Abstract
In this work, an attempt has been made to differentiate muscle non-fatigue and fatigue conditions using sEMG signals and texture representation of the time-frequency images. The sEMG signals are recorded from the biceps brachii muscle of 25 healthy adult volunteers during dynamic fatiguing contraction. The first and last curls of these signals are considered as the non-fatigue and fatigue zones, respectively. These signals are preprocessed and the time-frequency spectrum is computed using short time fourier transform (STFT). Gray-Level Co-occurrence Matrix (GLCM) is extracted from low (15–45 Hz), medium (46–95 Hz) and high (96–150 Hz) frequency bands of the time-frequency images. Further, the features such as contrast, correlation, energy and homogeneity are calculated from the resultant matrices. The results show that the high frequency band based features are able to differentiate non-fatigue and fatigue conditions. The features such as correlation, contrast and homogeneity extracted at angles 0°, 45°, 90°, and 135° are found to be distinct with high statistical significance (p < 0.0001). Hence, this framework can be used for analysis of neuromuscular disorders.
1 Introduction
Muscle fatigue is a time-related progressive impairment of maximal force generating capacity of the muscles. It arises due to sustained or intense contraction, Parkinson’s disease, carcinoma, endocrine disturbances, malnutrition and immobilization [1]. Repeated fatigue may lead to irreversible impairment of muscles. Hence, it is necessary to analyse fatigue conditions for the clinical diagnosis of muscle disorders. Surface electromyography (sEMG) is a non-invasive technique which is commonly used to analyse muscle fatigue [2].
sEMG is a complex bio-electric signal which represents the contraction of the muscle in the body. It offers useful information to understand the human movement which helps in the assessment of muscular activation and internal loads on muscles, tendons, and other tissues. These signals are random, non-stationary and multi-component in nature [3].
The main component of fatigue analysis is the identification of prominent features of the signal [1]. Various methods have been proposed in the literature based on the extraction of sEMG features in the time [4], frequency [5], and t-f domain [2], [3], [6]. The time domain features contain amplitude, rhythmicity and entropy information. The frequency domain features includes spectrum normalized power, frequency sub-band powers and mean frequency. In the t-f domain, the features are extracted from the time-frequency representation of sEMG signals and are capable of characterizing the non-stationary and multi-component nature of sEMG signals [3]. These include instantaneous frequency and sub-band energies.
Recently, time-frequency spectrum computed from the EEG signals is considered as images and the features extracted are used for the automatic detection of epileptic seizure in EEG data [7], [8], [9], [10], [11], [12]. These features includes Haralick features [8], texture features such as first order moment, second order moment [10], GLCM [7], [10], [12], texture feature coding method [4] and local binary pattern [7], [9] and Histogram features such as mean, variance, skewness and kurtosis [11].
In this work, sEMG signals are recorded from biceps brachii muscles in bipolar configuration. These signals are represented in t-f domain using the spectrogram of STFT. The corresponding images obtained are subdivided into three images based on the frequency bands and converted into 8- bit grayscale images. Finally, GLCM texture features are extracted and sEMG signals are analysed.
2 Methodology
Signals are acquired from 25 normal subjects with no history of neuromuscular problems. The experiment is carried out on the biceps. Ag–AgCl disc-type disposable surface electrodes are placed on the belly of the muscle in differential electrode configuration, with the distance between the electrodes equal to 3 cm. The reference electrode is placed at the proximal end of the elbow. The subjects are made to stand on a wooden platform to provide isolation. The subjects perform repetitive flexion and extension of the elbow with a 6 kg load until they experiences fatigue [3]. These signals are recorded at a sampling rate of 1000 Hz. The sEMG signals corresponding to the first and the last curls are used in this study.
2.1 STFT spectrogram
STFT is a commonly used method for spectrogram image formation [2]. Each signal is divided into smaller segments and subjected to discrete fourier transform (DFT). The spectrum values from each segment are stacked side-by-side to form the spectrogram image. The spectrogram shows dominant frequency information against time where the frequency components are equally spaced along the vertical axis with constant bandwidth.
By definition, STFT spectrogram is the normalized, squared magnitude of the STFT coefficients. STFT coefficients for a signal is calculated as [7]
where x[m]w[n − m] is a short time part of input signal x[m] at time n.
The frequency range of sEMG varies from dc to 10 kHz with the dominant frequency content lying in the range of 50–150 Hz [3]. Each spectrogram is divided into three sub-images corresponding to the dominant frequency bands as follows: Low frequency band (15–45 Hz), Medium frequency band (46–95 Hz) and High frequency band (96–150 Hz).
2.2 GLCM features
The sub images are converted into 8-bit grayscale images and the texture features are extracted. GLCM is a method to analyse the texture images which estimates image properties related to second-order statistics [10]. This corresponds to a directional pattern counter with a specific distance d and angle θ between neighbouring pixel pairs for grayscale images [7]. The distance parameter d is set to 1 and the angle parameters θ are 0°, 45°, 90°, and 135°. The corresponding displacement vectors are [0 1], [−1 1], [−1 0] and [−1 −1]. The features such as contrast, correlation, energy and homogeneity are extracted and the formula is given below:
where p(i, j) is the intensity value of the pixel at point (i, j), μ and σ represents the mean and standard deviation.
Contrast is a measure of the intensity variation between a pixel and its neighbour over the whole image. Correlation explains the relation between a pixel and its neighbour over the whole image while energy is the sum of squared elements in the GLCM. Homogeneity calculates the closeness of the distribution of elements in the GLCM to its diagonal.
3 Results and discussion
The representative sEMG signals recorded during the experiment is shown in Figure 1.
The amplitude and frequency components of the recorded signals are found to be subject dependent. The amplitude is observed to be higher in fatigue condition. This may be due to the participation of more motor units during the contraction. Similar variations are observed in most of the cases.
The spectrogram of sEMG signals has been represented as a t–f image. Figure 2 shows the spectrogram of sEMG signals for non-fatigue and fatigue conditions. The colormap represents the strength of the magnitude of sEMG signals. The variations in the spectral components are associated with physiological parameters such as random firing rate, motor unit recruitment patterns, muscle fibre conduction velocity and volume conductor effects. A reduction in the frequency components is observed in fatigue condition and this may be due to the synchronisation of motor units.
It is observed that, during non-fatigue conditions, the texture pattern exhibits brighter pixels with more variations compared to fatigue conditions. In order to discriminate between these texture patterns, GLCM features are employed. The t-f images are converted to 8-bit grayscale images. For each of these images, the angle parameter is varied from 0° to 135° with a 45° increment. Thus four GLCMs are obtained for each image. Features such as contrast, correlation, energy and homogeneity are extracted for each angle thereby constructing a 16-dimensional feature vector for each image. The average value of these features is shown in Table 1. The class discrimination ability of the feature sets is quantified using t-test. It is observed that only the energy feature calculated from all the angles is distinct with high statistical significance (p < 0.0001).
Features | Non-fatigue | Fatigue | p-Value | |
---|---|---|---|---|
Contrast | 0 | 6.1236 | 4.5720 | 0.0588 |
45 | 59.5267 | 40.9858 | 0.0265 | |
90 | 53.8319 | 36.7704 | 0.0287 | |
135 | 59.4551 | 41.0085 | 0.0270 | |
Correlation | 0 | 0.9959 | 0.9957 | 0.8075 |
45 | 0.9627 | 0.9638 | 0.2671 | |
90 | 0.9665 | 0.9677 | 0.2209 | |
135 | 0.9627 | 0.9637 | 0.2961 | |
Energy | 0 | 0.4098 | 0.5906 | <0.0001 |
45 | 0.3972 | 0.5821 | <0.0001 | |
90 | 0.4000 | 0.5843 | <0.0001 | |
135 | 0.3972 | 0.5821 | <0.0001 | |
Homogeneity | 0 | 0.9494 | 0.9670 | 0.0033 |
45 | 0.8758 | 0.9219 | 0.0003 | |
90 | 0.8899 | 0.9312 | 0.0003 | |
135 | 0.8758 | 0.9218 | 0.0003 |
Figure 3 shows the 8-bit grayscale sub-image representation of t-f images. For each of these sub-images, the GLCM features are extracted. Due to the non-uniform nature of the sEMG spectrograms, this local feature extraction gives better results than global features.
The average value of the features extracted from the high frequency band is shown in Table 2. Among the three frequency bands considered, the high frequency band is able to discriminate non-fatigue and fatigue conditions. The three features namely correlation, energy and homogeneity extracted from this band in all the four angles is found to be highly significant (p < 0.0001). There is a greater value of energy for fatigue condition compared to non-fatigue condition indicating reduction in the action potential of the muscles. The higher correlation for non-fatigue condition indicates normal muscle activity.
Features | Non-fatigue | Fatigue | p-Value (less than) | |
---|---|---|---|---|
Contrast | 0 | 5.8881 | 7.0186 | 0.1355 |
45 | 20.3979 | 11.5909 | 0.001 | |
90 | 15.9509 | 5.9354 | 0.001 | |
135 | 20.3952 | 11.5927 | 0.001 | |
Correlation | 0 | 0.9679 | 0.7277 | 0.0001 |
45 | 0.9303 | 0.6082 | 0.0001 | |
90 | 0.9557 | 0.8403 | 0.0001 | |
135 | 0.9302 | 0.6084 | 0.0001 | |
Energy | 0 | 0.6386 | 0.9216 | 0.0001 |
45 | 0.6292 | 0.9199 | 0.0001 | |
90 | 0.9679 | 0.7277 | 0.0001 | |
135 | 0.9303 | 0.6082 | 0.0001 | |
Homogeneity | 0 | 0.9557 | 0.8403 | 0.0001 |
45 | 0.9302 | 0.6084 | 0.0001 | |
90 | 0.6386 | 0.9216 | 0.0001 | |
135 | 0.6292 | 0.9199 | 0.0001 |
4 Conclusion
In this work, second order texture statistics have been used to analyse muscle non-fatigue and fatigue conditions after mapping the sEMG signals into a 2D grayscale image. The time-frequency spectrum of the signals is computed using STFT. Each spectrum is divided into three frequency bands such as low, medium and high. GLCM features such as contrast, correlation, energy and homogeneity are calculated for different angles from these sub-images. The visual results suggest that the non-fatigue and fatigue conditions have different spectral patterns. The features calculated from the sub band images carry more information than the whole spectral images. The texture features namely correlation, energy and homogeneity calculated from the high frequency sub images gives significant results compared to other frequency sub images (p < 0.0001). Hence, it appears that time-frequency images based texture features could be used for analysis of neuromuscular disorders.
Author’s Statement
Research funding: The author state no funding involved. Conflict of interest: Authors state no conflict of interest. Material and Methods: Informed consent: Informed consent is not applicable. Ethical approval: The conducted research is not related to either human or animal use.
References
[1] Venugopal G, Navaneethakrishna M, Ramakrishnan S. Extraction and analysis of multiple time window features associated with muscle fatigue conditions using sEMG signals. Expert Syst Appl. 2014;41:2652–9.10.1016/j.eswa.2013.11.009Search in Google Scholar
[2] Zawawi TNST, Abdullah AR, Shair EF, Halim I, Rawaida O. Electromyography signal analysis using spectrogram. IEEE Student Conference on Research and Development (SCOReD); 2013. p. 319–24.10.1109/SCOReD.2013.7002599Search in Google Scholar
[3] Karthick PA, Ramakrishnan S. Surface electromyography dased muscle fatigue progression analysis using modified B distribution time-frequency features. Biomed Signal Proces 2015;2:42–51.Search in Google Scholar
[4] Abdullah RS, Pinar A. Detection of surface electromyography recording time interval without muscle fatigue effect for biceps brachii muscle during maximum voluntary contraction. J Electromyogr Kines. 2010;20:773–6.10.1016/j.jelekin.2010.02.006Search in Google Scholar PubMed
[5] Gonzalez IM, Malanda A, Navarro-Amezqueta I, Gorostiaga EM, Mallor F, Ibanez J. EMG spectral indices and muscle power fatigue during dynamic contractions. J Electromyogr Kines. 2010;20:233–40.10.1016/j.jelekin.2009.03.011Search in Google Scholar PubMed
[6] Yousefi H, Askari S, Dumont GA, Bastany Z. Automated decomposition of needle EMG signal using STFT and Wavelet Transforms. Iranian Conference on Biomedical Engineering (ICBME). 2014;358–63.10.1109/ICBME.2014.7043951Search in Google Scholar
[7] Sengu A, Guo Y, Akbulut Y. Time-frequency texture descriptors of EEG signals for ef?cient detection of epileptic seizure. Brain Informat. 2015;1–8.Search in Google Scholar
[8] Boubchir L, Al-Maadeed S, Bouridane A. Haralick feature extraction from time-frequency images for epileptic seizure detection and classification of EEG data. International Conference on Microelectronics (ICM). 2014;32–35.10.1109/ICM.2014.7071799Search in Google Scholar
[9] Boubchir L, Al-Maadeed S, Bouridane A, Cherif A. Classification of EEG signals for detection of epileptic seizure activities based on LBP descriptor of time-frequency images. International Conference on Image Processing. 2015;3758–62.10.1109/ICIP.2015.7351507Search in Google Scholar
[10] Boubchir L, Al-Maadeed S, Bouridane A, Cherif A. Time-frequency image descriptors–based features for EEG epileptic seizure activities detection and classification. International Conference on Acoustics, Speech, and Signal Processing (ICASSP). 2015;867–71.10.1109/ICASSP.2015.7178093Search in Google Scholar
[11] Fu K, Qu J, Chai Y, Dong Y. Classification of seizure based on the time-frequency image of EEG signals using HHT and SVM. Biomed Signal Proces. 2014;13:15–22.10.1016/j.bspc.2014.03.007Search in Google Scholar
[12] Samiee K, Kiranyaz S, Gabbouj M, Saramaki T. Long-term epileptic EEG classification via 2D mapping and textural Features. Expert Syst Appl. 2015;42:7175–85.10.1016/j.eswa.2015.05.002Search in Google Scholar
©2016 N. Punitha et al., licensee De Gruyter.
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.