An Efficient Framework for Personalizing EMG-Driven Musculoskeletal Models Based on Reinforcement Learning.

Area of Science:

• Biomedical Engineering
• Robotics
• Neuroscience

Background:

• Electromyography (EMG)-driven musculoskeletal models (MMs) are crucial for advanced prosthetic control.
• Personalizing these models is essential for accurate and intuitive prosthetic function.
• Current personalization methods can be time-consuming and computationally intensive.

Purpose of the Study:

• To develop a novel, rapid framework for personalizing upper limb musculoskeletal models using EMG signals.
• To enhance the accuracy of hand and wrist motion estimation for prosthetic applications.
• To compare the proposed framework's efficiency and effectiveness against existing methods.

Main Methods:

• Utilized a generic upper-limb musculoskeletal model as a baseline.
• Employed an artificial neural network-based policy trained with reinforcement learning (RL) for parameter fine-tuning.
• Compared the RL-based framework against a baseline model and simulated annealing (SA) for optimization.
• Conducted both offline evaluations with non-disabled subjects and online evaluations with human subjects, including an amputee.

Main Results:

• Personalized MMs significantly reduced motion estimation errors compared to the generic MM in both offline and online tests.
• The RL-based framework achieved model optimization in under one second, drastically outperforming SA (over 13 minutes).
• The personalized models demonstrated comparable kinematics estimation accuracy to slower methods.

Conclusions:

• The developed RL-based personalization framework offers a practical and highly efficient solution for EMG-driven musculoskeletal models.
• This approach can significantly benefit the daily use of upper limb prostheses and other assistive devices.
• Rapid personalization of MMs is key to improving prosthetic performance and user experience.