Gait intent recognition using mechanomyography for application in prosthetic lower limb control
File(s)
Author(s)
Needham, Ashwin Paul Henry
Type
Thesis
Abstract
It has always been supposed that a prosthesis must infer its function itself and in doing so replace the brain of the user. With the emergence of the bionic not only does one have the function the user requires, but potential for the control they crave. By fusing man and machine how does one accomplish the non-trivial task of interfacing the brain and conscious will of the subject with the bionic? It has long been recognised that the subconscious neural pathways that would have previously controlled the lower limb in walking are not severed when the limb is. Mapping the residual muscle activity to these walking modes is a question of science, and not imagination.
In this study I use microphonic mechanomyographic muscle activity sensors to achieve this long sought-after goal using pattern recognition as applied through an SVM machine learning algorithm. In doing so one must counter the problem of data volume, as seen through the eyes of an amputee, with sympathy to their stamina and comfort.
I contemplate the challenge through 3 distinct clinical scenarios. Firstly, one provides an off the shelf, preprogramed device with an algorithm trained on a cohort of able-bodied subjects to create a subject-independent data pool. Secondly, a clinician acclimatises the amputee to the algorithm through a limited training routine. Thirdly, we train the knee device using the amputee. In exploring these I transition through various levels of significance until an average of 90.1% accuracy in identifying gait modes is achieved, with 89.3% from the amputee differentiating groups of 2 of 5 walking modes from each other, including level walking at various cadences and ramp ascent and descent. In further work, a bespoke diagnostic socket is designed, manufactured and tested for my amputee subject, a significant advance on the status quo in the industry.
In this study I use microphonic mechanomyographic muscle activity sensors to achieve this long sought-after goal using pattern recognition as applied through an SVM machine learning algorithm. In doing so one must counter the problem of data volume, as seen through the eyes of an amputee, with sympathy to their stamina and comfort.
I contemplate the challenge through 3 distinct clinical scenarios. Firstly, one provides an off the shelf, preprogramed device with an algorithm trained on a cohort of able-bodied subjects to create a subject-independent data pool. Secondly, a clinician acclimatises the amputee to the algorithm through a limited training routine. Thirdly, we train the knee device using the amputee. In exploring these I transition through various levels of significance until an average of 90.1% accuracy in identifying gait modes is achieved, with 89.3% from the amputee differentiating groups of 2 of 5 walking modes from each other, including level walking at various cadences and ramp ascent and descent. In further work, a bespoke diagnostic socket is designed, manufactured and tested for my amputee subject, a significant advance on the status quo in the industry.
Version
Open Access
Date Issued
2022-10
Date Awarded
2023-03
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Vaidyanathan, Ravi
Publisher Department
Mechanical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Master of Philosophy (MPhil)