Low-density surface electromyographic patterns under electrode shift: Characterization and NMF-based classification

https://doi.org/10.1016/j.bspc.2020.101890Get rights and content

Abstract

Electrode shift causes high variability and non-stationarity in surface electromyographic (sEMG) patterns, which seriously impairs the robustness of sEMG-based prosthetic control in daily use. Existing methods for electrode shift are mostly suitable for high-density sEMG configuration and do not work well for low-density sEMG. Nowadays, a quantitative characterization of the influence of electrode shift on low-density sEMG and an effective classification method to handle the influence of electrode shift are still lacking. The present study first designed an experiment to produce electrode shift in an 8-channel sEMG recording system. By using tSNE and other quantitative indices, we observed that rotating electrodes’ position led to great changes in the TD feature space, which subsequently decreased the classification accuracy. Further, we showed that existing algorithms against electrode shift in a high-density electrode configuration have limited effect for low-density sEMG recordings. To combat the influence of electrode shift and to improve the classification accuracy, the nonnegative matrix factorization (NMF) algorithm was used to reduce the non-stationarity of sEMG features and the self-enhancing linear discriminant analysis (SE-LDA) method was adopted to update the classifier based on the changes of sEMG features space. Results showed that the NMF with SE-LDA method achieved an accuracy of 70.58 ± 18.08% for the 10-motion classification problem with electrode shift, which was much higher than the accuracy of 54.84 ± 12.24% achieved by the classical TD features with the LDA classifier. The robust and effective new method against the electrode-shift problem has a great potential for the design of practical sEMG-based prosthetic control.

Introduction

Surface electromyography (sEMG) has been extensively used as a control input for upper-limb prosthesis control, which could largely increase the functional capacity of amputees and improve their quality of life [1]. sEMG-based prosthesis control technology provides a natural mapping from upper-limb muscle motions to prosthesis functions with high accuracy and easy operation [2]. For example, by using classical time-domain (TD) features and the linear discriminant analysis (LDA) classifier, an accuracy of > 90% was achieved in the circumstance of 6 motions [3]. However, the high accuracy of sEMG-based motion recognition obtained in laboratory is in sharp contrast to the high rejection rate in practical uses of intelligent prosthetics [4]. Besides the problems of prosthetic weight, battery life, power [5] etc., the low robustness of existing sEMG pattern recognition methods are also an important reason behind the high rejection rate. In daily use, the accuracy of sEMG pattern recognition methods is seriously degraded by several factors, such as electrode shift, changes in arm posture and muscle contraction force, poor electrode contact [[6], [7], [8], [9], [10]]. These factors could remarkably change the sEMG patterns. Consequently, most popular pattern recognition methods, which focus more on the high accuracy in the stationary feature space but less on the robustness to non-stationary environments, cannot achieve good performance under electrode shift.

In recent years, the issue of low robustness of sEMG-based prosthesis control has received increasing attention. Particularly, many methods have been proposed to solve the problem of adverse effects of electrode shift, which is the major source of low robustness of sEMG pattern recognition. For example, Variogram (Variog), a statistical measure of the spatial correlation, could provide robust features for high-density sEMG under electrode shift [11]. Gray-level co-occurrence matrix (GLCM), which provides a well description of spatial distribution of pixels in image processing, can discard information sensitive to shift and keep as much as useful information in electrode shift [12]. Common Spatial Pattern (CSP), which is based on the multiple channels signal analysis, can maximize the difference between the variance of two classes [13]. Electrode shift causes changes in the feature space, so that a classifier trained before electrode shift cannot achieve good performance when being used after electrode shift. Above-mentioned methods have showed improved robustness to electrode shift, but they were developed for high-density sEMG. High-density electrodes have high recognition efficiency and control quality [14], but with the increasing number of the electrodes (> 60 electrodes), the system will need more time to be worn and have a higher risk of breaking single electrode [15]. Furthermore, a high-density sEMG system has much higher requirements for signal acquisition, amplification, transmission and computation, which makes the system more complex and expensive. For these reasons, current high-density sEMG systems are mainly used in laboratory environment, but rarely in daily use.

To date, commercial EMG devices normally use low-density electrodes (for example, Myo armband, Myo, Thalmic Labs, with 8 electrodes) and they are more preferred in daily use for their lower price, easiness-to-wear and acceptable performance. However, it is still unclear whether these high-density electrodes-based methods, like Variog, GLCM and CSP, are still robust against electrode shift for low-density electrodes. Also, to the best of our knowledge, effective methods to cope with the problem of electrode shift in low-density electrode system are still lacking.

The present study is aimed to investigate the electrode shift effects on low-density sEMG signals and to develop a robust pattern recognition method to solve the problem of electrode shift. Firstly, we designed an experiment to produce electrode shift in a 8-channel sEMG configuration. Then we applied t-distributed stochastic neighbor embedding (t-SNE) [16], the quantitative measurement index space distance ratio (SDR), and the confusion matrix, to provide a comprehensive characterization of electrode shift effects on TD features. Last, the nonnegative matrix factorization (NMF) algorithm was used to reduce the non-stationarity of TD features and a self-enhancing linear discriminant analysis (SE-LDA) method was adopted to further improve the classification accuracy.

Section snippets

Experimental design

Twenty-five able-bodied subjects (18 males and 7 females, aged 22.04 ± 2.07) in this experiment. Ethical approval of the study was sought and obtained from the Medical Ethics Committee, Health Science Center, Shenzhen University (No. 2019004). All subjects were informed of the experimental procedure and signed informed consent before the experiment. Eight-channel sEMG signals were collected from able-bodied subjects' right forearms by SX230 EMG sensor (Biometrics Ltd., Newport, UK) with the

Characterization of TD and NMF Features (Single Subject Analysis)

We first checked the feature distributions without electrode shift (Dataset 1). The distributions of TD and NMF features from Dataset 1 are shown in Fig. 2A and B, respectively. It can be observed that the intra-class variance is increased and the inter-class variance is decreased after NMF decomposition, which is not good for classification. As introduced later in Section 3.3, TD features can achieve higher accuracy than NMF features, if there is no electrode shift.

We can further observe from

Changes in distributions of electrode shift

Electrode shift makes great changes in the feature space of low-density sEMG. But the characterization and visualization of TD features under electrode shift are not easy. Because there are 32 TD features from 8 channels, the high-dimensional feature space makes it difficult describe and visualize the electrode shift effects. Therefore, the robustness of a pattern recognition method against electrode shift can only be evaluated by the final classification accuracy, which limits the

Conclusions

In this paper, the problem of electrode shift for prosthetics control with low-density sEMG was studied. The concluding remarks are as follows. First, the non-stationary characteristics of sEMG impair the robustness of prosthetic control in daily use. Existing methods, which were originally proposed to deal with the electrode shift in high-density electrode environment, do not work well for low-density sEMG. Second, the changes of TD feature space caused by electrode shift were visualized by

Acknowledgments

This work was supported by National Natural Science Foundation of China (No. 61701316, No. 81871443), New Entry Teachers’ Scientific Research project in Shenzhen University (No. 2018012), Shenzhen Peacock Plan (No. KQTD2016053112051497) and Science, Technology and Innovation Commission of Shenzhen Municipality Technology Fund (No. JCYJ20170818093322718).

Declaration of Competing Interest

None of the authors have potential conflicts of interest to be disclosed.

References (24)

  • P. Parker et al.

    Myoelectric signal processing for control of powered limb prostheses

    J. Electromyogr. Kinesiol.

    (2006)
  • R.A. Jacobs

    Increased rates of convergence through learning rate adaptation

    Neural Netw.

    (1988)
  • X. Chen et al.

    Application of a self-enhancing classification method to electromyography pattern recognition for multifunctional prosthesis control

    J. Neuroeng. Rehabil.

    (2013)
  • B. Hudgins et al.

    A new strategy for multifunction myoelectric control

    IEEE Trans. Biomed. Eng.

    (1993)
  • G. Li et al.

    Quantifying pattern recognition—based myoelectric control of multifunctional transradial prostheses

    IEEE Trans. Neural Syst. Rehabil. Eng.

    (2010)
  • D. Farina et al.

    The extraction of neural information from the surface EMG for the control of upper-limb prostheses: emerging avenues and challenges

    IEEE Trans. Neural Syst. Rehabil. Eng.

    (2014)
  • M.A. Oskoei et al.

    Myoelectric control systems—a survey

    Biomed. Signal Process. Control

    (2007)
  • P. Kaufmann et al.

    Fluctuating EMG signals: investigating long-term effects of pattern matching algorithms

    Annu. Int. Conf. IEEE Eng. Med. Biol., IEEE

    (2010)
  • A.J. Young et al.

    The effects of electrode size and orientation on the sensitivity of myoelectric pattern recognition systems to electrode shift

    IEEE Trans. Biomed. Eng.

    (2011)
  • D.T. MacIsaac et al.

    Influences of dynamic factors on myoelectric parameters

    IEEE Eng. Med. Biol. Mag.

    (2001)
  • P.C. Doerschuk et al.

    Upper extremity limb function discrimination using EMG signal analysis

    IEEE Trans. Biomed. Eng.

    (1983)
  • A.J. Young et al.

    Improving myoelectric pattern recognition robustness to electrode shift by changing interelectrode distance and electrode configuration

    IEEE Trans. Biomed. Eng.

    (2012)
  • View full text