Using the frequency signature to detect muscular activity in weak and noisy myoelectric signals

https://doi.org/10.1016/j.bspc.2019.02.026Get rights and content

Highlights

  • The method is based on the frequency characteristics of the weak and noisy EMG signals.

  • A clustering approach is proposed to detect muscular activity.

  • The method was tested on a set of simulated EMG data.

Abstract

The detection of muscular activity for signals characterized by low amplitude and low signal-to-noise ratio – weak and noisy – is a challenge in biomedical data processing. The aim of this paper is to introduce a method based only on the frequency characteristics of the weak and noisy EMG to detect muscular activity. The algorithm is window-based and consists of two processing steps: i) estimation of zero-crossings and mean instantaneous frequency of the signal; ii) clustering by a k-means approach to separate the muscular activity from the silent phases. We assessed the method on 320 simulated EMG signals that have been generated from a small number of synthetic motor units working at a low firing rate and then manipulated by adding Gaussian noise to simulate four different levels of low signal-to-noise ratio (SNR). Tests were carried on by changing the window dimension – fifteen different window lengths – and the amount of overlap of the window along the signal – four different values of overlapping. The performance of the algorithm was evaluated by calculating the temporal bias of the onset detection, the percentage error made when estimating the activity duration, and the F1 score as a measure of accuracy. The results showed that the algorithm performance does not depend from SNR but depends on both window length and overlap. The detection accuracy ranges from 96% to 98% depending on combinations of window length and overlap, while for specific combinations of window length and overlap, the amount of temporal bias fell below 20 ms. These results open promising scenarios for the application of this algorithm to real weak and noisy EMG data.

Introduction

The detection of skeletal muscle activation is a critical issue in myoelectric signal (EMG) processing for clinical [1], motor control [2], ergonomics [3] and sports applications [4]. Different approaches were developed to estimate EMG onset–offset based on several automatic computer-based algorithms, including: single-threshold [5] and double-threshold [6] detection, adaptive threshold method [7], advanced statistical procedures [8], artificial neural network [9] and fuzzy logic [10] techniques. Unfortunately, these methods often fail when the muscle activity is recorded by damaged (i.e amputation) [12], pathological muscles (i.e myopathology) [13,14], or by complex muscles (i.e upper trapezius) during sustained and prolonged activity [15]. In these situations, the signals are characterized by a low activity level due to low firing rate, low number of motor units recruited, low activation threshold and very low signal-noise ratio (SNR) [11].

For these weak and noisy [19] EMG signals, several approaches have been recently proposed to improve SNR and to minimize erroneous onset detection [16]: Xu et al. [17] developed an adaptive algorithm for the determination of the onset and offset of the muscle contractions based on the generalized likelihood ratio test (GLR); Merlo et al. [11] proposed a time-frequency analysis to identify the single motor unit activation potentials in noisy EMG, as they used the continuous wavelet transform (CWT), permitting to detect the intervals of activation through an optimal threshold definition when the SNR is low (2 dB); Zhang and Zhou [18], instead, proposed an analysis based on Sample Entropy to distinguish the EMG signal from spurious background spikes. A method based on the non-linear properties of the Teager- Kaiser energy (TKE) operator, applied to surface EMG signal, was proposed for the first time by Li et al [16] considering simultaneously the amplitude and the instantaneous frequency of the signal: it improves SNR and favours a more robust detection of the onset muscle activity. This method was perfected by Yang et al. [19] applying a filtering method that is borrowed from an advanced image enhancement technique in the TKE domain. Solnik and colleagues [20] compared the results obtained by the application of the TKE with the three more-used classical onset methods for detection (visual detection, threshold-based method and approximated generalized likelihood-ratio), showing that the application of TKE operator in signal conditioning significantly reduces the mean detection error with respect to the classical methods.

However, all the mentioned methods depend, either fully (i.e GLR) or partially (i.e TKE or CWT), on the amplitude characteristics of the weak EMG signal, and their performance often depends on the SNR level; this detrimentally affects the detection performance when SNR is less than 10 dB.

To date, despite the frequency analysis of the EMG is widely used to track muscular changes and to evaluate muscle fatigue [21,22], the possibility to extract the intervals of muscular activity only from the changes in frequency domain was only partially investigated: to our knowledge only one preliminary study was proposed in the literature, which uses frequency characteristics such as zero-crossings and mean instantaneous frequency to detect the muscular activity of a real EMG signal, by an empirical threshold approach [23].

In our work the aim is to evaluate the possibility to extract the intervals of muscular activation in weak and noisy signals, where amplitude-based parameters display a reduced efficacy driven by the low SNR characteristics, leveraging on frequency features only. In particular, we developed a cluster-based algorithm which is based on the extraction of the number of zero-crossings and the mean instantaneous frequency [23], which we can comprehensively denote as frequency signature. This kind of approach to muscle activity detection allows to overcome the problems related to the definition of empirical thresholding, which in turns makes comparisons of results among different operators and laboratories harder.

The algorithm was validated on a set of simulated EMG signals and the performance was assessed showing good results in terms of detection and robustness with respect to both noise level and implementation settings.

Section snippets

EMG simulated signals

The EMG signals were simulated using an EMG simulation software based on the model developed by Hamilton-Wright and Stashuk [24,25]. The details of the theoretical bases of the EMG modelling method can be found in [24,25].

The EMG signals were supposed to be generated from a small muscle (100 motor units) with low firing rate (10 pps). The muscular contraction was generated in terms of percentage of the maximum voluntary contraction (5% MVC). The maximum recruitment threshold was set at 50% of

Statistical analysis

For each parameter, descriptive statistics was calculated (mean ± standard deviation) and statistical analysis was done: to verify the normal distribution of data, Kolmogorov-Smirnov test was done. Bias and error time (τ, terror) parameters were then considered separately as dependent variables for a 3-way ANOVA test considering window length (win200, win400, win600,win800, win1000, win1200, win1400, win1600, win1800, win2000, win2200, win 2400, win2600, win3000), overlap (0%, 25%, 50%,75%) and

Results

For both τ and terror the statistical analysis has shown a global effect for window length, overlap, SNR and the interaction between window length and overlap, between window length and SNR and between overlap and SNR. The statistical values are reported in the following Table 2.

Discussion and conclusions

The detection of the muscular activity in signals characterized by low amplitude and low signal-to noise ratio is a challenge in biomedical data processing for application in different fields such as prosthesis control and ergonomics.

To date the main approaches proposed to detect muscular activity are essentially based either on the amplitude or on the combination of both frequency and amplitude characteristics of the signals, and their performance are strictly related to the SNR level.

References (28)

  • H. Freund

    Motor unit and muscle activity in voluntary motor control

    Physiol. Rev.

    (1983)
  • S. Kumar

    Electromyography in Ergonomics

    (2017)
  • J. Vance et al.

    EMG activity as a function of the performer’s focus of attention

    J. Mot. Behav.

    (2004)
  • P. Bonato et al.

    A statistical method for the measurement of muscle activation intervals from surface myoelectric signal during gait

    IEEE Trans. Biomed. Eng.

    (1998)
  • Cited by (8)

    • Percentage estimation of muscular activity of the forearm by means of EMG signals based on the gesture recognized using CNN

      2020, Sensing and Bio-Sensing Research
      Citation Excerpt :

      In general, the applicability of EMG signals is mainly based on muscle activity, that is when the muscle is exercising a work and the intensity of activation that is generated. This topic has several points of focus of use, such as onset detection, i.e. when the activation of the muscle begins when performing a force [16–18], which is described as one of the challenges in the analysis of electromyographic signals, due to the difficulty of estimating the real start of muscle activity, taking into account the noise that may exist in the signal. The study of muscle activity has also been applied in health care, e.g. to analyze body postures and how different muscles interact in these, in order to see how it affects the health of people [19].

    • Generalization of a wavelet-based algorithm to adaptively detect activation intervals in weak and noisy myoelectric signals

      2020, Biomedical Signal Processing and Control
      Citation Excerpt :

      The maximum likelihood method proposed by Xu and colleagues [29] improved the generalized likelihood ratio test with an adaptive threshold technique based on the SNR estimation. D’Anna et al. [30] proposed a clustering-based method that receives as inputs specific frequency characteristics of the EMG signal. The work authored by Merlo and colleagues [31] has drawn the attention of the scientific community, as the method therein proposed showed accurate timing performance for low SNR values (i.e. around 2 dB).

    • Comparison of Intramuscular and Surface Electromyography Recordings towards the Control of Wearable Robots for Incomplete Spinal Cord Injury Rehabilitation

      2020, Proceedings of the IEEE RAS and EMBS International Conference on Biomedical Robotics and Biomechatronics
    View all citing articles on Scopus
    View full text