Myographic data-driven insights for the integration of inter-muscular dynamics in gait prediction models
File(s)
Author(s)
Guez, Annika
Type
Thesis
Abstract
Human walking, also referred to as gait, is one of the most basic and intuitive movements for able-bodied adults. Losing the ability to walk, due to physiological or neurological injury, can therefore severely impact the quality of life of an individual. Assistive robotics have the potential to support patients during rehabilitation. For effective gait training, they need to maintain user engagement and enhance their sense of agency by accurately predicting motion intent, and providing natural gait kinematic outputs. To this effect, electromyography (EMG) , a sensing modality that tracks muscle activity, can be used as input to these gait prediction models. However, due to its sensitivity to electrode placement, lack of inter-session reproducibility, and noise levels, its deployment in clinical applications is limited.
The work presented in this thesis investigates multiple aspects of EMG-based gait prediction, to improve the integration of EMG in robotic aided rehabilitation. This includes: (1) sensor placement and spatial robustness for enhanced model performance, (2) a data-driven selection of muscle groups most suited as inputs to lower-limb kinematic predictions, and (3) the investigation of muscle inter-connectivity and functionality for more meaningful gait models. Results revealed that the inclusion of the trunk, as well as the targeted selection of muscles that drive movement, led to higher prediction accuracies across all investigated subjects. Furthermore, our analyses underline the importance of including inter-muscular and inter-joint dynamics for the selection of both inputs and outputs in gait prediction models.
This thesis bridges the gap between an engineering perspective and a more holistic understanding of human inter-muscular dynamics, combining myographic machine learning models with insights from clinical literature. With a targeted use of muscle activity inputs and physiology-inspired outputs, this thesis contributes towards a safer and more informed integration of EMG in gait models, for future applications in lower-limb assistive devices.
The work presented in this thesis investigates multiple aspects of EMG-based gait prediction, to improve the integration of EMG in robotic aided rehabilitation. This includes: (1) sensor placement and spatial robustness for enhanced model performance, (2) a data-driven selection of muscle groups most suited as inputs to lower-limb kinematic predictions, and (3) the investigation of muscle inter-connectivity and functionality for more meaningful gait models. Results revealed that the inclusion of the trunk, as well as the targeted selection of muscles that drive movement, led to higher prediction accuracies across all investigated subjects. Furthermore, our analyses underline the importance of including inter-muscular and inter-joint dynamics for the selection of both inputs and outputs in gait prediction models.
This thesis bridges the gap between an engineering perspective and a more holistic understanding of human inter-muscular dynamics, combining myographic machine learning models with insights from clinical literature. With a targeted use of muscle activity inputs and physiology-inspired outputs, this thesis contributes towards a safer and more informed integration of EMG in gait models, for future applications in lower-limb assistive devices.
Version
Open Access
Date Issued
2024-12-29
Date Awarded
01/10/2025
License URL
Advisor
Vaidyanathan, Ravi
Sponsor
UK Research and Innovation
UK Dementia Research Institute
Grant Number
EP/S023283/1
UK DRI-7003
UK DRI-7005
Publisher Department
Department of Computing
Department of Mechanical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)