Myoelectric digit action decoding with multi-output, multi-class classification: an offline analysis.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroprosthetics

Background:

• Current regression-based methods for prosthetic finger control using surface electromyogram (EMG) signals have limited success.
• Simultaneous and independent control of multiple degrees of freedom (DOFs) is the ultimate goal for myoelectric control.

Purpose of the Study:

• To propose and evaluate a novel paradigm, action decoding, for independent multi-digit movement intent prediction.
• To enable more precise and intuitive control of prosthetic hands.

Main Methods:

• Action decoding utilizes multi-output, multi-class classification to predict movement intent for each DOF (open, close, stall).
• Analysis of a public dataset from 10 able-bodied and 2 transradial amputee participants.
• Systematic offline analysis of algorithmic parameters (feature selection, classification algorithm, multi-output strategy).

Main Results:

• Demonstrated feasibility of action decoding for predicting movement intent for all five digits and thumb rotation.
• Identified key algorithmic parameters influencing decoding performance.
• Outcomes will inform real-time implementation for improved prosthetic control.

Conclusions:

• Action decoding offers a promising, paradigm-shifting approach for prosthetic hand control.
• The findings pave the way for enhanced real-time control experiments with upper-limb amputees.