Isokinetic elbow joint torques estimation from surface EMG and joint kinematic data: using an artificial neural network model

https://doi.org/10.1016/S1050-6411(98)00030-3Get rights and content

Abstract

Because the relations between electromyographic signal (EMG) and anisometric joint torque remain unpredictable, the aim of this study was to determine the relations between the EMG activity and the isokinetic elbow joint torque via an artificial neural network (ANN) model. This 3-layer feed-forward network was constructed using an error back-propagation algorithm with an adaptive learning rate. The experimental validation was achieved by rectified, low-pass filtered EMG signals from the representative muscles, joint angle and joint angular velocity and measured torque. Learning with a limited set of examples allowed accurate prediction of isokinetic joint torque from novel EMG activities, joint position, joint angular velocity. Sensitivity analysis of the hidden node numbers during the learning and testing phases demonstrated that the choice of numbers of hidden node was not critical except at extreme values of those parameters. Model predictions were well correlated with the experimental data (the mean root-mean-square-difference and correlation coefficient γ in learning were 0.0290 and 0.998, respectively, and in three different speed testings were 0.1413 and 0.900, respectively). These results suggested that an ANN model can represent the relations between EMG and joint torque/moment in human isokinetic movements. The effect of different adjacent electrode sites was also evaluated and showed the location of electrodes was very important to produce errors in the ANN model.

Introduction

Over the years, the neurophysiology and biomechanics of muscle systems have been investigated quite extensively in order to characterize the relations between muscle activity (electromyography, EMG) and various dynamical and/or kinematic aspects of the ensuing movement behavior. Inman et al. [1] first observed changes in EMG signal amplitudes according to variations in the applied muscle load. Since then many reports have been published on EMG–force/torque relations 2, 3, 4. However, the influences of experimental condition, cross-talk, and electrode placement have conspired to ensure that no definitive relationship has yet been found [5].

Rectified, low-pass filtered EMG has been used in conjunction with more detailed musculoskeletal model to estimate muscle force and/or joint moment 3, 6, 7, 8, 9. This approach has been employed much less than the other methods because it is difficult to find a set of neuromuscular excitation signals (or muscle forces) that, when provided as inputs to EMG–force models, produce a coordinated movement (i.e. a simulation of the movement) [5]. Typically, these models have been based on the Hill muscle model, estimation of neural activation, and joint dynamics' model [Fig. 1(a)]. It was hoped that piecemeal examination of the basic dynamical parameters would result in progressively better quantitative models of the musculoskeletal system. A problem with this approach is that assumptions have to be made at each step about the largely unknown nonlinear properties of the musculoskeletal and nervous system.

An alternate approach is to construct a complex nonlinear model—such as artificial neural networks (ANN)—to `catch' all information in the neuro-musculo–skeletal system [Fig. 1(b)]. Early work using ANN focused mainly on cognitive process and pattern recognition 10, 11. The existing applications of ANN in biomechanics have dealt primarily with joint angles and moments estimation in gait simulation [12] and estimation of muscle recruitment in static condition [13]. Recently, Koike and Kawato [14] used a complicated ANN model to estimate isometric joint torques and trajectory from surface EMG in upper limb motions. These results showed that the ANN could represent the relations between muscle activities and kinetic data. However, there are many questions still unanswered. For example, can a simple network model be easily developed using the experimental data? How sensitive is the model to the choice of specific network parameters of the hidden node number? Can the model make prediction of the joint torque in a simple one joint motion? Can this network tolerate different electrode locations? In this study, we try to demonstrate our answers to the above questions in an isokinetic exercise experiment.

Section snippets

Artificial neural network

Neural computing has resulted from attempts to create a mathematical model of the information processing capabilities of the biological neural system. A three layers, fully connected, feed-forward network (Fig. 2) was chosen for representing the EMG–torque relationship in this study. Inputs to the network are the normalized EMG magnitudes of biceps brachii (EMGf) and lateral triceps brachii (EMGe), elbow joint angle (θ), and elbow joint angular velocity (ω). Thus, four input nodes were created

Results

During the learning periods, the ANN model variants learned to approximate more closely the muscle responses as the learning process was iterated. The learning process of a subject (Subject B in Table 2) was illustrated as an example (Fig. 4). In Fig. 4, an ANN with 10 hidden nodes was interrupted during learning to compare the model response with the actual joint torque. After 60 000 epoch learning, the mean RMSD and correlation coefficient (γ) were 0.0290 and 0.998, respectively. With

Discussions

Because of the research and clinical importance of predicting joint moments/torques and current limitation in determining joint torque from muscular activation, an alternative approach by ANN model was proposed. In this study we constructed the ANN model and cross-validated it by the experimental data. In addition, the influences of the number of the hidden nodes and different sites of electrodes were also discussed.

The creation of an ANN model requires the specification of certain parameter

Conclusion

Considering the results and the above discussion, we found that the ANN model can be used successfully to map a set of EMG signals onto an output value of joint torque in the isokinetic exercise. In the sensitivity analysis of the number of hidden nodes, no significant difference was found from 5 to 20 hidden nodes. The ANN model estimated of joint torques from the experimental EMG signals, and it allowed for direct comparison with the experimental results from a dynamometer. In the evaluation

Acknowledgements

The authors acknowledge the financial support of the National Science Council of ROC (Grants NSC 84-2331-B-002-177-M08). We are grateful for the help of Mei-Hwa Jan, Chairwoman of Department of Physical Therapy, in instrumentation.

Jer-Jun Luh was born in Tiapei, Taiwan, R.O.C., on May 1 1967. He received the B.S. degree in physical therapy and the M.S. degree in electrical engineering from the National Taiwan University, Taipei, Taiwan, R.O.C., in 1990 and 1992, respectively. He is currently working toward the Ph. D. degree in electrical engineering at National Taiwan University.

His academic interests are in the areas of dynamics of neuromuscular system, motion control, and the development of assistive devices.

References (30)

  • K. Hornik

    Multilayer feedforward networks are universal approximator

    Neural Net

    (1989)
  • Y. Hirose et al.

    Back-propagation algorithm which varies the number of hidden units

    Neural Net

    (1991)
  • D.A. Winter et al.

    Cross-talk in surface electromyography: theoretical and practical estimates

    J Electromyogr Kinesiol

    (1994)
  • B.M. van Bolhuis et al.

    The relative action of elbow-flexor muscles in isometric flexion and in flexion/extension movements

    J Biomech

    (1997)
  • B. Bigland et al.

    The relation between force, velocity and integrated electrical activity in human muscle

    J Physiol Lond

    (1954)
  • Cited by (99)

    • Effect of shoulder angle variation on sEMG-based elbow joint angle estimation

      2018, International Journal of Industrial Ergonomics
      Citation Excerpt :

      The overall control architecture of these applications can be generalized as: (1) preprocessing the sEMG signals to remove the noise or artifacts, (2) extracting various types of features, (3) feeding these features into a trained estimation model to identify an angle, and (4) conveying a control signal transformed from the output of the model to the device. In a traditional way, elbow angle can be estimated using sEMG signals from several arm muscles (Luh et al., 1999; Tang et al., 2014b; Raj and Sivanandan, 2015). But since shoulder angle information cannot be acquired from sEMG of these arm muscles directly, it is difficult to deal with the adverse effect of shoulder angle using a traditional sEMG-based estimation method.

    • Dynamic modeling of SEMG-force relation in the presence of muscle fatigue during isometric contractions

      2016, Biomedical Signal Processing and Control
      Citation Excerpt :

      The relation between muscles’ electrical activity and force is of special importance in many applications including gait analysis, orthopedics, rehabilitation, ergonomic design, haptic technology, tele-presence surgery and human-machine interactions [1–26].

    • Can shoulder joint reaction forces be estimated by neural networks?

      2016, Journal of Biomechanics
      Citation Excerpt :

      Isokinetic knee torque as predicted by a neural network using EMG, joint kinematics and several other variables showed higher accuracy than a forward stepwise regression model (Hahn, 2007). Luh et al. (1999) showed that moments around a single joint axis can be estimated by a neural network (NN), using segment kinematics and surface EMG as inputs. These results inspired us to investigate a NN approach in the direct prediction of the glenohumeral joint reaction force under unconstrained daily conditions, based on ambulatory obtainable variables like body segment kinematics and EMG.

    View all citing articles on Scopus

    1. Download : Download full-size image
    Jer-Jun Luh was born in Tiapei, Taiwan, R.O.C., on May 1 1967. He received the B.S. degree in physical therapy and the M.S. degree in electrical engineering from the National Taiwan University, Taipei, Taiwan, R.O.C., in 1990 and 1992, respectively. He is currently working toward the Ph. D. degree in electrical engineering at National Taiwan University.

    His academic interests are in the areas of dynamics of neuromuscular system, motion control, and the development of assistive devices.

    1. Download : Download full-size image
    Jin-Shin Lai was born in Taipei, Taiwan, R.O.C., in 1949. He received the M.D. degree from Medicine School of National Taiwan University, Taipei, Taiwan, in 1974. He completed residency training in the Department of Physical Medicine and Rehabilitation, National Taiwan University Hospital, Taiwan, in 1978.

    He was a Lecturer and Associate Professor of the Medical School, National Taiwan University in 1980 and 1984, respectively. Now, he is the Chairman of the Department of Physical Medicine and Rehabilitation, National Taiwan University Hospital, the Chairman and Full Professor of the Physical Medicine and Rehabilitation, Medical School, National Taiwan University, and the President of the Rehabilitation Medicine Association, Taiwan. His professional interests include sport medicine, rehabilitation medicine, and rehabilitation engineering.

    1. Download : Download full-size image
    Gwo-Ching Chang was born in Ping-Tung, Taiwan, in 1967. He received the B.S. degree in biomedical engineering the Chung Yung Christian University, Chung-Li, Taiwan, in 1990, the M.S. degree in biomedical engineering from the National Cheng Kung University, Tainan, Taiwan, 1992, and the Ph.D. degree in electrical engineering from the National Taiwan University, Taipei, Taiwan.

    He is currently working at Chunghwa Telecom Co., Ltd., Taiwan. His research interests include biomedical signal processing, telecommunication, and system security.

    1. Download : Download full-size image
    Cheng-Kung Cheng was born in Taipei, Taiwan, R.O.C., in 1959. He received the B.S. degree from National Cheng Kung University in 1981 and the M.S. and Ph.D. degree in orthopaedic biomechanics from the University of Iowa, Iowa City, in 1984 and 1988, respectively.

    After completing a Postdoctoral Fellowship with the Department of Physical Therapy at the University of Alberta, Edmonton, Canada, he joined the National Taiwan University, Taipei, Taiwan, R.O.C., where he was an Associate Research Professor in the Center for Biomedical Engineering of the College of Medicine from 1988 to 1995. From 1995 to 1997, he was an Associate Professor and became a Full Professor and Chairman in 1997 in the Institute of Biomedical Engineering, National Yang Ming University, Taiepi, Taiwan. His research interests include orthopaedic biomechanics, ergonomics, musculoskeletal research, and rehabilitation engineering.

    1. Download : Download full-size image
    Te-Son Kuo was born in Taiwan, Republic of China, on January 8, 1938. He received the B.S. degree in electrical engineering from the National Taiwan University, Taipei, Taiwan, in 1960, and the M.S. and the Ph.D. degrees in electrical engineering from Georgia Institute of Technology, Atlanta, Georgia, in 1967 and 1970, respectively.

    From 1963 to 1964, he had one year study on Digital Computers in the Philips International Institute of Technological Studies, Eindhoven, Netherlands. He was a Visiting Assistant Professor of Electrical Engineering Dept. at Texas A &M University, College Station, Texas, in 1970. From 1970 to 1973 he was an Associate Professor and became a Full Professor in 1973, then he served as the Department Head fom 1975 to 1981, all in the Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan, where he is now a Full Professor. He has been also holding a joint appointment with the center for biomedical Engineering, National Taiwan University since February 1993. His fields of interest are computer aided design, control system, and biomedical engineering.

    View full text