An EMGbased feature extraction method using a normalized weight vertical visibility algorithm for myopathy and neuropathy detection
 Patcharin Artameeyanant^{1},
 Sivarit Sultornsanee^{2} and
 Kosin Chamnongthai^{1}Email author
DOI: 10.1186/s4006401637722
© The Author(s) 2016
Received: 2 July 2016
Accepted: 30 November 2016
Published: 20 December 2016
Abstract
Background
Electromyography (EMG) signals recorded from healthy, myopathic, and amyotrophic lateral sclerosis (ALS) subjects are nonlinear, nonstationary, and similar in the time domain and the frequency domain. Therefore, it is difficult to classify these various statuses.
Methods
This study proposes an EMGbased feature extraction method based on a normalized weight vertical visibility algorithm (NWVVA) for myopathy and ALS detection. In this method, sampling points or nodes based on sampling theory are extracted, and features are derived based on relations among the vertical visibility nodes with their amplitude differences as weights. The features are calculated via selective statistical mechanics and measurements, and the obtained features are assembled into a feature matrix as classifier input. Finally, powerful classifiers, such as knearest neighbor, multilayer perceptron neural network, and support vector machine classifiers, are utilized to differentiate signals of healthy, myopathy, and ALS cases.
Results
Performance evaluation experiments are carried out, and the results revealed 98.36% accuracy, which corresponds to approximately a 2% improvement compared with conventional methods.
Conclusions
An EMGbased feature extraction method using a NWVVA is proposed and implemented to detect healthy, ALS, and myopathy statuses.
Keywords
EMG signal Complex network Normalized weight vertical visibility algorithm Network measurements kNearest neighbor Multilayer perceptron neural network Support vector machineBackground
Currently, amyotrophic lateral sclerosis (ALS), or neuropathy, a rapidly progressive, invariably fatal neurological disease that affects the neurons responsible for controlling voluntary muscles in the arms, legs, and face (Ahdab et al. 2013), is diagnosed in approximately 6000 people each year (ALS Association 2016). In the USA alone, the number of patients is estimated to be as many as 20,000. This disease belongs to a group of motor neuron disorders and eventually leads to death. According to previous studies, patients who are diagnosed live an average of 3 years, and 20, 10, and 5% of them die in 5, 10, and 20 years, respectively. Myopathy is a neuromuscular disorder that causes muscle cramps, stiffness, and spasms, and muscle weakness is the primary symptom due to dysfunction of muscle fibers and eventually causes death. In accordance with the 2005 statistics data of the USA (Oskarsson 2011), approximately 2.97 million patients have been diagnosed with myopathy. In diagnosing both aforementioned diseases, medical doctors first interview patients, although sometimes the patients are extremely weak and unavailable to even speak. In such cases, electromyography (EMG) is used to analyze muscle signals to assist a specialized neurological expert to diagnose both myopathy and ALS (Kincaid 2015; Weiss et al. 2015; Gitiaux et al. 2016). However, the number of neurological experts is quite limited, and therefore, an automatic system to assist diagnosis is urgently required. Such a system could be used not only for assisting diagnosis but also for periodic detection and monitoring. In performing diagnoses based on EMG signals, a primary issue is that the system must correctly classify an EMG signal as ALS or myopathic, because different therapies and drugs are used to treat the two disorders.
In studying and developing this kind of system, EMG signals is regarded as an excellent approach for acquiring data (Yousefi and HamiltonWright 2014), which records the corresponding electrical to activity of motor units in the neuromuscular system. Analysis of EMG signals is generally performed in two cases. The first is for prosthetic device control and human–machine interactions (Naik and Kumar 2011; Naik et al. 2014, 2016a; Arjunan et al. 2014, 2015; Guo et al. 2015; Naik and Nguyen 2015). The second is for diagnosing disorders (Xie et al. 2014). Neuromuscular disorders are related to pathological changes in the structure of the motor unit and can be generally divided into two categories: muscular (myopathy) and neuronal (neuropathy) (Nikolic 2001) disorders. The need for distinct classification between myopathy and neuropathy originates from the differences between the causes of the diseases, which is a critical factor in determining treatment. The development of a highly accurate diagnostic system based on EMG readings would provide a promising way to improve the assessment of neuromuscular disorders (Gokgoz and Subasi 2015). Highly accurate classification problems depend on the crucial step of feature extraction. If features are extracted sufficiently well, it is possible to obtain outstanding classification performance.
Previous studies related to feature extraction of EMG signals have been proposed in three main domains, the frequency domain, the time–frequency domain, and the complex network domain. In frequency analyses, fast Fourier transform (FFT) and autoregressive (AR) spectral models have been employed to extract features (Guler and Kocer 2005; Subasi et al. 2006; Kocer 2010; Sultornsanee et al. 2011). Power spectral analysis of FFT and AR can represent the characteristics of the signal. However, different subjects have different signal strengths in addition to nonlinearity and chaos. Various types of wavelets have been used to analyze EMG signals in the time–frequency domain (Gokgoz and Subasi 2015; Hu et al. 2005; Istenic et al. 2010; Subasi 2012a, b, 2013a, b). The advantage of the method is the ability to perform analyses in various subbands. However, computational complexity might occur at the initial stages, such as when selecting the mother wavelet. Additionally, the level of decomposition is related to the number of subbands. Using many subbands with various features in each subband results in a high dimension of input for the classifier. Mishra et al. (2016) and Naik et al. (2016b) utilized an empirical mode decomposition technique to analyze EMG signals, which was proven to be quite versatile over a broad range of applications for extracting signals from data generated in noisy nonlinear and nonstationary processes.
Finally, for the complex network domain, Campanharo et al. (2011) studied the duality between the time series and networks and proposed a map of the time series resulting in networks with distinct topological properties. Thus, nonlinear signals can be transformed into a complex network using a visibility algorithm. Lacasa et al. (2007) proposed a visibility algorithm to convert a time series signal into a graph. The resulting graph inherited several properties of the series in its structure. Luque et al. (2009) employed a horizontal visibility algorithm, which is a geometrically simpler and an analytically solvable version of the visibility algorithm. All the aforementioned works on the complex network domain are pure theoretical concepts without evidence of implementation in signal analysis. Tang et al. (2013) used visibility graphs from higher frequency bands to classify electroencephalogram (EEG) signals. They concluded that their approach is better than the simple entropy method. Additionally, Zhu et al. (2012) employed visibility graphs with nonlinear feature extraction algorithms on the EEG signal, although their algorithms were slower than FFT analysis, which is not suitable for practical purposes. Subsequently, Zhu et al. (2014) introduced the fastweighted horizontal visibility algorithm (FHVA). The FHVA can be employed using signals that have high amplitude variations. However, the FHVA is not suitable for EMG signals because the algorithm uses a horizontal relationship, which does not distinguish features sufficiently well; thus, the classification results using this method are incorrect.
In our previous works, Artameyanant et al. (2014) proposed a feature extraction technique based on transforming the signal into a complex network using a vertical visibility algorithm. The method yielded excellent accuracy results. However, a rapidly decreasing/increasing signal configuration could yield the same features. Therefore, a classification error could occur. The authors then improved upon the work in Artameyanant et al. (2014) by presenting a weightvisibility algorithm for transforming the signal into a complex network (Artameyanant et al. 2015). The method solved the problem of the same features being yielded for a different type of signal. However, the drawback was the loss of the link in the calculation caused by the same amplitude of the signal. Additionally, the EMG signal of each subject for the same type of disease can vary in signal strength. Thus, the various strengths of signals for different patients can induce classification problems. In this paper, we overcome the drawbacks of our previous work with two steps of feature extraction. First, we propose normalizing the signal with respect to the maximum/minimum value of each epoch. The normalized signal corresponds to the visual inspection of the same scale of the signal pattern by neurological expertise for classification. Second, we introduce an adjustedweight vertical visibility algorithm to obtain the adjacency matrix for network measurements. The proposed work shows that feature extraction based on network measurements of the adjustedweight vertical visibility algorithm can be used as an analysis tool for EMG signals. Some distinct characteristics inherited in the signal are extracted and employed as a feature vector. Performance is evaluated using several types of classifiers: kNN, MLPNN, and SVM. The proposed method yields outstanding average accuracy results.
The organization of the paper is as follows. In “EMG signal analysis and basic concept” section, we analyze the research problem and outline the basic concept. In “Proposed method of EMGbased feature extraction” section, we explain the proposed method according to the basic concept. We describe the datasets and experimental results in “Datasets and experimental results” section, and discuss errors and tradeoffs in “Discussion” section. Finally, the research is concluded in “Conclusions” section.
EMG signal analysis and basic concept
To select efficient tools for feature extraction and classification, we analyze the EMG signal and explain our ideas in this section.
Proposed method of EMGbased feature extraction
Based on the aforementioned basic concept, our proposed method of EMGbased feature extraction for ALS and myopathy detection begins with preprocessing, followed by feature extraction and classification processes. Process overview of the proposed method is explained in “Overview processes of proposed method” section, and all proposed processes, including preprocessing, feature extraction, and classification, are described in “Preprocessing”–“Classification” sections, respectively.
Overview processes of proposed method
Preprocessing
Feature extraction
In the feature extraction process, as shown by the second dashed rectangle in the flowchart in Fig. 6, a normalized epoch is first sampled based on the sampling theory, and the sampled pulses are then extracted for vertical visibility features including the number of node links and weights using the normalized weight vertical visibility algorithm (NWVVA) in the process of matrix creation. Those links and weights are put into matrix form, the feature matrices and obtained features are filtered and considered by statistical machines for selected powerful features in the next step of statistical feature extraction. Effective statistical features are selected during the step of the last process. The following section is divided into twosubsections, extraction of candidate features and feature finalization.
Matrix creation
In the example shown in Fig. 9, nine sampling points are obtained. For a given sampling point or node, all other surrounding sampling peaks to which straight lines from the considered point can be drawn without any obstacles are defined as related to the sampling point, and these related nodes are counted and used to create an adjacency matrix. As show the sample of Fig. 9a, the sampling point 4 is related with sampling points 1 and 5 via high sight (the node is looking up) and 2 and 3 via low sight (the node is looking down). However, the sampling point 6–9 are hidden by sampling point 5, and therefore, that no relation is counted from them. To account for both the relation link and amplitude features, element W_{ ij } of the weighted adjacency matrix is obtained as follows. If there exists a link between node i and j, W_{ ij } is first set to 1 to account for the relation link, and then the absolute difference of the normalized amplitudes between nodes i and j is added, which produces element (W_{ ij }) of the matrix. If node i and j have no link, W_{ ij } is set to zero. All diagonal elements (W_{ ij }), which indicate links with itself are set to zero.
The procedure of the aforementioned concept can be described based on normalized weight vertical visibility algorithm (NWVVA) as follows:
In the example shown in Fig. 9b, the values of the relations of the weighted adjacency matrix are determined as follows. There exists a link between the 1st and 2nd data points whose amplitudes are 0.87 and 0.49, respectively. Hence, W_{ 12 } is the absolute value of (0.87–0.49); adding “1” equals 1.38, while the same procedure is applied to obtain other elements of the matrix. As a result, a weight adjacency matrix for this signal is obtained as shown in Fig. 9b.
Statistical feature extraction
In statistical feature extraction, it is complicated and redundant to classify epochs by some classifiers using perceptron data and the features extracted in the previous process. Because these features hold statistical characteristics in each target classified groups (normal, ALS, and myopathy), it is better for users to utilize statistical mechanics and statistical measurements as inputs to the appropriate classification tools. However, because not all statistical mechanics and measurements are effective for classification, a process of selecting effective statistical mechanics is needed in the learning state, which could be done in advance. Such a way to select effective statistical mechanics and measurements is introduced as a guideline as follows.
In the learning state, users should first calculate candidate features of the number of links and weights obtained via vertical visibility by using possible statistical mechanics and measurements and then consider selecting only the effective features based on the selected set of training signals. The selected statistical mechanics and measurements are then used to find final features in the testing state.
The statistical mechanics and measurements as candidates for selection in the learning state are introduced in “Average degree”–“Kurtosis” sections, respectively, as follows.
Average degree
In the sample in Fig. 9, L = 16, and N = 9, hence AD = 3.55.
Average clustering coefficient
In the sample in Fig. 9, N = 9, L _{ i }: 4, 3, 3, 4, 6, 3, 3, 4, 2. As a result, Ci = 0.82, 1.29, 1.32, 0.82, 0.39, 1.29, 1.32, 0.82, and 1.02 for i = 1, 2, …,9, respectively; hence ACC = 1.01.
Transitivity
In the sample in Fig. 9, number of triangles in the network and connected triples are 74.73 and 92, respectively; hence T = 0.81.
Assortativity
A positive value of As indicates that the nodes tend to link to other nodes of an identical or similar degree. In the example in Fig. 9, AS = −0.32.
Density
In the example in Fig. 9, L = 16 and N = 9, hence Den = 0.44.
Central point dominance
In the example in Fig. 9, B _{ u } = 10, 0, 0, 10, 34, 0, 0, 2, 0 and max(B _{ u } = 34), hence CPD = 25.77.
Closeness centrality
Average shortest path (ASP)
In the example in Fig. 9, d _{ ij } = \( \left[ {\begin{array}{*{20}l} 0 \hfill & {0.72} \hfill & {0.66} \hfill & {0.96} \hfill & 1 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ {0.72} \hfill & 0 \hfill & {0.88} \hfill & {0.74} \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ {0.66} \hfill & {0.88} \hfill & 0 \hfill & {0.68} \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ {0.96} \hfill & {0.74} \hfill & {0.68} \hfill & 0 \hfill & {0.96} \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ 1 \hfill & 0 \hfill & 0 \hfill & {0.96} \hfill & 0 \hfill & {0.72} \hfill & {0.66} \hfill & {0.96} \hfill & 1 \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & {0.72} \hfill & 0 \hfill & {0.88} \hfill & {0.74} \hfill & 0 \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & {0.66} \hfill & {0.88} \hfill & 0 \hfill & {0.68} \hfill & 0 \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & {0.96} \hfill & {0.74} \hfill & {0.68} \hfill & 0 \hfill & {0.96} \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 1 \hfill & 0 \hfill & 0 \hfill & {0.96} \hfill & 0 \hfill \\ \end{array} } \right] \) and N = 9, and hence ASP = 0.37.
Global efficiency (E)
In the example in Fig. 9, d _{ ij } = \( \left[ {\begin{array}{*{20}l} 0 \hfill & {0.72} \hfill & {0.66} \hfill & {0.96} \hfill & 1 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ {0.72} \hfill & 0 \hfill & {0.88} \hfill & {0.74} \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ {0.66} \hfill & {0.88} \hfill & 0 \hfill & {0.68} \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ {0.96} \hfill & {0.74} \hfill & {0.68} \hfill & 0 \hfill & {0.96} \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill \\ 1 \hfill & 0 \hfill & 0 \hfill & {0.96} \hfill & 0 \hfill & {0.72} \hfill & {0.66} \hfill & {0.96} \hfill & 1 \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & {0.72} \hfill & 0 \hfill & {0.88} \hfill & {0.74} \hfill & 0 \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & {0.66} \hfill & {0.88} \hfill & 0 \hfill & {0.68} \hfill & 0 \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & {0.96} \hfill & {0.74} \hfill & {0.68} \hfill & 0 \hfill & {0.96} \hfill \\ 0 \hfill & 0 \hfill & 0 \hfill & 0 \hfill & 1 \hfill & 0 \hfill & 0 \hfill & {0.96} \hfill & 0 \hfill \\ \end{array} } \right] \) and N = 9, hence E = 0.54
Network diameter (D)
In the example in Fig. 9, max(d _{ ij }) = 4, and hence D = 0.96.
Average weight
In the example in Fig. 9, \( \mathop \sum \limits_{i} \mathop \sum \limits_{j} W_{ij} = 39.64 \), and hence AW = 4.40.
Skewness
In the example in Fig. 9, μ = 0.49 and s = 0.62, and hence skewness = 0.56.
Kurtosis
In the example in Fig. 9, μ = 0.49 and s = 0.62, and hence kurtosis = 1.64.
These calculation results obtained by the selected statistical mechanics explained above are evaluated by ANOVA (Wassernman 2013), and the evaluated results are used to construct feature vectors. These vectors would be classified into healthy, myopathy and ALS statuses, which is explained in the next section.
Selection of effective statistical features
In the aforementioned examples, average degree, average cluster coefficient, density, average weight, skewness, and kurtosis are statistically selected as six effective features. However, users are recommended to undertake this type of pretesting or training using their own samples to obtain effective features for their datasets.
Classification
Datasets and experimental results
Datasets
In our experiments, the databases 1 (Physionet 2016) and 2 (Nikolic 2001) used in the conventional methods are employed under the objective of fair comparison with the results of conventional methods, and the results classified by the kNN, MLPNN, and SVM classifiers are shown and compared with those in the conventional methods, as follows.
The p values of the statistical mechanics for databases 1 and 2 using ANOVA
Feature extraction  p value  

Database 1  Database 2  
AD  p < 0.001  p < 0.001 
ACC  p < 0.001  p < 0.001 
Den  p < 0.001  p < 0.001 
AW  p < 0.001  p < 0.001 
Skewness  p < 0.001  p < 0.001 
Kurtosis  p < 0.001  p < 0.001 
Experimental results

Specificity the number of correctly classified normal subjects divided by the number of total normal subjects.

Sensitivity (myopathy) the number of correctly classified subjects suffering from myopathy divided by the number of total subjects suffering from myopathy.

Sensitivity (neuropathy) the number of correctly classified subjects suffering from neuropathy divided by the number of total subjects suffering from neuropathy.

Total classification accuracy the number of correctly classified subjects divided by the number of total subjects.
Summary of the classification performance of the proposed method
Test (%)  Statistical parameter  Database 1  Database 2  

kNN  MLPNN  SVM  kNN  MLPNN  SVM  
20  Specificity  94.86 ± 1.48  96.08 ± 1.24  98.68 ± 1.02  94.50 ± 1.18  95.37 ± 1.64  97.86 ± 0.84 
Sensitivity (neuropathy)  96.46 ± 0.68  97.18 ± 0.96  98.98 ± 0.58  95.20 ± 1.69  96.26 ± 1.86  98.26 ± 0.94  
Sensitivity (myopathy)  98.28 ± 1.27  98.82 ± 1.20  99.86 ± 0.86  97.90 ± 1.20  97.23 ± 1.02  98.98 ± 0.62  
Total classification accuracy  96.53 ± 0.92  97.36 ± 0.74  99.17 ± 0.68  95.87 ± 0.85  96.28 ± 0.62  98.36 ± 0.48  
40  Specificity  92.54 ± 1.84  95.23 ± 1.84  98.42 ± 1.62  91.95 ± 2.01  94.78 ± 1.82  97.08 ± 1.58 
Sensitivity (neuropathy)  94.98 ± 1.92  96.02 ± 1.69  98.21 ± 1.82  94.55 ± 1.57  95.72 ± 2.02  98.26 ± 1.42  
Sensitivity (myopathy)  98.18 ± 0.74  98.18 ± 0.92  99.26 ± 0.94  98.15 ± 0.91  97.56 ± 1.24  98.86 ± 1.08  
Total classification accuracy  95.23 ± 0.86  96.47 ± 0.76  98.63 ± 0.82  94.88 ± 0.78  96.02 ± 0.89  98.06 ± 0.68  
50  Specificity  92.46 ± 2.26  95.46 ± 1.92  97.04 ± 2.12  91.24 ± 2.58  94.18 ± 2.26  96.12 ± 1.98 
Sensitivity (neuropathy)  93.78 ± 1.69  96.28 ± 1.96  98.98 ± 1.65  92.68 ± 1.72  95.23 ± 1.68  98.62 ± 1.28  
Sensitivity (myopathy)  98.75 ± 0.94  98.92 ± 0.81  99.98 ± 1.46  98.04 ± 0.83  97.98 ± 1.24  99.24 ± 0.98  
Total classification accuracy  94.99 ± 0.83  96.88 ± 0.68  98.63 ± 0.94  93.99 ± 0.97  95.79 ± 0.72  97.99 ± 0.82  
60  Specificity  91.89 ± 2.34  93.49 ± 3.02  96.78 ± 2.98  90.97 ± 2.00  92.43 ± 2.98  96.02 ± 2.46 
Sensitivity (neuropathy)  91.94 ± 2.28  95.82 ± 2.46  97.69 ± 2.46  91.10 ± 2.43  94.39 ± 2.04  96.92 ± 1.48  
Sensitivity (myopathy)  97.64 ± 0.87  97.46 ± 1.04  99.43 ± 1.24  96.73 ± 0.64  95.45 ± 1.12  98.42 ± 1.12  
Total classification accuracy  93.82 ± 0.93  95.59 ± 0.68  97.96 ± 0.86  92.93 ± 0.75  94.09 ± 0.74  97.12 ± 0.84  
80  Specificity  88.46 ± 2.46  92.63 ± 1.92  94.98 ± 1.95  87.73 ± 2.05  91.94 ± 2.02  94.76 ± 3.98 
Sensitivity (neuropathy)  88.12 ± 1.24  93.82 ± 1.28  97.84 ± 1.74  87.65 ± 1.61  92.83 ± 1.76  97.14 ± 1.92  
Sensitivity (myopathy)  98.06 ± 1.29  95.92 ± 1.24  98.68 ± 1.84  97.60 ± 1.09  93.92 ± 1.82  97.64 ± 2.04  
Total classification accuracy  91.54 ± 0.83  94.12 ± 0.98  97.16 ± 0.95  90.99 ± 0.82  92.89 ± 0.94  96.51 ± 0.96 
Summary of the classification performances of previous methods
Method (feature + classification)  Total classification accuracy (%) 

Database 1  
RQA + SVM (Sultornsanee et al. 2011)  98.28 
VVA + SVM (Artameyanant et al. 2014)  99.07 
WVA + MLPNN (Artameyanant et al. 2015)  94.73 
Proposed method  99.17 
Database 2  
AR + WNN (Subasi et al. 2006)  90.70 
CWT + SVM (Istenic et al. 2010)  70.40 
AR + neurofuzzy system (Kocer 2010)  90.00 
94.00  
DWT + ESVM (Subasi 2013a)  97.00 
DWT + PSOSVM (Subasi 2013b)  97.41 
DWT + random forest (Gokgoz and Subasi 2015)  96.67 
Proposed method  98.36 
Discussion
This paper proposes a method of EMGbased feature extraction using a normalized weight vertical visibility algorithm for ALS and myopathy detection. Due to the effectiveness of specific features of the vertical visibility algorithm with normalized weights, which are well matched with the patterns of ALS and myopathy signals, the proposed method yields better classification accuracy results compared with conventional methods as shown in Table 3. For studies targeting applications in medicine, which is critical for improving human life, the experimental results should ideally be perfect without any errors. However, the proposed method contributes to a new approach, which currently corresponded to best accuracy results that approached 100%. Research on this topic should be accepted and continue to be studied until the results successfully meet the final goal. Regarding errors in the experiments, their causes and how to prevent errors are analyzed and discussed as follows.
The proposed method of EMGbased feature extraction using a normalized weight vertical visibility algorithm for myopathy and ALS detection improves classification accuracy and advantages. To obtain improved accuracy, computational complexity and time implicitly become disadvantages as tradeoffs. Although the increase in computational time is often considered in comparisons with conventional methods, the necessary computational time in the proposed method is on the order of milliseconds, which is practically acceptable due to prominent improvements in current computing technologies.
During the final classification step of the proposed method, some popular classifiers such as kNN, MLPNN, and SVM classifiers, were recommended and tested here. Users are recommended to find their own appropriate tools, which should match their applications. As shown in Table 2, the kNN, MLPNN, and SVM classifiers yielded excellent accuracies as approximately 96, 97, and 98%, respectively. Although the results show that the SVM classifier, which yielded the highest accuracy, should be recommended as the classification tool in terms of accuracy, the accuracy differences compared with the other classifiers were not extremely high. In some applications that require highly efficient training with low complexity, kNN classifiers should be considered as another choice. On the other hand, MLPNN classifiers, which are theoretically designed as a tool to address complicated classification with slightly high complexity, could be a compromise in some applications that require some level of complexity.
Conclusions
This paper proposes a method of EMGbased feature extraction using a normalized weight vertical visibility algorithm for myopathy and neuropathy detection. In the proposed method, EMG signals representing muscle responses were sampled based on the sampling theory for reversible discrete pulses, and the features of the obtained pulses were then extracted via a vertical visibility algorithm with their normalized weights. An adjacent matrix, whose elements represent links between nodes and their weights, was accordingly created and employed to extract statistical features using statistical mechanics and measurements. These statistical features were finally classified using kNN, MLPNN, and SVM classifiers into normal, ALS, and myopathic cases. To evaluate the performance of the proposed method, experiments were performed on conventional 2 databases, and the results revealed 98.36% accuracy, which is approximately 2% improvement compared with conventional methods.
Abbreviations
 ALS:

amyotrophic lateral sclerosis
 EMG:

electromyography
 NWVVA:

normalized weight vertical visibility algorithm
 kNN:

knearest neighbor
 MLPNN:

multilayer perceptron neural network
 SVM:

support vector machine
 FFT:

fast Fourier transform
 AR:

autoregressive spectral models
 FHVA:

fast weighted horizontal visibility algorithm
 AD :

average degree
 ACC :

average cluster coefficient
 T :

transitivity
 As :

assortativity
 Den :

density
 CPD :

central point dominance
 CC :

closeness centrality
 ASP :

average shortest path
 E :

global efficiency
 D :

network diameter
 AW :

average weight
Declarations
Authors’ contributions
PA, SS, and KC made substantial contributions to the conception and design of the method proposed. PA, SS, and KC designed experimental procedure. PA drafted the manuscript. KC revised the manuscript to meet the expected standards of scientific publishing. All authors read and approved the final manuscript.
Acknowledgements
Financial support via a faculty department program scholarship awarded to the first author by Vongchavalitkul University, Thailand, is gratefully acknowledged. The authors also thank Asst. Prof. Dr. Pinit Kumhom, an English expert, for editing the manuscript.
Competing interests
The authors declare that they have no competing interests.
Availability of data and materials
The datasets supporting the conclusions of this article are available in the http://www.physionet.org/physiobank/database/emgdb/ and http://www.emglab.net.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 Ahdab R, Creange A, SaintVal C, Farhat WH, Lefaucheur JP (2013) Rapidly progressive amyotrophic lateral sclerosis initially masquerading as a demyelinating neuropathy. Neurophysiol Clin Clin Neurophysiol 43(3):181–187View ArticleGoogle Scholar
 ALS Association (2016) Who gets ALS? http://www.alsa.org/aboutals/factsyoushouldknow.html. Accessed 30 Sept 2016
 Arjunan SP, Kumar DK, Naik G (2014) Computation and evaluation of features of surface electromyogram to identify the force of muscle contraction and muscle fatigue. BioMed Res Int 2014:197960View ArticlePubMedPubMed CentralGoogle Scholar
 Arjunan SP, Kumar D, Naik G (2015) Independence between two channels of surface electromyogram signal to measure the loss of motor units. Meas Sci Rev 15(3):152–155View ArticleGoogle Scholar
 Artameyanant P, Sultornsanee S, Chamnongthai K, Higuchi K (2014) Classification of electromyogram using vertical visibility algorithm with support vector machine. Presented at the 2014 signal and information processing association annual summit and conference, Siem Reap, Cambodia, 9–12 December 2014
 Artameyanant P, Sultornsanee S, Chamnongthai K (2015) Classification of electromyogram using weight visibility algorithm with multilayer perceptron neural network. Presented at the 7th international conference on knowledge and smart technology, Bang San, Thailand, 28–31 January 2015
 Barabasi AL (2012) Network science. Ebook version. http://barabasilab.neu.edu/networksciencebook/download/network_science_November_Ch1_2012.pdf. Accessed 16 May 2013
 Boccaletti S, Latora V, Moreno Y, Chaves M, Hwang DU (2006) Complex networks: structure and dynamics. Phys Rep 424:175–308. doi:10.1016/j.physrep.2005.10.009 ADSMathSciNetView ArticleGoogle Scholar
 Campanharo ALSO, Sirer MI, Malmgren RD, Ramos FM, Amaral LAN (2011) Duality between time series and networks. PLoS ONE 6(8):1–12View ArticleGoogle Scholar
 Costa LDF, Boas PRV, Silva FN, Rodrigues FA (2010) A pattern recognition approach to complex networks. J Stat Mech Theory Exp. doi:10.1088/17425468/2010/11/P11015 Google Scholar
 Cover T, Hart P (1967) Nearest neighbor pattern classification. IEEE Trans Inf Theory 13:21–27View ArticleMATHGoogle Scholar
 Gitiaux C, Chemaly N, QuijanoRoy S, Barnerias C, Desguerre I, Hully M, Chiron C, Dulac O, Nabbout R (2016) Motor neuropathy contributes to crouching in patients with Dravet syndrome. Neurology 87(3):277–281View ArticlePubMedGoogle Scholar
 Gokgoz E, Subasi A (2015) Comparison of decision tree algorithms for EMG signal classification using DWT. Biomed Signal Process Control 18:138–144View ArticleGoogle Scholar
 Guler NF, Kocer S (2005) Classification of EMG signals using PCA and FFT. J Med Syst 29(3):241–250View ArticlePubMedGoogle Scholar
 Guo Y, Naik GR, Huang S, Abraham A, Nguyen HT (2015) Nonlinear multiscale maximal Lyapunov exponent for accurate myoelectric signal classification. Appl Soft Comput 36:633–640View ArticleGoogle Scholar
 Haykin S (1994) Neural networks a comprehensive foundation. Macmillan, New York, pp 178–277MATHGoogle Scholar
 Hu X, Wang Z, Ren X (2005) Classification of surface EMG signal using relative wavelet packet energy. Comput Methods Progr Biomed 17:189–195View ArticleGoogle Scholar
 Istenic R, Kaplanis PA, Pattichis CS, Zazula D (2010) Multiscale entropybased approach to automated surface EMG classification of neuromuscular disorders. Med Biol Comput 48:773–781View ArticleGoogle Scholar
 Kincaid JC (2015) Nerve conduction studies and needle EMG. Nerves and nerve injuries, vol 1. Elsevier, London, pp 125–145View ArticleGoogle Scholar
 Kocer S (2010) Classification of EMG signals using neurofuzzy system and diagnosis of neuromuscular diseases. J Med Syst 34:321–329View ArticlePubMedGoogle Scholar
 Krebel U (1999) Pairwise classification and support vector machines. Advances in kernel methodssupport vector learning. MIT, Cambridge, pp 255–268Google Scholar
 Lacasa L, Luque B, Ballesteros F, Luque J, Nuno JC (2007) From time series to complex networks: the visibility graph. PNAS 105(13):4972–4975ADSMathSciNetView ArticleMATHGoogle Scholar
 Luque B, Lacasa L, Ballesteros F, Luque J (2009) Horizontal visibility graphs: exact results for random time series. Phys Rev E 80:1–11View ArticleGoogle Scholar
 Mishra VK, Bajaj V, Kumar A, Singh GK (2016) Analysis of ALS and normal EMG signals based on empirical mode decomposition. IET Sci Meas Technol 10(8):963–971View ArticleGoogle Scholar
 Naik GR, Kumar DK (2011) Estimation of independent and dependent components of noninvasive EMG using fast ICA: validation in recognizing complex gestures. Comput Methods Biomech Biomed Eng 14(12):1105–1111View ArticleGoogle Scholar
 Naik GR, Nguyen HT (2015) Nonnegative matrix factorization for the identification of EMG finger movements: evaluation using matrix analysis. IEEE J Biomed Health Inform 19(2):478–485View ArticlePubMedGoogle Scholar
 Naik GR, Kumar DK, Palaniswami M (2014) Signal processing evaluation of myoelectric sensor placement in lowlevel gestures: sensitivity analysis using independent component analysis. Expert Syst 31(1):91–99View ArticleGoogle Scholar
 Naik GR, Selvan SE, Gobbo M, Acharyya A, Nguyen HT (2016a) Principal component analysis applied to surface electromyography: a comprehensive review. IEEE Access 4:4025–4037. doi:10.1109/ACCESS.2016.2593013 View ArticleGoogle Scholar
 Naik GR, Selvan SE, Nguyen HT (2016b) Singlechannel EMG classification with ensembleempiricalmodedecompositionbased ICA for diagnosing neuromuscular disorders. IEEE Trans Neural Syst Rehabil Eng 24(7):734–743View ArticlePubMedGoogle Scholar
 Newman MEJ (2003) The structure and function of complex networks. SIAM Rev 45(2):167–256. doi:10.1137/S003614450342480 ADSMathSciNetView ArticleMATHGoogle Scholar
 Newman MEJ (2010) Networks: an introduction. Oxford University Press, New York, pp 1–771View ArticleGoogle Scholar
 Nikolic M (2001) Detailed analysis of clinical electromyography signals EMG decomposition, findings and firing pattern analysis in controls and patients with myopathy and amyotrophic lateral sclerosis. A Ph.D. Dissertation submitted to the University of Copenhagen: the Faculty of Health Science. August
 Oskarsson B (2011) Myopathy: five new things. Neurology 76:14–19. doi:10.1212/WNL.0b013e31820c3648 View ArticleGoogle Scholar
 Physionet (2016) https://www.physionet.org/physiobank/database/emgdb. Accessed 9 Feb 2016. doi:10.13026/C24S3D
 Subasi A (2012a) Classification of EMG signals using combined features and soft computing techniques. Appl Soft Comput 12:2188–2198View ArticleGoogle Scholar
 Subasi A (2012b) Medical decision support system for diagnosis of neuromuscular disorders using DWT and fuzzy support vector machines. Comput Biol Med 42:806–815View ArticlePubMedGoogle Scholar
 Subasi A (2013a) A Decision support system for diagnosis of neuromuscular disorders using DWT and evolutionary support vector machines. SIViP. doi:10.1007/s117600130480z Google Scholar
 Subasi A (2013b) Classification of EMG signals using PSO optimized SVM for diagnosis of neuromuscular disorders. Comput Biol Med 43:576–586View ArticlePubMedGoogle Scholar
 Subasi A, Yilmaz M, Ozcalik HR (2006) Classification of EMG signals using wavelet neural network. J Neurosci Methods 156:360–367View ArticlePubMedGoogle Scholar
 Sultornsanee S, Zeid I, Kamarthi S (2011) Classification of electromyogram using recurrence quantification. Procedia Comput Sci 6:375–378View ArticleGoogle Scholar
 Tang X, Xia L, Liao Y, Liu W, Peng Y, Gao T, Zeng Y (2013) New Approach to epileptic diagnosis using visibility graph of highfrequency signal. Clin EEG Neurosci 44:1–5View ArticleGoogle Scholar
 Wassernman L (2013) All of statistics a concise course in statistical inference. Springer, BerlinGoogle Scholar
 Weiss JM, Weiss LD, Silver JK (2015) Neuromuscular junction disorders, easy EMG: a guide to performing nerve conduction studies and electromyography. Elsevier, London. ISBN: 9780323286640Google Scholar
 Xie HB, Guo T, Bai S, Dokos S (2014) Hybrid soft computing systems for electromyographic signals analysis: a review. Biomed Eng Online 13(8). http://www.biomedicalengineeringonline.com/content/13/1/8. Accessed 16 Jan 2016
 Yousefi J, HamiltonWright A (2014) Characterizing EMG data using machineleaning tools. Comput Biol Med 51:1–13View ArticlePubMedGoogle Scholar
 Zhu G, Li Y, PP Wen (2012) An efficient visibility graph similarity algorithm and its application on sleep stage classification. In: BI 2021 LNCS, vol 7670, pp 185–195
 Zhu G, Li Y, Wen PP (2014) Epileptic seizure detection in EEGs signals using a fast weighted horizontal visibility algorithm. Comput Method Progr Biomed 115:64–74View ArticleGoogle Scholar