-
bilistic muscle characterization using QEMG:
Application to neuropathic muscle. Muscle
Nerve. 2010;41(1):18–31.
20. Hamilton-Wright A, McLean L, Stashuk DW,
Calder KM. Bayesian aggregation versus major-
ity vote in the characterization of non-specific
arm pain based on quantitative needle electro-
myography. J Neuroeng Rehabil. 2010;7:8.
21. Wijnberg ID, van Dam KG, Graaf-Roelfsema
E, Keizer HA, van Ginneken MM, Barneveld
A, Breda E, van der Kolk JH. (Over)training
effects on quantitative electromyography and
muscle enzyme activities in standardbred hors-
es. J Appl Physiol. 2008;105(6):1746–53.
22. Daube JR, Rubin DI. Needle electromyogra-
phy. Muscle Nerve. 2009;39(2):244–70.
23. Pino LJ, Stashuk DW. Using motor unit po-
tential characterizations to estimate neuro-
muscular disorder level of involvement. Conf
Proc IEEE Eng Med Biol Soc. 2008;2008:
4138–41.
24. Preston D, Shapiro B. Needle electromyogra-
phy fundamentals, normal and abnormal pat-
terns. Neurol Clin. 2002;20:361–96.
484
25. Sonoo M. New attempts to quantify concen-
tric needle electromyography. Muscle Nerve.
2002;25(S11):S98–102.
26. Zalewska E, Hausmanowa-Petrusewicz I.
The SIIR index--a non-linear combination of
waveform size and irregularity parameters for
classification of motor unit potentials. Clin
Neurophysiol. 2005;116(4):957–64.
27. Zalewska E, Hausmanowa-Petrusewicz I,
Stålberg E. Modeling studies on irregular
motor unit potentials. Clin Neurophysiol.
2004;115(3):543–56.
28. Stålberg E, Trontelj JV. The study of normal
and abnormal neuromuscular transmission
with single fibre electromyography. J Neurosci
Methods. 1997;74(2):145–54.
29. Parsaei H, Stashuk D W, Rasheed S, Farkas
C, Hamilton-Wright A. Intramuscular EMG
signal decomposition. Crit Rev Biomed Eng.
2010;435–65.
30. Parsaei H, Nezhad FJ, Stashuk D, Hamilton-
Wright. Development of an SVM classifier for
detecting merged motor unit potential trains.
Proceedings of the XVII th Congress of ISEK,
June 18 – 21, Niagara Falls, Canada; 2008.
31. Parsaei H, Nezhad F, Stashuk D, Hamilton-
Wright A. Validation of motor unit potential
trains using motor unit firing pattern infor-
mation. Conf Proc IEEE Eng Med Biol Soc.
2009;2009:974–7.
32. Parsaei H, Stashuk D. MUP shape-based vali-
dation of a motor unit potential train. Conf Proc
IEEE Eng Med Biol Soc. 2009;2009:2551–4.
33. Christodoulou C, Pattichis C. Unsupervised
pattern recognition for the classification of
EMG signals. IEEE Trans Biomed Eng.
1999;46(2):169–78.
34. Pattichis CS, Pattichis MS. Time-scale analy-
sis of motor unit action potentials. IEEE Trans
Biomed Eng. 1999;46(11):1320–9.
35. Pattichis CS, Elia AG. Autoregressive and
cepstral analyses of motor unit action poten-
tials. Med Eng Phys. 1999;21(6–7):405–19.
36. Katsis C, Goletsis Y, Likas A, Fotiadis D, Sar-
mas I. A novel method for automated EMG de-
composition and MUAP classification. Artificial
Intelligence in Medicine. 2006;37(1):55–64.
37. Katsis CD, Exarchos TP, Papaloukas C, Go-
letsis Y, Fotiadis DI, Sarmas I. A two-stage
Farkas et al
method for MUAP classification based on
EMG decomposition. Comput Biol Med.
2007;37(9):1232–40.
38. Hamilton-Wright A, Stashuk D. Clinical Char-
acterization of Electromyographic Data Using
Computational Tools. Conf. Proc. IEEE Com-
putational Intelligence and Bioinformatics and
Computational Biology (CIBCB 2006); 2006;
2006:1–7.
39. Duda RO, Hart PE, Stork DG. Pattern classi-
fication. 2nd ed. New York: Wiley-Interscience;
2000.
40. Pino L, Stashuk D, Boe S, Doherty T. Motor
unit potential characterization using “pattern
discovery”. Med. Eng Phys. 2008;30(5):563–73.
41. Schizas CN, Middleton LT, Pattichis CS. Neu-
ral network models in EMG diagnosis. IEEE
Trans Biomed Eng. 1990;42(5):486–96.
42. Quinlan JR. C4.5: Programs for machine learn-
ing. 1st ed. San Francisco: Morgan Kaufmann;
1992.
43. Wong AKC, Wang Y. High-order pattern dis-
covery from discrete-valued data. IEEE Trans.
Knowl. Data Eng. 1997;9(6):877-893.
44. Stewart CR, Nandedkar SD, Massey JM, Gil-
christ JM, Barkhaus PE, Sanders DB. Evalua-
tion of an automatic method of measuring fea-
tures of motor unit action potentials. Muscle
Nerve. 1989;12(2):141–8.
45. Stålberg E, Stålberg S, Melander M, Arimura
K. A personal computer based system used in
electromyography for interpretation and re-
porting. Comput Methods Programs Biomed.
1991;34(2-3):219–27.
46. Buchthal F, Pinell P, Rosenfalck P. Action po-
tential parameters in normal human muscle
and their physiological determinants. Acta
Physiol Scand. 1954;32(2–3):219–29.
47. Podnar S, Vodusek DB. Standardisation of
anal sphincter EMG: high and low threshold
motor units. Clin Neurophysiol. 1999;110(8):
1488–91.
48. Podnar S, Vodusek DB. Standardization of anal
sphincter electromyography: uniformity of the
muscle. Muscle Nerve. 2000;23(1):122–5.
49. Podnar S, Vodusek DB. Standardization of
anal sphincter electromyography: utility of mo-
tor unit potential parameters. Muscle Nerve.
2001;24(7):946–51.
Critical Reviews™ in Biomedical Engineering
A Review of Clinical QEMG
50. Podnar S, Mrkaić M. Size of motor unit po-
tential sample. Muscle Nerve. 2003;27(2):
196–201.
51. Podnar S. Usefulness of an increase in size of
motor unit potential sample. Clin Neurophysi-
ol. 2004;115(7):1683–8.
52. Podnar S. Criteria for neuropathic abnormality
in quantitative anal sphincter electromyogra-
phy. Muscle Nerve. 2004;30(5):596–601.
53. Podnar S. Comparison of different outlier crite-
ria in quantitative anal sphincter electromyogra-
phy. Clin Neurophysiol. 2005;116(8):1840–5.
54. Podnar S. Comparison of parametric and non-
parametric reference data in motor unit poten-
tial analysis. Muscle Nerve. 2008;38(5):1412–9.
55. Xie H, Huang H, Wang Z. Multiple feature
domains information fusion for computer-aid-
ed clinical electromyography. In: Gagalowicz
A, Philips W, editors. Computer analysis of im-
ages and patterns. Berlin/Heidelberg: Springer;
2005. p. 304–12.
56. Dobrowolski A, Tomczykiewicz K, Komur P.
Spectral analysis of motor unit action poten-
tials. IEEE Trans Biomed Eng. 2007;54(12):
2300–2.
57. Komur P, Dobrowolski AP, Dabrowski T, Tom-
czykiewicz K. Automated diagnostic method
supporting EMG examination. Conf Proc
IEEE Eng Med Biol Soc. 2008;2008:1116–9.
58. Dobrowolski AP, Jakubowski J, Tomczykiewicz
K. Linear discriminant analysis of MUAP sca-
lograms. Conf Proc IEEE Eng Med Biol Soc.
2008;2008:1100–3.
59. Pfeiffer G. The diagnostic power of motor unit
potential analysis: an objective bayesian ap-
proach. Muscle Nerve. 1999;22(5):584–91.
60. Whitlock MC. Combining probability from
independent tests: the weighted Z-method
is superior to Fisher’s approach. J Evol Biol.
2005;18(5):1368–73.
61. Fuglsang-Frederiksen A, Johnsen B, Vingtoft
S, Rønager J. Vingtoft S, Fuglsang-Frederiksen
Volume 38, Number 5, 2010
485
A, Rønager J, Petrera J, Stigsby B, Willison RG,
et al. KANDID an EMG decision support sys-
tem evaluated in a European multicenter trial.
Muscle Nerve. 1993;16(5):520–9.
62. Revett K, Gorunescu F, Gorunescu M, El-
Darzi E, Ene M. A hybrid breast cancer di-
agnosis system: a combined approach using
rough sets and probabilistic neural networks.
Conf. Proc. IEEE Int. Conf. on “Computer as
a tool”-Eurocon 2005, Belgrade, Serbia, 2005;
2005: 1124- 1127.
63. Gevaert O, De Smet F, Timmerman D,
Moreau Y, De Moor B. Predicting the prog-
nosis of breast cancer by integrating clinical
and microarray data with Bayesian networks.
Bioinformatics. 2006;22(14):e184–90.
64. Economou GP, Lymberopoulos D, Karavatsel-
ou E, Chassomeris C. A new concept toward
computer-aided medical diagnosis--a proto-
type implementation addressing pulmonary
diseases. IEEE Trans Inf Technol Biomed.
2001;5(1):55-66.
65. Economou G, Goumas P, Spiropoulos K. A
novel medical decision support system. J Com-
put Control Eng. 1996;7(4):177–83.
66. Suojanen M, Andreassen S, Olesen KG.
A method for diagnosing multiple diseas-
es in MUNIN. IEEE Trans Biomed Eng.
2001;48(5):522–32.
67. Xiang Z, Minter RM, Bi X, Woolf PJ, He Y.
miniTUBA: medical inference by network in-
tegration of temporal data using Bayesian anal-
ysis. Bioinformatics. 2007;23(18):2423–32.
68. Kononenko I. Machine learning for medical
diagnosis: history, state of the art and perspec-
tive. Artif Intell Med. 2001;23(1):89–109.
69. Šprogar M, Lenič M, Alayon S. Evolu-
tion in medical decision making. J Med Syst.
2002;26(5):479–89.