Optimization of myoelectric control via context-informed incremental learning

Abstract

Myoelectric control systems translate muscle activity into control signals for external devices, offering intuitive interfaces for assistive technologies and human–computer interaction. However, these systems are highly sensitive to distributional shift—the mismatch between calibration and deployment conditions—caused by changes in limb posture, electrode placement, and, inescapably, user behaviour. This thesis introduces and evaluates Context-Informed Incremental Learning (CIIL), a generalizable adaptation framework that leverages situational task context to guide real-time model updates without requiring explicit supervision.

CIIL is validated across classification and regression control schemes, statistical models, static neural networks, and temporal neural network models, and a range of closed-loop control tasks, including target acquisition and object manipulation in virtual reality. Through experiments involving over 60 participants, CIIL is shown to outperform conventionally calibrated models and unsupervised high-confidence adaptation both in favourable conditions and under severe distributional changes, such as electrode shifts. The framework enables stable adaptation by assigning pseudo-labels based not on model confidence but on the alignment between the user’s behaviour and the inferred task context in a more tolerant strategy than existing environment-dependent incremental learning approaches.

A systematic review of prior incremental learning approaches is provided, and CIIL’s compatibility with diverse model classes—including feedforward and temporal architectures—is empirically demonstrated. Extensions to velocity-based control and proportional regression further highlight CIIL’s flexibility. The thesis also positions CIIL within existing context-derived adaptation literature and discusses its unique suitability for non-unique control tasks.

Limitations are acknowledged, including the exclusive use of able-bodied participants. The work concludes by outlining future directions, including principled system analysis using control theory, evaluation of user psychological metrics such as agency, and clinical deployment in stroke or neuromuscular populations. Together, this thesis advances the field by demonstrating that situational context can serve not only as a behavioural cue but also as a powerful supervisory signal for stable, scalable, and user-aligned myoelectric adaptation.