Role of Muscle Synergies in Real-Time Classification of Upper Limb Motions using Extreme Learning Machines

Chris Wilson Antuvan, Federica Bisio, Francesca Marini, Shih‐Cheng Yen, Erik Cambria, Lorenzo MasiaJournal of NeuroEngineering and Rehabilitation 2016 · DOI

Abstract: This work demonstrates better robustness of online decoding of upper-limb motions and motor intentions when using synergy feature. Furthermore, we have quantified the performance of the decoder trained using ELM for online control, providing a potential and viable option for real-time myoelectric control in assistive technology.

Introduction

Myoelectric control—the use of electromyography (EMG) signals to drive prosthetic limbs or exoskeletons—faces significant hurdles in real-world applications. While laboratory "offline" tests often show high accuracy, these systems frequently struggle in "online" scenarios. Real-time control is complicated by signal noise, muscle fatigue, and the user's constant adaptation to visual feedback. This paper investigates whether muscle synergies—coordinated patterns of muscle activation—can provide a more robust control signal than raw EMG data when paired with a high-speed Extreme Learning Machine (ELM) classifier.

Methodology

The study involved seven healthy subjects performing five distinct upper-limb motions and a rest state. The researchers compared two feature extraction strategies:

  1. Standard EMG Features: Time-domain features (Mean Absolute Value and Variance) from seven muscle channels.
  2. Synergy Features: Non-negative Matrix Factorization (NMF) was used to decompose the seven EMG channels into a lower-dimensional "synergy space," representing the brain's modular control of movement.

Classification was handled by an Extreme Learning Machine (ELM). Unlike traditional neural networks, ELMs use random input weights and solve for output weights linearly, allowing for the extremely fast training and recalibration necessary for real-time assistive devices.

Electrode placement on the seven targeted upper-limb muscles.

Results and Key Insights

The study revealed a notable "offline paradox." While raw EMG features achieved a near-perfect 99.37% accuracy on pre-recorded data, their performance dropped significantly during real-time interaction. In contrast, synergy features, which had lower offline scores (~65%), proved far more effective in the loop.

  • Online Accuracy: Synergy-based decoding reached 91.79%, outperforming the EMG-based approach at 84.09%.
  • Operational Efficiency: Synergy features enabled faster Motion Selection Times (0.11s vs 0.15s) and significantly shorter Motion Completion Times (2.09s vs 3.09s).
  • Robustness: Because synergies represent muscle coordination rather than raw intensity, they are less sensitive to the "noise" of human adaptation and effort variation during task execution.

Discussion and Conclusion

By reducing the dimensionality of the control signal (from 7 muscles to 4 synergies), the system effectively filters out non-essential signal fluctuations. This results in a more stable and intuitive interface for the user. While this study focused on healthy subjects and discrete motions, it establishes a viable pipeline for real-time myoelectric control.

Future work will involve testing this ELM-Synergy framework with stroke patients and amputees, as well as integrating the system into physical soft wearable exosuits for simultaneous, multi-joint control.

Summary updated

Abstract

This work demonstrates better robustness of online decoding of upper-limb motions and motor intentions when using synergy feature. Furthermore, we have quantified the performance of the decoder trained using ELM for online control, providing a potential and viable option for real-time myoelectric control in assistive technology.
BibTeX
@article{antuvan2016role,
  author = {Chris Wilson Antuvan and Federica Bisio and Francesca Marini and Shih‐Cheng Yen and Erik Cambria and Lorenzo Masia},
  title = {{Role of Muscle Synergies in Real-Time Classification of Upper Limb Motions using Extreme Learning Machines}},
  journal = {Journal of NeuroEngineering and Rehabilitation},
  year = {2016},
  doi = {10.1186/s12984-016-0183-0},
}
Figure