FedEMG: Achieving Generalization, Personalization, and Resource Efficiency in EMG-Based Upper-Limb Rehabilitation...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Rehabilitation Technology

Background:

• Upper extremity amputation significantly impacts quality of life.
• Deep learning for electromyography (EMG)-based prosthetic control faces challenges in user generalization, personalization, and computational efficiency.
• Existing Federated Learning (FL) methods struggle to balance global knowledge with user-specific needs.

Purpose of the Study:

• To introduce Federated Electromyography (FedEMG), a novel Federated Prototype Learning (FPL) framework for real-time EMG-based gesture recognition in prosthetic control.
• To address the trade-off between personalization and generalization in prosthetic control systems.
• To enable efficient, real-time prosthetic control on resource-constrained devices.

Main Methods:

• Developed FedEMG, a Federated Prototype Learning (FPL) framework utilizing a prototype-based approach for knowledge transfer.
• Implemented an adaptive personalization mechanism to balance global and user-specific knowledge.
• Integrated a lightweight gesture detector and an efficient neural network architecture for real-time performance.

Main Results:

• FedEMG demonstrated superior accuracy in intra- and inter-subject gesture recognition across various non-IID datasets.
• The framework achieved high personalization and generalization without performance degradation.
• Evaluations confirmed efficient resource utilization, enabling real-time application.

Conclusions:

• FedEMG offers an effective solution for real-time EMG-based gesture recognition in upper-limb prosthetics.
• The proposed framework advances upper limb rehabilitation by providing improved and accessible prosthetic control.
• FedEMG successfully balances personalization, generalization, and computational efficiency for prosthetic applications.