An exploratory study to design a novel hand movement identification system

https://doi.org/10.1016/j.compbiomed.2009.02.001Get rights and content

Abstract

Electromyogram signal (EMG) is an electrical manifestation of contractions of muscles. Surface EMG (sEMG) signal collected from the surface of skin has been used in diverse applications. One of its usages is in pattern recognition of hand prosthesis movements. The ability of current prosthesis devices has been generally limited to simple opening and closing tasks, minimizing their efficacy compared to natural hand capabilities. In order to extend the abilities and accuracy of prosthesis arm movements and performance, a novel sEMG pattern recognizing system is proposed. To extract more pertinent information we extracted sEMGs for selected hand movements. These features constitute our main knowledge of the signal for different hand movements. In this study, we investigated time domain, time–frequency domain and combination of these as a compound representation of sEMG signal's features to access required signal information. In order to implement pattern recognition of sEMG signals for various hand movements, two intelligent classifiers, namely artificial neural network (ANN) and fuzzy inference system (FIS), were utilized. The results indicate that our approach of using compound features with principle component analysis (PCA) as dimensionality reduction technique, and FIS as the classifier, provides the best performance for sEMG pattern recognition system.

Introduction

Electromyogram signal (EMG) is the recording of the electrical potential generated by the muscles. The activities of the muscle can be monitored by means of surface electrodes placed on the skin. These signals contain important information concerning the total electrical activity associated with the muscle contraction. Surface EMG (sEMG) can be used in diverse applications. An important application is in the area of rehabilitation progression of amputees and in clinical settings where muscle disorder identification is required. Furthermore sEMG patterns can be utilized for hand prosthesis motion command synthesis. Normally these commands are used to perform hand movements caused directly due to the contraction of specific muscles. Generation of motion or force by muscle occurs when the fiber in the membrane is excited. An action potential then propagates along the surface membrane of the fiber triggering the chemical reactions that in turn cause fiber contraction. When a muscle contracts, the action potential generates an electric field that can be extracted by means of surface electrodes. Due to the different firing rates of muscles’ action potentials in different hand movements, each of these responses may be associated to a specific sEMG signal pattern. These patterns may be utilized for discriminating among different movements of hand. Hence the sEMG features extracted from different signal patterns are originated from different hand motion commands and can be utilized to design a system for identifying hand motions.

Such a system may have several parts, namely sEMG acquisition and preprocessing, features extraction, dimensionality reduction and classification. Proposed systems for distinguishing hand movements have an understandable limitation on the number of possible movements. The constraint is due to the inaccuracy that is indigenous to the design of these systems. The source of such inaccuracies is the non-stationary nature of the signal. To use sEMG in the pattern recognition application, one needs to extract the appropriate information of the signal such as the energy or its rate of change. These features specify time characteristics of the signals and in general provide good results but are not able to recognize beyond four movements [1], [2], [3], [4].

For designing multifunctional prosthesis hands one needs to extract more relevant information from sEMG signal, discriminating among specific sEMG patterns. Recording sEMG signals is intensely influenced by other adjacent muscles that may affect its quality, resulting in crosstalk or unwanted signals and transformed features of the desired signals. Therefore the need for greater number and more appropriate signal features along with a reliable and accurate recording system is obvious.

However, increasing the database and signal information through improvement of instrumentation techniques and increasing number of electrodes for sEMG signal recording may add to the complexity of subsequent analyses [2], [3]. A more plausible approach is to devise an analytical machine that is able to extract appropriate features from the signal.

Many investigators utilize time domain features, namely amplitude, zero crossing (ZC) and EMG autoregressive model (AR), while others employ spectral characteristics of EMG signal. [1], [5], [6], [7]. These features determine the time characteristics of muscle fiber action potential in a given recording time interval. Most work in the EMG signal classification assumes steady state characteristics, collected through and during a constant force and maintained contractions [2]. Hudgins et al. is the first to consider the information of the transient bursts of EMG signal that accompanies the onset of contractions [1]. He uses time domain features and ANN classifier for discriminating among four motion commands. Englehart et al. applies time–frequency domain features and shows that these representations, fusing time and frequency information simultaneously, are more appropriate for EMG patterns recognition [2], [3], [4]. For classification purposes, these studies utilize techniques based on linear discriminate functions, neural networks, and fuzzy systems.

In this work, a conventional scheme for sEMG pattern recognition system is utilized and at each analytical stage appropriate techniques are implemented. In order to increase the accuracy of the system and to obtain more signal information from each movement, three types of features, namely time, time–frequency and their combination, are used. Increasing the number of features increases the dimension of the feature set and adds to the system's complexity. Using the mentioned feature types results in the construction of the compound sEMG feature set that contains appropriate information. This matter was our major incentive for using compound features. However, to provide a reasonable comparison among the features and investigate the performance of them, we used identical number of features and dimensions in each stage. The usage of compound features has the advantage of retaining the characteristics of time structures of muscle fiber action potentials while attending to the TFRs features that contain further spectral and time information of sEMG signals for each movement.

In this study, we employ a two channel surface mounted electrode for a sEMG recording system. Utilizing a minimum number of electrodes may prove to be crucial as employing more electrodes may hamper desired signals due to possible crosstalk of adjacent muscles, and resulting diminished signal quality and accuracy. Our experimental protocol includes eight differentiating movements. We extract sEMG signals from each movement and train the pattern recognition system accordingly. In order to increase the performance and the reliability of the system we try to determine the best parameters of each feature extraction method by employing a minimum set of three hand movements. We then proceed to use the selected parameters to extract the features of the sEMG signals in the full set of eight hand movements.

In the following sections of this paper, we introduce the overall scheme of our proposed approach in designing the sEMG pattern recognition system. 2 Processing of sEMG signals, 3 sEMG pattern classification using intelligent techniques, 4 Using time domain features to design sEMG pattern recognition system, 5 Using time–frequency domain features to design sEMG pattern recognition system, 6 Using compound features to design sEMG pattern recognition system present theoretical background and analytical results based on extracted features from time domain, time–frequency domain and their combination, the compound representation, respectively. Section 7 provides an analysis of the proposed approach by increasing the number of movements from three to eight and re-evaluating the accuracy of the system, followed by the conclusion.

Section snippets

Problem definition

The general scheme of sEMG pattern recognition system utilized in this work is similar to one that has been introduced before [4], [8], [9]. We consider four major analytical steps, including sEMG pre-processing and conditioning, feature extraction, dimensionality reduction and classification.

The goal of the pre-processing step is to appropriately prepare and amplify the sEMG signal for the next analytical steps. Also in this module, artifact noise is eliminated. During the feature extraction

sEMG pattern classification using intelligent techniques

There are several techniques for classification but in this work we focused on two intelligent techniques that have generalization property, namely ANN and FIS. One of the simplest and widely used ANN for EMG pattern classification is multilayer perceptron (MLP) [22], [23], [24]. This network consists of a set of input layers, a number of hidden layers and a layer of output units. Each input unit is connected through a set of weights to hidden layer. Each of the hidden layers consists of a set

Using time domain features to design sEMG pattern recognition system

Time domain features extract time signatures of the sEMG signal. This signal has a semi-irregular shape in its temporal waveform due to a deterministic component but despite this, its inner structures are varied due to its random components. In utilizing raw and complete sEMG signal waveform, we may lose the important time domain features because the repeated movements record multiple time signatures of the signal, resulting in an inappropriate system performance for signal classification. To

Using time–frequency domain features to design sEMG pattern recognition system

Many signals in bioengineering are produced by biologic systems whose spectral characteristics may change rapidly with time. To analyze these rapid spectral changes one needs a two-dimensional, mixed time–frequency signal representation (TFR) with time represented along one axis and frequency along the other, indicating which frequency components are present at each instant of time.

TFRs have been used in different applications such as signal compression, coding, filtering and classification.

Using compound features to design sEMG pattern recognition system

After determining the best parameters for each of the features in the respected TFR method, to reach a further improvement, a combination scheme was imposed towards optimizing the pattern recognition system for the maximum number of hand movements. Our goal was to maximize the amount of information for the recorded signals for any given movement, thereby minimizing the confusing effects of the adjacent or secondary muscles in such recordings.

Two approaches were chosen for this purpose. These

Results

In order to attain the best approach for designing the sEMG pattern recognition system for hand movements, the accuracy of the proposed system was evaluated, using previously mentioned features of these signals. For this purpose, initially the accuracy of time domain and TFR and their compound feature sets were evaluated for the maximum number of hand movements. This task was performed distinctly for PCA and CS dimensionality reduction techniques and FIS and ANN classifiers. Fig. 6 depicts the

Conclusion

This investigation provided an exploratory study to propose a novel approach for designing a sEMG pattern recognition system. Our study employed three types of sEMG feature sets, namely time domain, time frequency domain and their combination. The analysis utilized PCA and CS as dimensionality reduction technique and fuzzy inference system and artificial neural network as classifiers. Eight hand movements were considered and sEMG pattern recognition system was designed for recognizing these

Conflict of interest statement

None declared.

Mahdi Khezri was born in Esfahan in 1980. He completed his bachelor's in Electrical Engineering in 2002, at Azad University of Najafabad and Master in Biomedical Engineering in 2006 at Sharif University of Technology. He is currently planning for Ph.D. degree in biomedical engineering. His research interests are in biomedical signal processing specially EMG signals in application of prosthesis, virtual reality and haptic system design, artificial intelligence in medicine.

References (25)

  • I.T. Jolliffe

    Principal Component Analysis, Series: Springer Series in Statistics

    (2002)
  • A.A. Miranda, Y.-A. Le Borgne, G. Bontempi. New Routes from Minimal Approximation Error to Principal Components, vol...
  • Cited by (32)

    • Detecting the universal adversarial perturbations on high-density sEMG signals

      2022, Computers in Biology and Medicine
      Citation Excerpt :

      Myoelectric control is a technique that uses sEMG signal to translate movement or intent into commands to manipulate a powered prosthesis or other external robotic device [3,4]. Myoelectric pattern recognition plays an essential role in myoelectric control, which allows dexterous movement of multiple degrees of freedom [5]. To take full advantage of both valuable spatial and temporal information about muscle activity, the 2-dimensional high-density electrode grids covering the entire muscle are utilized to collect high-density sEMG (HD-sEMG).

    • Multiuser gesture recognition using sEMG signals via canonical correlation analysis and optimal transport

      2021, Computers in Biology and Medicine
      Citation Excerpt :

      Myoelectric control is an interaction methodology that directly transforms the electrophysiological properties of muscles into commands that convey the intent of the user's movement [1]. It has been widely utilized in biomedical prosthetics, neuro-rehabilitation engineering and human-computer interaction [2–7]. For those applications, gesture recognition based on surface electromyography (sEMG) is popularly used since it is safe, non-invasive and easy to be implemented [8,9].

    • Electromyogram-controlled assistive devices

      2019, Bioelectronics and Medical Devices: From Materials to Devices - Fabrication, Applications and Reliability
    • Exploring relations between EMG and biomechanical data recorded during a golf swing

      2017, Expert Systems with Applications
      Citation Excerpt :

      Predictive models built using surface electromyography (EMG) data have been applied in a variety of fields. Prostheses control by mapping EMG signals into several classes of forearm motions (Bu, Okamoto, & Tsuji, 2005; Khezri & Jahed, 2009; Khushaba, Al-Ani, & Al-Jumaily, 2010; Khushaba, Kodagoda, Takruri, & Dissanayake, 2012), evaluation of local muscle fatigue in static (Al-Mulla, Sepulveda, & Colley, 2011a, 2011b) as well as dynamic exercises (Tenan, McMurray, Blackburn, McGrath, & Leppert, 2011; Venugopal, Navaneethakrishna, & Ramakrishnan, 2014), detection of driver hypo-vigilance (Sahayadhas, Sundaraj, Murugappan, & Palaniappan, 2015), hand gesture recognition (Mesa, Rubio, Tubia, De No, & Diaz, 2014; Phinyomark, Phukpattaranont, & Limsakul, 2012) and analysis of gait (Mitchell, Krotish, Shin, & Hirth, 2010; Nair, French, Laroche, & Thomas, 2010) are probably the main application fields. Anticipation of head motion for virtual-environment applications (Barniv, Aguilar, & Hasanbelliu, 2005; Shrirao, Reddy, & Kosuri, 2009), a computer interface based on EMG signals induced by wrist movements (Choi, Micera, Carpaneto, & Kim, 2009), prediction of forces on the lumbar spine (Hou, Zurada, Karwowski, Marras, & Davis, 2007), classification of different tremor types (Palmes, Tech Ang, Widjaja, Tan, & Au, 2010), motion pattern recognition (Tsai, Luh, & Lin, 2015) are other recent examples involving EMG data-based predictive modelling.

    • Predicting physiological parameters in fatiguing bicycling exercises using muscle activation timing

      2017, Biomedical Signal Processing and Control
      Citation Excerpt :

      Predictive models built using sEMG data have been used in many other fields. Prostheses control through classification of sEMG data into several classes of forearm motions [26,27], hand gesture recognition [28–30] and analysis of gait [31,17] are popular application fields. Anticipation of head movement for virtual-environment applications [32,33], a computer interface based on sEMG signals activated by wrist movements [34], prediction of forces affecting the lumbar spine [35], diagnostics of neuromuscular disorders [36], classification of different tremor types [37], motion pattern recognition [38], and prediction of knee abnormalities [39] are several other recent examples involving sEMG data-based predictive modelling.

    View all citing articles on Scopus

    Mahdi Khezri was born in Esfahan in 1980. He completed his bachelor's in Electrical Engineering in 2002, at Azad University of Najafabad and Master in Biomedical Engineering in 2006 at Sharif University of Technology. He is currently planning for Ph.D. degree in biomedical engineering. His research interests are in biomedical signal processing specially EMG signals in application of prosthesis, virtual reality and haptic system design, artificial intelligence in medicine.

    Mehran Jahed was born in Tehran in 1960. He completed his Bachelor's degree in EE in 1982 at Purdue University, West Lafayette, and Masters and Ph.D. in EE in 1987 and 1990 at University of Kentucky. He was a post doctoral fellow at Center for Excellence for Biomedical Engineering at University of Kentucky from 1990 to 1993. Since 1993 he has been a faculty member of EE department at Sharif University of Technology. His research interests are in biomedical modeling and control of the heart and neuromuscular systems, bio-robotics and tracking, virtual and augmented reality, and application of artificial intelligence in Medicine and biology.

    View full text