A New Labeling Approach for Proportional Electromyographic Control

Annette Hagengruber, Ulrike Leipscher, Bjoern M. Eskofier, Jörn Vogel2022 · DOI

TL;DR

This paper introduces a novel, sensor-free method for generating continuous labels from EMG signals, simplifying the training process for proportional electromyographic (EMG) control. By directly using EMG features, including the transient muscle contraction phase, this approach achieves higher accuracy and better fine motor control compared to traditional binary labeling methods, making it particularly promising for assistive devices.

What Problem Does This Paper Address?

Controlling complex devices like robots or advanced prosthetics often requires proportional and simultaneous control, where users can adjust output smoothly and control multiple functions at once. Electromyography (EMG), which measures muscle electrical activity, is a promising technology for this, but training EMG-based control systems is challenging. Traditional methods often rely on external sensors (like cameras or force sensors) to generate "ground truth" labels for training, which can be impractical or impossible for individuals with physical disabilities. Even when using visual cues, synchronizing EMG data with these cues can lead to delays and inaccurate labels. Furthermore, deciding whether to include the brief, initial phase of muscle contraction (the transient phase) or just the steady state during training can significantly impact control performance, but there's no clear consensus on the best approach. This paper tackles the challenge of creating accurate, continuous EMG control labels without external sensors and investigates the impact of including the transient phase.

Methods

The core of this work is a new labeling approach that directly extracts continuous labels from the pre-processed EMG feature stream recorded during a simple training procedure. Here's a breakdown:

  • EMG Feature Extraction: Raw EMG signals from eight sensors on the arm are processed to extract four time-domain features: waveform-length, slope sign-change, zero-crossing-rate, and sEMG-amplitude. These features are then normalized.
  • Label Generation: The normalized features from all sensors are summed and then normalized again to create a continuous label between 0 and 1 for each direction of control (e.g., left/right, up/down).
  • Transient Phase Inclusion: A key innovation is the direct inclusion of the transient phase of muscle contraction in the labeling process. This means the initial rise in EMG signal, not just the sustained contraction, is used to generate labels. This is achieved by using an activity threshold to identify the onset of muscle activation.
  • Gaussian Process Regression: The generated continuous labels are used to train a Gaussian Process (GP) regression model. This model learns to map the EMG features to continuous control commands (e.g., velocity in 2D).
  • Labeling Strategies Evaluated: The study compares five labeling strategies:
  • (A) Binary label, including transient phase.
  • (B) Continuous label, including transient phase (the proposed method).
  • (C) Binary label, using only steady-state phase.
  • (D) Continuous label, using only steady-state phase.
  • (E) Continuous label, including both rising and falling transient phases.

Datasets

The study utilized data collected from 10 participants during a supervised training procedure. This procedure involved participants performing isometric muscle contractions against a handle, guided by visual cues indicating four directions of movement. The EMG data collected during these training sessions served as the basis for generating the labels and training the control models. No pre-existing public datasets were used; the data was generated specifically for this study.

Key Insights and Ideas

  • Sensor-Free Labeling is Feasible and Effective: The proposed method demonstrates that continuous, accurate EMG control labels can be generated solely from EMG signals, eliminating the need for external sensors and simplifying the training process.
  • Transient Phase is Crucial for Accuracy: Including the transient phase of muscle contraction in the labeling process significantly improves control accuracy, particularly for fine motor movements. This suggests that the initial dynamic changes in muscle activity provide valuable information for precise control.
  • Continuous Labels Outperform Binary Labels: Continuous labels, especially when incorporating the transient phase, lead to better performance in terms of success rate and fine motor control compared to traditional binary labels.
  • Simplicity Enhances Usability: The straightforward training procedure, which doesn't require specific ramped contractions or reaching maximum voluntary contraction (MVC) levels, makes the system more accessible and less demanding for users, especially those with disabilities.

Key Results

  • Aiming Task (AT) Performance:
    • The proposed continuous labeling methods (B and E) achieved the highest success rates (98.5% and 99%, respectively), significantly outperforming the baseline binary method (A) with 88.5% success.
  • Fine motion time was significantly reduced with continuous labels (especially method B) compared to binary labels (A).
  • Path efficiency was higher with continuous labels (over 83%) compared to binary labels (under 79%).
  • Maximum cursor speed during gross motion was higher with binary labels, but this did not translate to better overall performance.
  • Tracking Task (TT) Performance:
    • No significant difference was found in the average distance to the target across labeling methods.
  • Method B (continuous, transient phase) resulted in the lowest average cursor speed, which was closest to the target's speed, suggesting more precise tracking.
  • User Preference: Participants rated method B (continuous, transient phase) as providing better control than method A (binary, transient phase).

Limitations and Future Work

  • Comparison Scope: The study focused on comparing its new approach against specific binary and continuous labeling variations. It did not compare against all possible labeling strategies or external sensor-based methods.
  • Application Domain: The study was conducted in a 2D environment on a screen. While the authors assume the results are transferable to 3D applications and robotic control, further validation is needed.
  • User Group: The study was performed with healthy subjects. While the method is designed for individuals with physical disabilities, direct validation with this population is a crucial next step.
  • Prosthesis Control: The effectiveness of this labeling approach for actual prosthesis control remains to be investigated.
  • Falling Transient Phase: While method E included both rising and falling transient phases, its performance was similar to method B (rising phase only). The authors hypothesize that the falling phase might not offer significant advantages, possibly due to difficulties in abrupt muscle relaxation for some users.
  • Future Work: The authors suggest deploying the approach with robotic systems in 3D, evaluating it with users suffering from severe muscular atrophy, and exploring its use in shared control systems. They also note that the continuous labels generated by their method could be beneficial for online adaptation of predictive models, addressing issues like sensor shift and drift.

Summary updated

BibTeX
@misc{hagengruber2022new,
  author = {Annette Hagengruber and Ulrike Leipscher and Bjoern M. Eskofier and Jörn Vogel},
  title = {{A New Labeling Approach for Proportional Electromyographic Control}},
  year = {2022},
  doi = {10.3390/s22041368},
}
Figure