Portal de Periódicos da CAPES

Variational mode decomposition based differentiation of fatigue conditions in muscles using surface electromyography signals

2020; Institution of Engineering and Technology; Volume: 14; Issue: 10 Linguagem: Inglês

10.1049/iet-spr.2020.0315

1751-9683

Divya Bharathi Krishnamani, Alagar Karthick, Ramakrishnan Swaminathan,

EEG and Brain-Computer Interfaces

IET Signal ProcessingVolume 14, Issue 10 p. 745-753 Research ArticleFree Access Variational mode decomposition based differentiation of fatigue conditions in muscles using surface electromyography signals Divya Bharathi Krishnamani, Corresponding Author Divya Bharathi Krishnamani divyak0593@gmail.com Non-Invasive Imaging and Diagnostics Laboratory, Biomedical Engineering Group, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai, IndiaSearch for more papers by this authorKarthick P.A., Karthick P.A. Physiological Measurements and Instrumentation Laboratory, Department of Instrumentation and Control Engineering, National Institute of Technology Tiruchirappalli, Tiruchirappalli, IndiaSearch for more papers by this authorRamakrishnan Swaminathan, Ramakrishnan Swaminathan Non-Invasive Imaging and Diagnostics Laboratory, Biomedical Engineering Group, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai, IndiaSearch for more papers by this author Divya Bharathi Krishnamani, Corresponding Author Divya Bharathi Krishnamani divyak0593@gmail.com Non-Invasive Imaging and Diagnostics Laboratory, Biomedical Engineering Group, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai, IndiaSearch for more papers by this authorKarthick P.A., Karthick P.A. Physiological Measurements and Instrumentation Laboratory, Department of Instrumentation and Control Engineering, National Institute of Technology Tiruchirappalli, Tiruchirappalli, IndiaSearch for more papers by this authorRamakrishnan Swaminathan, Ramakrishnan Swaminathan Non-Invasive Imaging and Diagnostics Laboratory, Biomedical Engineering Group, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai, IndiaSearch for more papers by this author First published: 11 February 2021 https://doi.org/10.1049/iet-spr.2020.0315AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinked InRedditWechat Abstract Surface electromyography (sEMG) signals are stochastic, multicomponent and non-stationary, and therefore their interpretation is challenging. In this study, an attempt has been made to develop an automated muscle fatigue detection system using variational mode decomposition (VMD) features of sEMG signals and random forest classifier. The sEMG signals are acquired from 103 healthy volunteers during isometric (45 subjects) and dynamic (58 subjects) muscle fatiguing contractions and preprocessed. The band-limited intrinsic mode functions (BLIMFs) are extracted from non-fatigue and fatigue segments of the signals using the VMD algorithm. Hjorth features, such as activity, mobility and complexity are extracted from each BLIMF and are given to the random forest classifier. The performance of these features is evaluated using leave-one-subject-out cross-validation. The results show that the complexity feature performs better than others and it has resulted in an accuracy of 83% in dynamic contractions and 80% in isometric contractions. The performance is increased by about 8% in a dynamic condition when the most significant complexity features (p < 0.001) are used and by about 12% for isometric when the authors use all significant features. Therefore, the proposed approach could be used to detect fatigue conditions in various neuromuscular activities and real-time monitoring in the workplace. 1 Introduction Biomedical signals represent the physiological conditions of various organs and tissues in the human body. The research on these signals aims in identifying and extracting reliable biomarkers that uncover the functional state of the underlying human system. The analysis of biosignals plays a vital role in various fields of health care, such as diagnostics, therapeutics, and rehabilitation. The recent advancements in machine learning allow researchers and scientists to develop a decision support system using the extracted features. In recent years, several advanced signal processing techniques and machine learning algorithms have been reported in the literature [1, 2]. However, the extraction of prominent, reliable features still considered as a complex and challenging task due to the inherent randomness, nonstationary and multicomponent variations associated with the biosignals. The surface electromyography (sEMG) is a non-invasive technique that records the electrical activity of contraction of skeletal muscles. These signals find a wide range of applications in the field of rehabilitation, ergonomics, exercise physiology, sports biomechanics and biofeedback [3]. Muscle fatigue is a neuromuscular condition that is closely associated with all the aforementioned fields. It is a condition in which the muscles fail to contract and generate the required or expected force. It often results from prolonged sustained muscle contraction or unhealthy work practice. The repeated muscle fatigue can cause injury and sometimes, it permanently damages the muscles [4]. Therefore, it is important to monitor and detect muscle fatigue condition. sEMG signals are widely used for the assessment of muscle fatigue because of its non-invasiveness [5]. Several studies have analysed sEMG signals under isometric contractions which finds application in various sports exercises and physical rehabilitation [6]. In isometric contractions, force is produced by maintaining the joint angle and muscle length constant. For such contractions, analysis is made by assuming the sEMG signal to be stationary. The amplitude and spectral parameters have shown to reflect changes in the sEMG signals during this contraction. Various time and frequency domain methods have been used to analyse the characteristic changes in these signals [7-9]. On the other hand, dynamic contractions play a vital role in real-life activities such as locomotion and lifting the load. sEMG signals are found to be complex, multicomponent and highly nonstationary during this type of contraction [10, 11]. The non-stationary nature causes increased variability in amplitude and frequency parameters of the sEMG signal. This arises due to the physiological factors such as recruitment and de-recruitment of motor units that are active in the area of detecting electrodes, synchronisation of motor units within the muscle, movement of the innervation zone in relation to the surface electrodes as there is the change in muscle shape [1, 12]. Various time–frequency methods such as Stockwell transform, Cohen class distribution have also been used to address the multicomponent and non-stationarity nature of sEMG signals during dynamic contractions [13-15]. Empirical mode decomposition (EMD) proposed by [16] is another widely used time-frequency method to analyse non-stationarity of time series. It decomposes the signal into a finite number of intrinsic mode function (IMF) and a residual term. Each of these IMFs is subjected to Hilbert transform to obtain instantaneous frequency and amplitude [17, 18]. Various studies have utilised the EMD-based time–frequency approach for the detection of muscle fatigue condition from the sEMG signals [19, 20]. However, this method is reported to be more sensitive to noise and sampling. Variational mode decomposition (VMD) is an extended version of EMD which is robust to sampling and noise. It adaptively decomposes the signal into a set of band-limited oscillations called modes [21]. VMD eliminates the mode mixing problems in EMD by using an optimisation method to extract the modes localised on central frequencies [22]. VMD and EMD have been applied in signals corrupted with white Gaussian noise and it is found that VMD gives better performance compared with EMD [23]. In addition, VMD is proven to be effective for removing noise in non-linear and non-stationary signals [23]. This work utilises VMD-based time–frequency analysis to assess the spectral band variations in fatigue condition for both isometric and dynamic contractions. Several classification algorithms have been reported for automated detection of the muscle fatigue condition from the features of sEMG signals. This paper explores VMD features for classifying non-fatigue and fatigue conditions. Classifiers namely, k-nearest neighbour (k-NN), logistic regression (LR), Naïve-Bayes (NB), support vector machine (SVM) are employed in the literature [1, 24]. The k-NN is the simplest classification technique which is a type of instance-based learning or lazy learning. Classification algorithms such as NB and LR are probabilistic. SVM performs non-linear mapping of the input feature vectors to a high-dimensional space using kernel trick method. Random forest (RF) is an ensemble machine learning technique that avoids overfitting of data. Recently, it has been used for detecting myopathic conditions and congestive heart failure from the characteristics of sEMG, ECG time series, respectively [25, 26]. In this study, RF classifier is adapted for muscle fatigue detection. In this work, sEMG signals are acquired from biceps brachii muscles in both isometric and dynamic contraction using the standard experimental protocol. These signals are subjected to VMD and a fixed number of band-limited intrinsic mode functions (BLIMFs) are obtained. Features such as activity, mobility and complexity are extracted from each BLIMF for non-fatigue and fatigue conditions. Finally, these features are evaluated using the performance of a classifier. 2 Methods 2.1 Signal acquisition and preprocessing sEMG signals are recorded from biceps brachii muscle using BIOPAC MP36 acquisition system [1, 27]. This system has a signal-to-noise ratio of 89 dB and common-mode rejection ratio of 110 dB. The sampling rate and gain used for the acquisition are 10 kHz and 1000, respectively. One hundred and three untrained healthy adult volunteers participated in this study out of which 45 subjects with the mean age of 23.5 years performed isometric contraction. Fifty-eight subjects took part in a dynamic contraction whose average age is 26.12 years. The subjects have no history of neuromuscular disorders. They are advised to take complete rest for 12 h prior to the experiment. The procedure is explained to the participants and informed consent is obtained. This experiment follows the principle of the declaration of Helsinki. The recording follows the bipolar configuration by placing two Ag–AgCl electrodes on the muscle belly with an interelectrode distance of 3 cm. The reference electrode is placed on the joint of the elbow. The subjects are instructed to stand erect on the wooden platform to avoid electric shocks [10, 13]. sEMG signals are acquired for the entire duration of the exercise. The upper hand of the subject is kept vertical to the ground and close to the body. Among the total participated subjects, 45 are asked to perform an isometric contraction. The subjects are instructed to maintain 90° elbow position with 6 kg dumbbell load under this protocol. The forearm is kept in a supination position when holding the dumbbell. The experiment continues until there is a drop of 10° angle or when the subject is unable to hold the load. The rest of the subjects are asked to perform dynamic contractions. They are instructed to perform continuous biceps curl exercise with 6 kg dumbbell load until they experience fatigue or unable to continue the exercise. The subjects are asked to maintain the curl speed at their comfortable pace. Instructions are constantly given to avoid arm movements during acquisition [10]. Studies on sEMG signals analysis have shown that the use of an optimal segment length of 500 ms is effective [2]. For the signals acquired from isometric contraction, first and last 500 ms are considered as non-fatigue and fatigue segment. On the other hand, for dynamic contraction, the sEMG signals associated with the muscle non-fatigue and fatigue conditions are extracted from the first and last curl of the exercise. The peak of the first and last curl is identified using the moving average root mean square value and a segment of 500 ms is extracted before and after the peak for the analysis [13]. The motion artefacts, high-frequency noises and power line interference are removed using a band-pass filter with the frequency range of 10–400 Hz and a notch filter of 50 Hz [1]. The study consists of extracting BLIMFs from non-fatigue and fatigue segments using VMD. Then, Hjorth parameters such as activity, mobility and complexity are extracted from each BLIMF and non-fatigue and fatigue segments are classified using random forest classifier. The description of the workflow is depicted in Fig. 1. It includes the sEMG signal acquisition, preprocessing of signals, non-fatigue and fatigue segmentation, computation of BLIMFs using VMD algorithm, extraction of Hjorth parameters and development of the random forest-based model and its performance evaluation. Fig. 1Open in figure viewerPowerPoint Automated detection system of muscle fatigue condition 2.2 Variational mode decomposition VMD decomposes a real-valued input signal x(t) into a discrete number of modes φk that have specific sparsity properties while reproducing the input. Each BLIMF is compact around a centre frequency ωk, and its bandwidth is estimated using H1 Gaussian smoothness of the shifted signal. In short, VMD represents a signal x(t) with modes localised on the centre frequency [22]. The constrained variational optimisation problem using alternate direction method of multipliers (ADMM) can be expressed as given below: min φ k , ω k ∑ k = 1 K ∂ t δ t + j π t ∗ φ k t e − j ω k t 2 2 s . t . ∑ k = 1 K φ k t = x ( t ) (1) where K represents the number of modes, {φk} = {φ1, …, φK}, {ωk} = {ω1, …, ωK} are the notations for the set of modes and centre frequencies, respectively, δ is the Dirac distribution, * and ∂ denote the convolution and partial differential operators, respectively. The modes φk(t) in the frequency domain are estimated using ADMM in the form of the Wiener filter structure as given below: φ ^ k ω = φ ^ ω − ∑ i ≠ k φ ^ i ω + ( λ ^ ( ω ) / 2 ) 1 + 2 α ω − ω k 2 (2) where α represents the variance of white noise and ^ represents the Fourier transform of the components. Finally, the modes are obtained in the time domain by computation of the inverse Fourier transform of the filtered signal, and the centre frequencies are estimated by the following equation: ω k = ∫ 0 ∞ ω | φ ^ k ω | 2 d ω ∫ 0 ∞ | φ ^ k ω | 2 d ω (3) The modes having localised centre frequency properties are viable for analysing the sub-components of a signal. The number of modes can be selected due to the aforementioned mathematical background. Consequently, VMD has an advantage in decomposition and in adjusting the number of modes for further analysis. In this study, the number of BLIMFs is empirically chosen to be 15. The frequency resolution and computational load depend on the choice of the number of modes. The frequency resolution will be higher for the higher number of modes. 2.3 Feature extraction Hjorth parameters usually describe the statistical properties of a signal. This signal analysis proposed by Hjorth [28] has three kinds of parameters, such as activity, mobility and complexity. Activity is the measure of the squared standard deviation of the amplitude and is referred to as mean power or variance. Mobility is given by the standard deviation of the slope with reference to the standard deviation of the amplitude while complexity is expressed as the number of standard slopes actually generated during the average time required for the generation of one standard amplitude [29-31]. Hjorth parameters are used in this study to analyse the statistical properties of BLIMFs in non-fatigue and fatigue conditions. The expression of these features extracted from BLIMF is given in (4)–(6), where φ represents the BLIMF and i corresponds to BLIMF number. Activit y i = 1 N ∑ n = 0 N − 1 φ i n − φ i ¯ 2 (4) Mobilit y i = Activity d φ i / d t Activit y i (5) Complexit y i = Mobility d φ i / d t Mobilit y i (6) The significance of the extracted features is evaluated using the Wilcoxon rank-sum test. It is a non-parametric test for finding the equality between two groups by comparing their medians. The statistical significance is evaluated using a p-value which represents the probability. A very small value of p indicates the significant difference between the two classes. In this study, p < 0.001 is considered for evaluating statistical significance [32]. 2.4 Classification Random forest classifier proposed by [33] is considered in this work for differentiating non-fatigue and fatigue conditions. It is a powerful classifier and an ensemble learning technique widely used in various studies. Every tree can be regarded as an individual classifier and the classification output is voted by all the decision trees. In this ensemble technique, each decision tree is trained with different subsets of the training data using random samples drawn with replacement from the original training set by means of Bootstrap aggregating. Therefore, data may be used more than once in the training process of these classifiers in some instances. This phenomenon increases the stability and robustness of the classifier for the data that are with slight variations. In other words, a random vector independent from the previous one is produced and distributed to all trees in this classifier. Then, each tree is grown using random vector and training set thus resulting in a collection of tree-structured classifiers at the input vector. The generalisation error is given by margin function that measures the extent to which the average number of votes at random vectors for the right output exceeds the average vote for any other output. The parameters strength and correlation measure accuracy of individual classifier and dependence between classifiers, respectively. RF classifier with random features is formed by selecting a small group of input variables on each node, randomly. The generalisation error in RF classifier is given below: E = P X , Y m g ( X , Y ) < 0 (7) where subscripts X and Y are random vectors which indicate the probability is over X, Y space and mg is the margin function that measures the degree by which the average number of votes for the right output exceeds any other output [26]. In this study, the number of decision trees is set to 30 for the RF classifier. In addition, leave-one-subject-out cross-validation is employed to evaluate the performance of the classifier. 3 Results and discussions Figs. 2 and 3 show the representative sEMG signals of non-fatigue and fatigue conditions for both the isometric and dynamic contractions, respectively. These signals were recorded from age-matched volunteers. In general, the frequency components of sEMG vary with time and these variations are random and non-deterministic. In order to address these nonstationary behaviours, the sEMG signals are decomposed into a fixed number of BLIMFs using the VMD algorithm. Fig. 2Open in figure viewerPowerPoint Representative sEMG signals recorded during isometric contraction (a) Non-fatigue condition, (b) Fatigue condition Fig. 3Open in figure viewerPowerPoint Representative sEMG signals recorded during dynamic contraction (a) Non-fatigue condition, (b) Fatigue condition The BLIMFs and their corresponding spectrums for both the muscle non-fatigue and fatigue conditions are shown in Figs. 4 and 5, respectively. The magnitude of the spectrum is measured in mV/Hz. From the figures, it can be observed that the BLIMFs are arranged from lowest to highest frequency components. Fig. 4Open in figure viewerPowerPoint VMD of the representative signal recorded in dynamic contractions (a) BLIMFs of sEMG from non-fatigue condition, (b) Corresponding spectrums Fig. 5Open in figure viewerPowerPoint VMD of the representative signal recorded in dynamic contractions (a) BLIMFs of sEMG from fatigue condition, (b) Corresponding spectrums Fig. 6 shows the variation of activity, mobility and complexity of BLIMFs in muscle non-fatigue and fatiguing dynamic contractions. All these features are found to be different in both the conditions across the modes. From Fig. 6a, it is observed that activity is higher in fatigue condition. The large difference is noticed at BLIMF2. The maximum percent change in the mean value of an activity is 400% in BLIMF2. Then, the percent change decreases from BLIMF2 to BLIMF15. On the other hand, mobility is higher in non-fatigue than fatigue conditions and is shown in Fig. 6b. It is also to be noted that mobility increases from BLIMF1 to BLIMF15 irrespective of non-fatigue or fatigue condition. It indicates that high-frequency components have higher mobility. The maximum difference in the mean value is observed in BLIMF7. Fig. 6c shows the variations of complexity in the muscle non-fatigue and fatigue conditions for all the modes. The complexity value extracted from the non-fatigue condition is higher for the first eight BLIMFs whereas it is lower for the other BLIMFs. Unlike activity and mobility, there is no consistent trend across BLIMFs is observed in the complexity feature. The large difference in the mean value of complexity is observed in BLIMF13. Fig. 6Open in figure viewerPowerPoint Variation of VMD features in dynamic contractions (a) Activity, (b) Mobility, (c) Complexity Fig. 7 demonstrates the result of isometric muscle fatiguing contraction. The variations of activity and mobility for all modes are shown in Figs. 7a and b, respectively. As in the case of dynamic contraction, the activity is higher and mobility is lower in fatigue conditions. Fig. 7c shows the complexity variations across different modes. In isometric contraction, the complexity is higher in first five BLIMFs in non-fatigue conditions and is lower in the other BLIMFs. However, the percent change in the mean value of the extracted feature is considerably lower than the dynamic contractions. It reveals that the activity and mobility that are extracted from BLIMFs are more sensitive in detecting the fatigue conditions associated with dynamic contractions. In dynamic contractions, activity extracted from BLIMF1-7,9,15, mobility associated with BLIMF2-14, and complexity computed from BLIMF1-5,13-15 are found to have significant differences (p < 0.001) between non-fatigue and fatigue conditions. In the case of isometric contraction, only a smaller number of features exhibit a significant difference (p < 0.001) between the two conditions. Activity from BLIMF1,15, mobility from BLIMF12,13 and complexity from BLIMF10,12-15 are found to be significant. Fig. 7Open in figure viewerPowerPoint Variation of VMD features in isometric contractions (a) Activity, (b) Mobility, (c) Complexity Table 1 shows the classification performance of features for both the isometric and dynamic contractions. The performance is evaluated based on a feature extracted from all the fifteen BLIMFs. It can be observed that the complexity feature detects the fatigue condition with about 80% accuracy in both dynamic and isometric contraction. The classification performance of other features, namely, activity and mobility are comparatively lower than complexity. It is important to notice that the sensitivity is more than 80% for both the muscle contractions when we use complexity feature. In addition, the positive predictive value (PPV) and negative predictive value (NPV) are also higher in the classification model based on complexity measure. In general, feature selection algorithms are employed to simplify the models by removing/reducing the irrelevant features and to avoid the curse of dimensionality in order to further enhance the classification results. Table 1. Classification performance of extracted features in muscle fatigue and non-fatigue conditions for both the contractions Features Sensitivity Specificity Accuracy PPV NPV Dynamic contractions activity 72.41 86.20 79.31 84.00 75.75 mobility 75.86 75.86 75.86 75.86 75.86 complexity 82.75 84.48 83.62 84.21 83.05 Isometric contractions activity 75.55 71.11 73.33 72.34 74.41 mobility 73.33 77.77 75.55 76.74 74.46 complexity 84.44 75.55 80.00 77.55 82.92 The results of the model based on features that exhibit a significant difference (p < 0.001) between the muscle non-fatigue and fatigue conditions are presented in Table 2. It can be noticed that the accuracy of about 80% in the activity-based model, 77% in the mobility-based model and 91% in complexity based model is achieved in dynamic fatiguing contractions. The accuracy is improved by 7% when we use only the eight selected complexity features (BLIMF1-5,13-15). The other parameters, namely PPV and NPV are also found to have a considerable improvement in model based on complexity features. It is important to note that only eight features BLIMF1-5,13-15 found to have a significant difference (p < 0.001). Table 2. Classification performance of model based on features with significant difference (p < 0.001) in muscle fatigue and non-fatigue conditions for both the contractions Features Sensitivity Specificity Accuracy PPV NPV Dynamic contractions activity 74.13 86.20 80.17 84.31 76.92 mobility 75.86 77.58 76.72 77.19 76.27 complexity 90.38 92.30 91.34 92.15 90.56 Isometric contractions activity 84.44 88.88 86.66 88.37 85.10 mobility 55.55 64.44 60.00 60.97 59.18 complexity 80.00 77.77 78.88 78.26 79.54 all features (significant) 91.11 93.33 92.22 93.18 91.30 In isometric contractions, among all significant features, the model based on activity from BLIMF1,15 is observed to have considerable improvement. The accuracy is improved by around 13%, PPV is about 16% and NPV is about 10% when we use the significant activity feature. Since only a smaller number of features exhibits a significant difference in isometric contractions, all these features are given to random forest. It is noted that the accuracy is further improved by 5%. The maximum accuracy of 92% (sensitivity: 91% specificity: 93%), maximum PPV of 93% and maximum NPV of 91% is achieved in the model based on features with a significant difference. Fig. 8 depicts the information on time consumption of the full and reduced feature set for isometric and dynamic contractions. The time consumption for the computation of activity, mobility and complexity from all the fifteen BLIMFs is found to be approximately 27, 31, and 37 ms, respectively. In the case of the model with the reduced feature set, the computation time is decreased by 30% for activity, 4% for mobility and 15% for complexity. The time consumption is low for the model with a reduced feature set in the detection of isometric fatiguing contractions. This experiment is carried out in the Matlab 2020 environment using a computer with Intel Core i5 processor running at 1.60 GHz. Fig. 8Open in figure viewerPowerPoint Comparison of computational time for the full and reduced feature set in dynamic and isometric contractions Table 3 shows the comparison of studies related to the detection of muscle fatigue conditions in both isometric and dynamic conditions. It includes the details of the number of participants involved in the study, feature extraction methods, classification algorithms, cross-validation techniques and detection accuracy. It can be seen that the combination of several signal processing methods and machine learning algorithms have been employed for the detection of this neuromuscular condition from the sEMG signals. These signal analysis methods aimed to characterise and utilise the stationary, nonstationary and nonlinear variations of myoelectric signals. Various time–frequency approaches such as short-time Fourier transform (STFT), smoothed pseudo-Wigner–Ville distribution (SPWVD), and continuous wavelet transform (CWT) were used to analyse the time-varying frequency characteristics of sEMG in muscle fatigue condition. It was found that the multilayer perceptron neural network (MLPNN) based on CWT features performed better with a classification accuracy of 91% [34]. Another study showed that a genetic algorithm (GA)-based pseudo wavelet function detects the muscle fatigue and non-fatigue segments using linear discriminant analysis (LDA) at the classification rate of 88.4% [35]. Recently, the applicability of feature, mean absolute value slope computed from the multiple time windows (MTW) namely, Hamming, trapezoidal and Slepian window were evaluated. The measures selected by information gain (IG) along with k-NN algorithm resulted in a 93% correct detection rate [1]. The aforementioned studies were based on sEMG associated with isometric fatiguing contractions. Table 3. Comparison of studies on automated detection of muscle fatigue conditions Authors No. of subjects involved Methods Classification algorithms Validation technique Fatiguing contraction type Max. accuracy, % Subasi and Kiymikaa Represents the study based on different dat