Mitigating the effect of electrode displacement for robust neural interfaces with wearable muscle sensors
Author(s)
Binenbojm de Sa Pereira, Joao
Type
Thesis or dissertation
Abstract
Brain-computer interfaces offer a direct communication pathway between the nervous system and external devices, promising transformative applications in neurorehabilitation and assistive technology. Among non-invasive modalities, surface electromyography has emerged as a prominent solution, leveraging the muscle as a biological amplifier to provide a high-fidelity, accessible readout of motor intent. However, widespread adoption is hindered by surface electromyography variability across conditions, with particularly severe performance degradation observed in subject-specific models due to electrode displacement across recording sessions. Existing solutions typically rely on prohibitively large-scale data collection or strenuous calibration procedures.
This thesis aimed to develop machine learning approaches that ensure robust intersession generalisation without resource-intensive data collection. Initially, spatial data augmentation using high-density surface electromyography channel subsets was evaluated using a publicly available gesture classification dataset. The approach yielded poor final performances, achieving less than 45% accuracy for classification between eight sustained hand gestures. Consequently, the Spatial Adaptation Layer was introduced: an interpretable, input-based domain adaptation approach designed to virtually correct electrode displacement.
Validated on two independent public datasets with surface high-density electromyography recordings during sustained hand gestures, the introduced layer achieved significant performance improvements over state-of-the-art methods. Specifically, it provided a >30% accuracy improvement when using a sparse calibration set of only one sustained gesture. To demonstrate its versatility, this framework was extended to the domain of motor unit decomposition. Validated on both phenomenological simulations and experimental data, the proposed adaptation method significantly reduced simulated electrode displacement (median post-calibration displacement of <0.25mm) and enabled the tracking of the majority of originally decomposed motor units. Collectively, these contributions establish a versatile, computationally efficient framework for handling electrode displacement, facilitating reliable neural interfaces for longitudinal use.
This thesis aimed to develop machine learning approaches that ensure robust intersession generalisation without resource-intensive data collection. Initially, spatial data augmentation using high-density surface electromyography channel subsets was evaluated using a publicly available gesture classification dataset. The approach yielded poor final performances, achieving less than 45% accuracy for classification between eight sustained hand gestures. Consequently, the Spatial Adaptation Layer was introduced: an interpretable, input-based domain adaptation approach designed to virtually correct electrode displacement.
Validated on two independent public datasets with surface high-density electromyography recordings during sustained hand gestures, the introduced layer achieved significant performance improvements over state-of-the-art methods. Specifically, it provided a >30% accuracy improvement when using a sparse calibration set of only one sustained gesture. To demonstrate its versatility, this framework was extended to the domain of motor unit decomposition. Validated on both phenomenological simulations and experimental data, the proposed adaptation method significantly reduced simulated electrode displacement (median post-calibration displacement of <0.25mm) and enabled the tracking of the majority of originally decomposed motor units. Collectively, these contributions establish a versatile, computationally efficient framework for handling electrode displacement, facilitating reliable neural interfaces for longitudinal use.
Version
Open Access
Date Issued
2026-01-06
Date Awarded
2026-08-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Farina, Dario
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)