Enhancing myoelectric prosthetic control with deep learning : continuous DOF estimation and generalization
Abstract
Description
The advancement of machine learning (ML) offers significant potential for improving
assistive technologies aimed at restoring function for individuals with upper limb loss.
Although ML techniques such as direct control and pattern recognition have been
extensively explored over the past decades, current commercial myoelectric prostheses
remain limited in functionality. They typically support a small number of movements,
lack simultaneous control, and suffer from low reliability due to issues like electrode
displacement and physiological variability, which necessitate frequent recalibration.
Recent research has therefore shifted toward deep learning (DL) and regression-based
approaches capable of decoding continuous joint kinematics. However, challenges
persist in achieving robust interpretation of user intent across multiple degrees of
freedom (DOFs), particularly in scenarios involving unseen movements or cross-subject deployment. Few studies have systematically examined these conditions,
despite their importance for clinical translation.
This thesis addresses these gaps by providing the first large-scale evaluation of deep
regression models for sEMG-based multi-DOF decoding under both within-subject and
cross-subject conditions. Specifically, three contributions are made: (1) an extensive
analysis of CNN-LSTM architectures for simultaneous and proportional joint control
across 22 DOFs, tested on a broader and more diverse dataset than in prior studies; (2)
a systematic investigation of model robustness under unseen movements and session
variability, demonstrating improved generalization through multi-scale temporal
feature representations; and (3) a novel cross-subject framework that integrates model-agnostic meta-learning (MAML), ensemble learning, and multihead decoding, yielding
substantial improvements under both few-shot and zero-shot conditions.
Our findings show that deep learning models, when combined with principled
generalization strategies, can outperform traditional approaches in accuracy,
robustness, and scalability, moving myoelectric prosthesis control closer to practical
real-world use.