*5.2.2.2. Non-transparent classification techniques*

These methods can be used with careful positioning of the needle and during low level of muscle activation. Only one MUPT can be extracted from the EMG signal detected during muscle contraction. Therefore, for each MUPT to be extracted a separate contraction must be

EMG signals are the linear summation of the MUPTs created by the MUs active in a muscle. EMG signal decomposition extracts individual MUPTs from an EMG signal. Unlike level or window triggering, EMG signal decomposition allows several MUPTs created by MUs concurrently active during a single muscle contraction to be analyzed. The accuracy of the MUPTs extracted by an EMG signal decomposition algorithm determines the type of analyses that can be successfully applied to the extracted MUPTs. The MUPTs extracted during EMG signal decomposition can be further analyzed to assist in diagnosing neuromuscular disorders.

EMG signal decomposition involves three main steps, described in the following paragraphs.

The first step is to detect the MUPTs comprising an EMG signal. Some EMG signal decompo‐ sition algorithms attempt to detect all the MUPTs that existed in the EMG signal while others attempt to extract only MUPTs that had a major contribution to the EMG signal. The following step is to determine the shapes of the different MUPs. This can be done by categorizing the MUPs in the signal based on their shapes and sizes. This categorization, if implemented properly, reveals clusters of MUPs with similar shapes and sizes. As a result, MUPs with different shapes and sizes should belong to different clusters. MUPs with similar shapes and sizes were most probably created by different discharges of the same MU, while MUPs with unique shapes and sizes (i.e. not belonging to cluster or to a cluster with very few members) are most probably superpositions. The main outcome of this step is to identify the number of MUs that contributed significant MUPs to the EMG signal (i.e. to estimate the number of MUPTs with significant MUPs) and to estimate the MUP template of each discovered MUPT.

The second step is to determine the class of every template. Superpositions of MUPs are harder to deal with in the first step as well as in this step. If the overlap is only slight, the constituents might still be recognizable. But if the overlap is complete it might be necessary to try different alignments of the templates to see which gives the closest fit. The motor unit discharge patterns can also be used to help determine which MUs are involved in a superimposed MUP [36]. As discharge rates are assumed to be rather orderly (i.e. IDIs can be assumed to follow a Gaussian distribution), the time at which a particular discharge took place can be estimated from the

The final step in decomposition is to validate the results to ensure they are consistent with the expected physiological behavior of MUs. If there are unexpected short IDI in any of the discharge patterns, or if there are detected MUPs that have not been assigned to a MUPT, then the decomposition is probably not correct or incomplete. On the other hand, if all the activity in the signal (i.e. the detected MUPs) has been adequately accounted for by the set of extracted MUPTs which in turn represent MUs with physiologically realistic discharge patterns, then there is a good chance that the decomposition is substantially complete and accurate [36].

time at which the preceding or following discharge took place.

performed.

**2.** EMG signal decomposition

104 Electrodiagnosis in New Frontiers of Clinical Research

MUPT characterization can be performed using probabilistic techniques. Probabilistic techniques provide a MUP characterization in terms of conditional probabilities that sum to 1 across all of the categories considered. For instance, a probabilistic technique can suggest that considering the features of a MUPT there is a 10% probability it was detected in a myopathic muscle, a 70% probability it was detected in a normal muscle and a 20% probability it was detected in a neurogenic muscle neurogenic if three categories are considered.

Various methods have been used in the literature to perform MUP template characterization, ranging from conventional to advanced classifiers. For example, linear discriminant analysis (LDA), decision trees and a standard Naive Bayes (NB) classifier were implemented and compared in [9]. LDA attempts to find a linear combination of features that maximizes the between class variance and minimizes the within class variance and it relies on this as a basis for optimal classification. Using these trivial classifiers has the advantage of being rather more transparent than using more advanced pattern recognition techniques like neural networks and support vector machines.

Artificial neural networks were first used for MUP template characterization in [10] and [11]. More progress in this direction was achieved in [12] as artificial neural networks were used along with radial basis functions and probabilistic neural networks in a two-phase classifier, which increased MUPT characterization accuracy. In the second phase of the classification, a C4.5 decision tree was used to determine whether the disorder was myopathic or neurogenic, if any. Another example of using neural networks in MUP analysis can be found in [63].

In [53] autoregressive (AR) modeling and cepstral analysis were applied to characterize MUP templates and the training dataset was built on normal MUP templates as well as MUP templates taken from myopathic muscles. It was concluded in [53] that using AR modeling and cepstral analysis along with time domain features (in particular duration) led to catego‐ rizations with high accuracy in the assessment of myopathic MUP templates (in this work two categories were used; normal and myopathic). In [54], MUP templates were classified into three categories; normal, myopathic and neurogenic using support vector machines (SVM).

Using artificial neural networks in classification could lead to over-fitting; a classifier that has difficulty in producing the same accuracy with new or more generalized data. As mentioned earlier, using a SVM and artificial neural networks does not provide enough transparency and renders it more difficult for clinicians to understand how a certain classification decision was made.

#### *5.2.2.3. Transparent rule-based MUPT classification techniques*

An example of a transparent rule-based classification technique is the two-stage classifier developed in [55]. This two-stage classifier is based on utilizing radial basis function artificial neural networks and decision trees. The combined use of an artificial neural network and a decision tree reduces the number of tuned parameters required and allows an interpretation of the classification decisions to be provided [55].

Techniques based on pattern discovery (PD) represent another example of transparent rulebased techniques used for MUPT characterization. Pattern discovery is an information theory based technique established to detect significant patterns in data and then to use these patterns for classification [5]. PD was first introduced by Wong and Wang [57]-[60]. PD is applied on discrete data and, as a result, a quantization step is required for each feature that has contin‐ uous values, which is the case with most MUPT features. The number of discretization bins can be identified according to the nature of the problem at hand and the dataset used. For instance, if the number of bins is three; low, medium and high for each feature, there might be some loss of accuracy resulting from placing "very high" and "slightly high" values in the same bin. On the other hand, using five bins; very low, low, medium, high and very high can solve such a problem but more training examples are needed to keep the same number of expected occurrences per bin. In the PD classification algorithm, the first step is to discover the "significant" patterns; patterns that are repeated more often than expected assuming a random occurrence [56]. Rules are composed of patterns that include a muscle category. The order of a rule is equal to the number of MUPT features plus the muscle category. For example, high amplitude values in neurogenic MUPTs are a 2nd order rule [5]. Each rule has an associated weight of evidence (WOE) that denotes how much evidence the rule holds in support of a certain category [9]. For rule selection during testing, the highest order rule for each category is selected first. WOEs of selected rules are added to be normalized and this process continues until there are no more rules or all features have already been included in the previously selected rules [56]. In addition to the well-known pros and cons of discretization, characterizations performed by PD are transparent as the technique is rule-based which makes it feasible to explain to a clinician the rationale behind the classification decisions. However, when PD is used, there is a decrease in accuracy due to the discretization performed on the continuous MUPT feature values.

Stalberg identified the outlier range to be outside of the mean ± 2 × standard deviations range. This way, a categorization of the muscle being examined can be obtained. As an example for standard feature values, MUP template values documented in [62] are still considered

Clinical Quantitative Electromyography http://dx.doi.org/10.5772/56033 107

The above muscle categorization techniques do not provide any measure of confidence. Probabilistic muscle categorization, on the other hand, addresses this weakness as it provides probabilities describing MUPT characterizations as well as the overall muscle characterization. More formally, a probabilistic MUPT characterization technique assigns a likelihood measure to each MUPT category under consideration. For each MUPT characterization, a set of n likelihood measures is obtained, where n is number of muscle categories under consideration

Characterizations of MUPTs must be aggregated to obtain the overall muscle characterization of the muscle from which these MUPTs were detected. As with MUPTs, a set of n muscle likelihood measures is obtained, where n is the number of muscle categories and the muscle is considered to belong to the category that has the highest category likelihood value. Muscle likelihood values can be considered confidence measures in a particular characterization. In [64], [65], the idea of implementing probabilistic characterization and aggregating MUP template characterizations using Bayes' rule was first introduced. MUP template characteri‐ zation was performed using Fisher's LDA. Other techniques used Bayes' rule to aggregate MUP template likelihood measures obtained from multiple classifiers like decision trees, LDA

A muscle characterization likelihood value (measure) indicates the probability that the muscle, from which the characterized MUPTs were detected, actually belongs to the given category, conditioned on the evidence provided by the set of MUPT characterizations. Thus, a muscle characterization likelihood measure can be considered a measure of confidence in making a particular categorization based on the available evidence. For instance, a muscle confidence score of 75% for a given category means that, out of all the muscles that are assigned that score,

Another relevant concept in the context of muscle categorization is that of the level of involve‐ ment (LOI). When the values of the arithmetic mean of MUPT features are used to aggregate MUPT probabilities, the conditional probabilities resulting from a muscle characterization

It is not easy to predict LOI due to the fact that confidence in making a correct muscle catego‐ rization decision at lower levels of disease involvement is low, which results in greater

technique correlate well with the level of involvement (LOI) of a disease [66].

standard values for MUP template duration, amplitude and shape.

(2 or 3) [4]. For each MUPT, this set of likelihood measures should sum to 1.

*5.2.3.1. Aggregation of MUPT characterizations*

*5.2.3.2. Measures of confidence and involvement*

75% of such muscles actually belong to that category.

variability and lower accuracy in the LOI measurement [4].

and Naive Bayes [56].

In [61], a fuzzy inference system was introduced. This system is based on using PD in combi‐ nation with fuzzy logic theory to yield a hybrid system. The idea was to reduce quantization error via assigning memberships for every MUPT feature value based on their position within their assigned bin. The fuzzy membership values allow the same MUPT feature value to be considered in more than one rule simultaneously. Using this technique for establishing rules is similar to the method by which humans manually interpret certain data values and then attempt to classify these values, which makes it very useful, at least in terms of transparency, in a clinical decision support system.

### *5.2.3. Muscle categorization*

The ultimate purpose of characterizing MUPTs is to characterize the muscle from which they were detected. The statistical method for muscle characterization can be performed by calculating mean values for sampled MUPT features and comparing them to expected normative mean values (Note: comparisons should be standardized with respect to age, gender, muscle, EMG detection technique, etc.). Using the expected normative mean and outlier threshold values the overall category of a muscle is then determined based on the mean values of the sampled MUPT feature values with respect to these thresholds. For instance, Stalberg identified the outlier range to be outside of the mean ± 2 × standard deviations range. This way, a categorization of the muscle being examined can be obtained. As an example for standard feature values, MUP template values documented in [62] are still considered standard values for MUP template duration, amplitude and shape.

The above muscle categorization techniques do not provide any measure of confidence. Probabilistic muscle categorization, on the other hand, addresses this weakness as it provides probabilities describing MUPT characterizations as well as the overall muscle characterization. More formally, a probabilistic MUPT characterization technique assigns a likelihood measure to each MUPT category under consideration. For each MUPT characterization, a set of n likelihood measures is obtained, where n is number of muscle categories under consideration (2 or 3) [4]. For each MUPT, this set of likelihood measures should sum to 1.

### *5.2.3.1. Aggregation of MUPT characterizations*

Techniques based on pattern discovery (PD) represent another example of transparent rulebased techniques used for MUPT characterization. Pattern discovery is an information theory based technique established to detect significant patterns in data and then to use these patterns for classification [5]. PD was first introduced by Wong and Wang [57]-[60]. PD is applied on discrete data and, as a result, a quantization step is required for each feature that has contin‐ uous values, which is the case with most MUPT features. The number of discretization bins can be identified according to the nature of the problem at hand and the dataset used. For instance, if the number of bins is three; low, medium and high for each feature, there might be some loss of accuracy resulting from placing "very high" and "slightly high" values in the same bin. On the other hand, using five bins; very low, low, medium, high and very high can solve such a problem but more training examples are needed to keep the same number of expected occurrences per bin. In the PD classification algorithm, the first step is to discover the "significant" patterns; patterns that are repeated more often than expected assuming a random occurrence [56]. Rules are composed of patterns that include a muscle category. The order of a rule is equal to the number of MUPT features plus the muscle category. For example, high amplitude values in neurogenic MUPTs are a 2nd order rule [5]. Each rule has an associated weight of evidence (WOE) that denotes how much evidence the rule holds in support of a certain category [9]. For rule selection during testing, the highest order rule for each category is selected first. WOEs of selected rules are added to be normalized and this process continues until there are no more rules or all features have already been included in the previously selected rules [56]. In addition to the well-known pros and cons of discretization, characterizations performed by PD are transparent as the technique is rule-based which makes it feasible to explain to a clinician the rationale behind the classification decisions. However, when PD is used, there is a decrease in accuracy due to the discretization performed on the

In [61], a fuzzy inference system was introduced. This system is based on using PD in combi‐ nation with fuzzy logic theory to yield a hybrid system. The idea was to reduce quantization error via assigning memberships for every MUPT feature value based on their position within their assigned bin. The fuzzy membership values allow the same MUPT feature value to be considered in more than one rule simultaneously. Using this technique for establishing rules is similar to the method by which humans manually interpret certain data values and then attempt to classify these values, which makes it very useful, at least in terms of transparency,

The ultimate purpose of characterizing MUPTs is to characterize the muscle from which they were detected. The statistical method for muscle characterization can be performed by calculating mean values for sampled MUPT features and comparing them to expected normative mean values (Note: comparisons should be standardized with respect to age, gender, muscle, EMG detection technique, etc.). Using the expected normative mean and outlier threshold values the overall category of a muscle is then determined based on the mean values of the sampled MUPT feature values with respect to these thresholds. For instance,

continuous MUPT feature values.

106 Electrodiagnosis in New Frontiers of Clinical Research

in a clinical decision support system.

*5.2.3. Muscle categorization*

Characterizations of MUPTs must be aggregated to obtain the overall muscle characterization of the muscle from which these MUPTs were detected. As with MUPTs, a set of n muscle likelihood measures is obtained, where n is the number of muscle categories and the muscle is considered to belong to the category that has the highest category likelihood value. Muscle likelihood values can be considered confidence measures in a particular characterization. In [64], [65], the idea of implementing probabilistic characterization and aggregating MUP template characterizations using Bayes' rule was first introduced. MUP template characteri‐ zation was performed using Fisher's LDA. Other techniques used Bayes' rule to aggregate MUP template likelihood measures obtained from multiple classifiers like decision trees, LDA and Naive Bayes [56].

#### *5.2.3.2. Measures of confidence and involvement*

A muscle characterization likelihood value (measure) indicates the probability that the muscle, from which the characterized MUPTs were detected, actually belongs to the given category, conditioned on the evidence provided by the set of MUPT characterizations. Thus, a muscle characterization likelihood measure can be considered a measure of confidence in making a particular categorization based on the available evidence. For instance, a muscle confidence score of 75% for a given category means that, out of all the muscles that are assigned that score, 75% of such muscles actually belong to that category.

Another relevant concept in the context of muscle categorization is that of the level of involve‐ ment (LOI). When the values of the arithmetic mean of MUPT features are used to aggregate MUPT probabilities, the conditional probabilities resulting from a muscle characterization technique correlate well with the level of involvement (LOI) of a disease [66].

It is not easy to predict LOI due to the fact that confidence in making a correct muscle catego‐ rization decision at lower levels of disease involvement is low, which results in greater variability and lower accuracy in the LOI measurement [4].
