Development of a clinically-targeted human activity recognition system to aid the prosthetic rehabilitation of individuals with lower limb amputation in free living conditions

Jamieson, Alexander

Thesis

Development of a clinically-targeted human activity recognition system to aid the prosthetic rehabilitation of individuals with lower limb amputation in free living conditions

Download PDF

Creator

Jamieson, Alexander

Rights statement

Strathclyde Thesis Copyright

Awarding institution

University of Strathclyde

Date of award

2022

Thesis identifier

T16359

Person Identifier (Local)

201871324

Qualification Level

Doctoral (Postgraduate)

Qualification Name

Doctor of Philosophy (PhD)

Department, School or Faculty

Department of Biomedical Engineering

Abstract

Aim: Healthcare Professionals (HCPs) that specialize in the care of Individuals with Lower Limb Amputation (ILLAs) typically evaluate the patient’s physical wellbeing through physical function tests or subjective questionnaires filled out by the patient. These evaluations give a limited understanding of the ILLA’s physical wellbeing, which can be evaluated more in-depth via wearable sensor-based Human Activity Recognition (HAR) of physical activities. The key objectives of this thesis were to determine which physical activities could be of interest to HCPs, develop a portable sensor-based system to capture those physical activities, then evaluate how reliably those activities could be captured with wearable sensors. Methodology: A focus group was conducted with ILLAs and HCPs to identify the relevant outcome measurements for clinical assessment. A novel HAR study was conducted with ILLAs and non-amputated individuals wearing a thigh-bound accelerometer (ActivPAL™, PAL Technologies, Glasgow, UK) to evaluate how reliably these outcome measurements could be captured in free-living conditions. Results: The key activity monitoring outcomes identified were walking activities on a variety of terrains. Using supervised machine learning, a Support Vector Machine could capture walking activities on flat terrain, walking on hills and walking on stairs. There was further potential to distinguish the activities on walking terrains based on whether they were hard or soft. With unsupervised machine learning, it was possible to distinguish walking on flat or sloping terrain with walking up and down stairs without the need for annotated training data using a novel formula-based algorithm. The ActivPAL proprietary algorithm was also validated for detecting walking and stationary activity of ILLAs in free-living conditions. Conclusion: The thesis validated an activity monitoring system that could capture a variety of walking activities performed by ILLAs. These findings form the basis of a clinical activity monitoring framework which would allow HCPs to monitor the walking activity of their patients and gain a greater understanding of their rehabilitation progress.

Advisor / supervisor