Motor unit potential characterization using “pattern discovery”
Introduction
Analysis of motor unit potentials (MUPs) helps clinicians infer if a neuromuscular disorder is present and if it is a result of pathology affecting motor neurons/motor axons (i.e. neuropathic) or related to muscle fiber injury or atrophy (i.e. myopathic). The status quo approach for an electromyographer involves using qualitative auditory and visual analysis of intramuscular needle-detected electromyographic (EMG) signals from voluntarily activated muscles. Qualitative analysis is often prone to error and limited by subjective misinterpretation in part because a large amount of potentially ambiguous information needs to be extracted and analyzed [1], [2]. A recent study authored by Kendall et al. showed that faculty and residents (blind to the underlying diagnosis) using video recorded needle based examinations for radiculopathy had an overall 46.9% agreement with the actual diagnosis [3]. In addition, qualitative methods are less able than quantitative methods to provide effective longitudinal comparisons. An electrophysiological characterization that is sensitive to small changes caused by a neuromuscular disease would allow measurement of the degree of the disease involvement and therefore allow evaluation of treatment effectiveness [4].
For an electrophysiological characterization of a neuromuscular system to be sufficiently sensitive, EMG signals must be quantitatively analyzed. One class of quantitative electromyography (QEMG) methods attempts to characterize the individual motor units of a muscle by analyzing isolated MUPs. Usually, an individual motor unit is represented by a MUP template calculated from a train of MUPs generated by the motor unit. The characterization methods discussed in this work can use MUP templates based on any of these QEMG methods (e.g. level or window triggering, or more sophisticated decomposition techniques). Henceforth, the term MUP will be used to refer to the calculated MUP template unless otherwise mentioned. As such, the focus of this study is on characterization of MUPs as normal, myopathic or neuropathic. Characterization is a clinical term referring to the discernment, description or attribution of distinguishing traits [5]. The degree of involvement of a disease process can be determined if each MUP in a set of MUPs sampled from a muscle is objectively characterized using a numeric value on a continuous scale that can be effectively combined into a neuromuscular characterization. More specifically, characterization of a single MUP using three numeric values reflecting the probability of it being detected from a normal, myopathic or neuropathic muscle is an initial step towards useful, robust neuromuscular characterization. This work describes and evaluates characterization processes that provide quantitative interpretation of information commonly extracted from individual MUPs during a QEMG examination.
The literature describes a number of different processes for characterizing MUPs. Stewart et al. developed and evaluated a system where using a pool of normal subjects the normative ranges were defined as the mean ± 2 standard deviations of the values of several different MUP features [6]. The error rate for characterizing mean MUP feature values as being detected from myopathic muscles was 44% and the neuropathic error rate was 36.7%.
Pattichis et al. [7] applied artificial neural network (ANN) models to the classification of MUPs sampled from normal, neuropathic and myopathic patients and achieved an error rate as low as 10%. The authors found that ANNs easily tended towards over-fitting (i.e., it was difficult to achieve generalization—the ability of the ANN to correctly classify unknown cases based on the training data.) In addition, they are one of the least transparent (basis of their decisions are easily understood) classifiers because of the often-large number of linear transformations applied to the feature values making ANNs essentially black box classifiers.
Pfeiffer & Kunze [8], [9] applied linear discriminant analysis (LDA) to classify MUPs by calculating the probability that a particular MUP was detected from a normal, myopathic or neuropathic muscle. The LDA technique works best using continuous feature values and was found to be an accurate method for MUP characterization. However, because of arithmetic transformation of feature values, it is not as easy to understand the basis of decisions made by an LDA classifier as it is with classifiers that simply use logical relationships between features, i.e., for three classes LDA uses two arithmetic expressions to provide a relationship among the features.
Methods to characterize a neuromuscular system and to measure the degree of involvement of a disorder require a MUP characterization process that is accurate, allows the basis of it decisions to be easily understood, produces a numeric value in support of or refutation of a characterization, and is able to achieve generalization. The above techniques do not completely satisfy these important requirements. It is hypothesized, that a pattern discovery (PD) classifier can meet these requirements. Thus, the current study compares the accuracy of a PD classifier when used for the characterization of MUPs (detected in voluntarily active muscles using an intramuscular needle electrode) with other classifiers commonly used for medical decision support. The ability of PD classification to meet the above-mentioned requirements relative to other classifiers is also discussed. The remainder of the paper is structured as follows: a requirements section describing in more detail the criteria needed for effective MUP characterization; a brief description of potential classifiers and how they meet the requirements for MUP characterization; methodology, which formally describes pattern discovery based classification, provides rationale for choosing the other classifiers for comparison and describes how simulated data was generated and clinical data gathered; Section 5 reporting the performance of the different classifiers; and finally Sections 6 (Discussion) and 7 (Conclusions).
Section snippets
Requirements for MUP characterization
The requirements for MUP characterization are based on ideas developed by Kononenko [10] and Sprogar et al. [11] who describe a set of requirements needed for machine learning systems used in medical decision support.
Mixed mode data discrete classifiers
Naïve Bayesian (NB), Decision Tree (DT), and pattern discovery (PD) classifiers were designed to handle mixed mode data. They are discrete classifiers since they handle nominal and discrete data types and require continuous features to be quantized (segmenting the range of a feature's values into distinct intervals). NB and DT classifiers are widely used while PD based classifiers have been developed recently.
A NB classifier is built by assuming that all of the features are conditionally
Methods
This section formally defines pattern discovery based classification, provides a rationale for choosing the other classifiers for comparison, and describes how MUP data was obtained and represented and how classifier performance was measured.
Clinical MUP data
PD had an average error rate (and standard deviation) of 30.3% (1.8%), LDA 29.0% (2.1%), J48 DT 30.1% (2.1%) and NB 29.8% (1.8%) across thirty trials. All four characterization methods had similar error rates with no statistically significant differences between each other according to the Tukey post hoc test at a significance level of 0.05. Table 1 shows NB had the highest sensitivity and LDA had the highest specificity and accuracy. However, PD had a significantly lower SSD compared to the
Discussion
Table 4 summarizes a comparison of the studied classifiers for four of the identified requirements based on the study results and consideration of the known properties of each classifier. PD has an advantage with respect to transparency. All four methods have similar accuracy. PD, DT and NB can handle mixed mode data types while LDA cannot because of its inability to handle nominal data. All four classifiers can produce a numeric value for characterization. In addition, while the computational
Conclusions
Unlike the other classifiers investigated, the PD classifier is able to explain itself by reporting the sets of feature values, along with a strength-of-evidence measure, supporting or refuting its characterizations. This work has demonstrated that the PD classifier meets the requirements for normal and neuropathic MUP characterization through its abilities to report its characterizations in a transparent manner, handle mixed mode data, discover dependencies among features, provide numerical
Conflict of interest
None.
References (35)
- et al.
Discriminant classification of motor unit potentials (MUPs) successfully separates neurogenic and myogenic conditions, a comparison of multi- and univariate diagnostic algorithms for MUP analysis
Electroencephalogr Clin Neurophysiol
(1995) Machine learning for medical diagnosis: history, state of the art and perspective
Artif Intell Med
(2001)Decomposition and quantitative analysis of clinical electromyographic signals
Med Eng Phys
(1999)- et al.
Within-subject reliability of motor unit number estimates and quantitative motor unit analysis in a distal and proximal upper limb muscle
Clin Neurophysiol
(2006) - et al.
Analysis of amplitude and area of concentric needle EMG motor unit action potentials
Electroencephalogr Clin Neurophysiol
(1988) - et al.
Explaining evidential analyses
Int J Approx Reason
(1989) - et al.
Judgement under uncertainty
Science
(1974) - et al.
Engineering psychology and human performance
(2000) - et al.
Interrater reliability of the needle examination in lumbosacral radiculopathy
Muscle Nerve
(2006) What does the neurologist expect from clinical neurophysiology?
Muscle Nerve
(2002)
Stedman's medical dictionary
Evaluation of an automatic method of measuring features of motor unit action potentials
Muscle Nerve
Neural network models in EMG diagnosis.
IEEE Trans Biomed Eng
The diagnostic power of motor unit potential analysis: an objective Bayesian approach
Muscle Nerve
Evolution in medical decision making
J Med Syst
Needle EMG abnormalities in neurogenic and muscle diseases
Pattern classification
Cited by (27)
Comparison of decision tree algorithms for EMG signal classification using DWT
2015, Biomedical Signal Processing and ControlVisualizing information for electromyographic characterization
2014, Procedia Computer ScienceMedical decision support system for diagnosis of neuromuscular disorders using DWT and fuzzy support vector machines
2012, Computers in Biology and MedicineCitation Excerpt :Similar studies of LDA, C4.5. DT, MLPNN and SVM as an EMG signal classifier can be found in different studies [10–16]. If we compare the results taken from this study with the previous results [6–16], the results taken from this study is 3% better than the previous results.
Quantitative Electromyography
2012, Aminoff's Electrodiagnosis in Clinical NeurologyAdaptive motor unit potential train validation using MUP shape information
2011, Medical Engineering and PhysicsChapter 25 Decision support for QEMG
2009, Supplements to Clinical NeurophysiologyCitation Excerpt :There are a number of methods for estimating the MUP conditional probabilities comprising a MUP characterization. Pino et al. (2007) compared MUP characterizations using LDA and PD. They discovered that each provided valid estimates of the required conditional probabilities.