Mitigating confounding factors in myoelectric control through adaptive modelling and learning

Abstract

Despite significant advancements in myoelectric control for upper limb prostheses, which utilise electromyography (EMG) signals to control prosthetic movements, these devices continue to face performance limitations in real-world settings. The decoding of EMG signals is fundamentally challenged by confounding factors, which introduce unwanted variability that alters signal characteristics and degrade the decoding capabilities of the gesture recognition model, leading to performance deterioration. Consequently, myoelectric control systems for upper limb prosthetics struggle to maintain robustness and dexterity in practical applications, resulting in a gap between controlled laboratory settings and everyday use. This thesis addresses these critical limitations through adaptive learning and modelling approaches that can effectively mitigate these confounding factors, with the ultimate goal of developing more reliable prosthetic interfaces.

The thesis focuses on two critical confounding factors: long-term adaptation and the position effect.

First, we focus on long-term adaptation to address the challenge of gradual data drift over time. We propose an active learning framework with simulated human-in-the-loop (HITL) methodology to enhance the adaptability of a myoelectric decoder over time.

We implement sampling strategies, including least confidence and smallest margin techniques, demonstrating significant performance improvements.

With just 3.2 minutes of strategically selected training data, this framework improved accuracy by 4-5% for able-bodied participants and a notable 2% for individuals with limb differences compared to random sampling. Based on the significance of our findings, we propose a structured pipeline for future real-time deployment.

Second, the position effect refers to variations in limb positions that alter EMG signals by influencing muscle co-activation patterns through gravitational forces and electrode-muscle fiber interactions. To address the position effect, we first introduce GREAT, a comprehensive surface electromyographic dataset that addresses a critical gap in understanding signal variability across different arm positions. The protocol’s design extends beyond single-plane movements to encompass the full Cartesian space, thereby significantly enhancing the representativeness of the collected EMG data. The dataset comprises EMG and hand kinematics data collected from 8 able-bodied participants performing 6 distinct hand gestures (power, lateral, pointer, open, tripod, and rest) across 9 unique arm positions arranged in a 3×3 spatial grid. This data provides a robust resource for developing position-invariant myoelectric control algorithms. Our analysis demonstrates significant performance degradation (approximately 10%) during inter-position gesture decoding, underscoring arm position variability critical influence on sEMG signal characteristics.

Building on this foundation, we investigate position domain generalisation by implementing a rigorous benchmark design with nested cross-validation for fair and reliable model comparison. We conduct a comprehensive evaluation across statistical models, including Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA), and deep learning models, such as Empirical Risk Minimisation (ERM), Maximum Mean Discrepancy (MMD), and CORrelation ALignment (CORAL), under a domain generalisation framework. We introduce our DomainRegularised class-conditional Gaussian Mixture Model (DR-cGMM) that incorporates limb position knowledge via domain covariance regularisation to generalise across varying arm positions. Our analysis concludes that DR-cGMM achieves the highest overall average accuracy 94.29% across all domain generalisation tasks and statistically surpasses deep learning domain generalisation methods in the within-subject, low-data regime.

This thesis advances the field of myoelectric control by directly addressing two critical confounding factors through complementary approaches: active learning for long-term adaptation and domain generalisation through domain-regularised statistical modelling for position effect mitigation. Together, these contributions offer critical insights for enhancing the robustness of myoelectric interfaces. By tackling these fundamental challenges, this thesis lays the groundwork for more reliable, intuitive, and adaptable EMG-based myoelectric control systems that remain robust against the confounding factors encountered in daily prosthetic use.