FedAssist: Federated Learning in AI-Powered Prosthetics for Sustainable and Collaborative Learning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning for prosthetic control relies on surface electromyography (sEMG) data.
• Decentralized approaches are needed to address data ownership and privacy concerns.
• Non-independent and identically distributed (non-IID) data presents a significant challenge in collaborative machine learning.

Purpose of the Study:

• To propose a federated learning (FL) framework, FedAssist, for developing deep learning-based sEMG decoding methods for AI-controlled prosthetics.
• To address the challenge of non-IID sEMG datasets within a decentralized learning paradigm.
• To enhance data privacy and ownership while enabling collaborative model training.

Main Methods:

• Development of the FedAssist federated learning framework.
• Implementation of collaborative local-level and global-level warm-start strategies.
• Evaluation of the framework on non-IID surface electromyography datasets.

Main Results:

• FedAssist demonstrates superior performance in non-IID scenarios compared to conventional learning paradigms.
• The proposed warm-start strategies effectively mitigate the challenges posed by non-IID sEMG data.
• The framework successfully preserves data ownership in a decentralized setting.

Conclusions:

• Federated learning, specifically the FedAssist framework, offers a promising approach for developing robust AI-controlled prosthetics.
• The methods developed advance decentralized machine learning for sEMG signal processing.
• This research has potential applications in improving prosthetic precision and rehabilitation effectiveness.