Ensemble Learning Method for the Continuous Decoding of Hand Joint Angles

Hai Wang, Qing Tao, Xiaodong Zhang2024 · DOI

TL;DR

This paper introduces a novel ensemble learning approach for accurately decoding continuous hand joint angles from surface electromyography (sEMG) signals. By combining multiple powerful machine learning models, they achieve state-of-the-art performance, paving the way for more intuitive and functional prosthetic hands.

What Problem Does This Paper Address?

Controlling prosthetic limbs, especially hands, with a high degree of dexterity and naturalness remains a significant challenge. Current sEMG-based control often relies on recognizing a limited set of pre-defined actions. However, for truly intuitive control, especially for tasks requiring fine motor skills, it's crucial to decode the continuous, proportional movements of individual hand joints. This allows for simultaneous and proportional control (SPC), mimicking the natural range of motion. The challenge lies in accurately translating the complex, noisy sEMG signals into precise joint angle estimations, especially considering the computational and power constraints of wearable prosthetic devices.

Methods

The core of this research is the application of ensemble learning to the sEMG-based hand joint angle decoding problem. Instead of relying on a single model, the authors explore combining multiple "base" models to achieve better performance and robustness.

Here's a breakdown of their approach:

  • Feature Extraction: They extract eight time-domain features from the sEMG signals. These features are chosen for their efficiency and low computational requirements, making them suitable for real-time applications. Examples include Mean Absolute Value (MAV), Zero-Crossing (ZC), and Root Mean Square (RMS).
  • Ensemble Models: They focus on three popular gradient boosting ensemble methods:
  • XGBoost: Known for its speed and performance, it uses a level-wise tree growth strategy.
  • LightGBM: Developed by Microsoft, it employs a novel tree-growing algorithm (GOSS) and histogram-based data representation for faster training and improved accuracy. It grows trees leaf-wise.
  • CatBoost: Developed by Yandex, it excels at handling categorical features and uses a symmetric binary tree structure.
  • Stacking Strategy: To further enhance performance, they implement a stacking ensemble technique. This involves using the predictions of the base models (XGBoost, LightGBM, CatBoost) as input features for a "meta-model" (in this case, a simple Artificial Neural Network - ANN). This allows the meta-model to learn how to best combine the strengths of the individual base models. They explore various combinations of base models within the stacking framework.
  • Control Group: For comparison, they also evaluate a Gaussian Process (GP) model, a common and well-established method for regression tasks.

Datasets

The study utilized data collected from ten healthy right-handed subjects (5 male, 5 female) with no prior experience in sEMG control. The data collection involved:

  • sEMG Signals: Recorded from eight forearm muscles using a wireless sEMG acquisition system.
  • Hand Joint Angles: Captured using a Vicon motion system.
  • Experimental Tasks: Subjects performed various functional tasks, including independent finger movements and real-world grasping actions (e.g., grasping a USB stick, holding a wine glass, reading books). This diverse set of tasks aims to capture a wide range of natural hand movements.

The data was pre-processed and split using a 10-fold cross-validation approach for robust evaluation.

Key Insights and Ideas

  • Ensemble Power: The research strongly supports the idea that combining multiple models (ensemble learning) significantly outperforms individual models for sEMG-based angle decoding.
  • Stacking's Advantage: The stacking approach, which learns to optimally combine the predictions of base models, proved to be more effective than using individual ensemble models or even a simple combination of three models.
  • CatBoost's Strength: Among the individual ensemble models, CatBoost demonstrated superior performance, suggesting its symmetric tree structure might be particularly well-suited for the nature of sEMG signals in this context.
  • LightGBM + CatBoost Synergy: The combination of LightGBM and CatBoost within the stacking framework yielded the highest accuracy, highlighting a powerful synergy between these two gradient boosting algorithms.
  • Efficiency for Prosthetics: The ensemble learning approach, particularly the chosen features and models, offers a good balance between high performance and computational efficiency, making it suitable for deployment on resource-constrained prosthetic devices.

Key Results

The study evaluated performance using Pearson Correlation Coefficient (CC) and Root Mean Squared Error (RMSE).

  • Best Performing Model: The stacking model combining LightGBM and CatBoost achieved the highest average CC of 0.897 ± 0.022 and the lowest RMSE of 7.09.
  • Comparison to GP: The proposed ensemble models significantly outperformed the Gaussian Process (GP) control group across all fingers.
  • Finger-Specific Performance: While generally high, performance varied slightly across fingers, with the middle finger showing slightly lower CC values, potentially due to muscle overlap.
  • Individual Subject Variability: The study also analyzed performance across subjects, showing that while most models performed well, there was some variability, with the best performance reaching a CC of 0.911 ± 0.87% and the lowest at 0.509 ± 0.87%.

Limitations and Future Work

  • Limited Subject Pool: The study involved only 10 subjects. The authors acknowledge the need to include more subjects with different dominant hands to assess the generalizability of the model.
  • Embodied System Deployment: The proposed model has not yet been deployed on an actual embodied prosthetic system. Future work will focus on testing its online performance in such a setting.
  • Feature Dimensionality: While the chosen features are efficient, the authors suggest exploring dimension reduction techniques to further decrease computational demand.
  • Amputee Validation: A crucial next step is to test the system's efficiency with actual amputees to verify its real-world utility.
  • ANOVA Analysis: While an ANOVA was performed to check for age-related discrimination, no significant effect was found. However, further investigation into other demographic factors might be beneficial.

Summary updated

BibTeX
@misc{wang2024ensemble,
  author = {Hai Wang and Qing Tao and Xiaodong Zhang},
  title = {{Ensemble Learning Method for the Continuous Decoding of Hand Joint Angles}},
  year = {2024},
  doi = {10.3390/s24020660},
}
Figure