Generalizing Upper Limb Force Modeling With Transfer Learning: A Multimodal Approach Using EMG and IMU for New Users...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Electromyography (EMG)-based force modeling is crucial for assistive, robotic, and rehabilitation devices.
• Current models primarily focus on intra-subject performance, creating a burden for end-user data acquisition.
• Generalizing models across individuals is key for widespread adoption but remains a significant challenge.

Purpose of the Study:

• To investigate the efficacy of transfer learning (TL) for generalizing EMG-based force modeling to new users.
• To reduce the data acquisition burden for end-users by adapting models with minimal new data.
• To evaluate TL performance against leave-one-subject-out (LOSO) and intra-subject modeling scenarios.

Main Methods:

• Developed a deep multimodal convolutional neural network (CNN) integrating high-density (HD) EMG and Inertial Measurement Unit (IMU) motion data.
• Employed a TL strategy: establishing a baseline model with existing user data, then fine-tuning with 10%, 20%, and 40% of new user data.
• Tested the model under isotonic, isokinetic, and dynamic conditions.

Main Results:

• TL significantly improved force modeling accuracy, increasing average R-squared values by 60.81% to 199.79% over LOSO.
• TL outperformed intra-subject modeling by 13.4% to 45.51% across different conditions.
• Demonstrated successful generalization to new experimental conditions for new users.

Conclusions:

• Transfer learning enables effective EMG-based force modeling adaptation to new users with substantially reduced data requirements.
• The proposed multimodal TL approach offers superior performance compared to conventional intra-subject and LOSO methods.
• TL facilitates the development of more adaptable and user-friendly EMG-controlled assistive and robotic systems.