Towards Robust Surface Electromyography for Upper-Limb Prostheses using Machine Learning

: (Alternative Format Thesis)

Abstract

The loss of part or all the upper limb has a major impact on the quality of life of an individual. With a rising number of amputees worldwide, from both acquired and congenital limb differences, reliable prostheses capable of replicating lost functionality and cosmetic aspects of the limb are in high demand. Myoelectric prostheses have been in development since the mid-1900s to fulfil this role. Despite significant research advances in the design of machine learning-based control schemes, suggesting the possibility of increased degree-of-freedom control and real-time operation, the translation of this control into the clinical domain has been limited.

The non-stationary nature of the electromyographic signal presents a principal challenge to integrating machine learning-based control into prostheses. In laboratory experiments where recordings are ideal, non-stationarity is often unaccounted for. The signal properties vary with changes in electrode and physiological parameters, such as muscle fatigue and skin-electrode impedance. These changes in the signal generation and recording can lead to a performance reduction in machine learning systems trained in a single recording environment. Several methods have been explored to adjust for these variations, such as the use of adaptive algorithms and multi-modal recordings. This thesis advances on these previous studies, exploring the themes of repeatable and robust electromyographic recording under varying environments.

In this work a novel low-channel count multi-environment electromyography dataset is described. An analysis of the inclusion of skin-electrode impedance and temperature parameters within the classification pipeline is performed. The dataset exceeds acceptable quality thresholds for electromyography data and is shown to provide a suitable baseline for benchmarking machine learning applications across challenging environmental variations. Inclusion of the additional parameters provides optimal classifier
selection within an ensemble method that improves inter-trial classification accuracy by 17.48%.

Separately, a method for repeatably placing an electromyographic sensor on the flexor carpi ulanris is presented. The proposed method is implementable independent of practitioner skill while providing reliable and acceptable recording. This benefits researchers from non-clinical backgrounds engaging in electromyography research. Additionally, a window selection paradigm informed by the signal-to-noise ratio is compared with existing machine learning-based methods, improving accuracy by ∼5% compared to a confidence-based method. The potential of context-informed retraining is demonstrated to compensate for signal variations between recording sessions.

Date of Award	20 May 2026
Original language	English
Awarding Institution	University of Bath
Supervisor	Benjamin Metcalfe (Supervisor) & Elena Seminati (Supervisor)