Abstract:
Population-based modelling in biomechanics has been growing in popularity over the last decade with the onset of machine learning techniques and high performance computing. This thesis explores these techniques in the area of clinical musculoskeletal biomechanics, whereby large data sets can train machine learning methods to make rapid predictions and translate biomechanical analysis to the clinic, and allowing previously intractable methods to inform clinical decisions. The three population-based modelling in musculoskeletal biomechanics studies presented relate; (i) lower lumbar spinal shape to ¹⁸F-NaF uptake (a surrogate measure of bone metabolism); (ii) anthropometric measurements to human lower limb fat-free muscle volume using machine learning techniques; and (iii) a population-based model trained on the Victorian Institute for Forensic Medicine database integrated with wearable inertial measurement unit (IMU) sensors to follow the rehabilitation of patients with a total knee replacement. Study (i) revealed that features of spine shape, which we interpreted as sacral tilt and vertebral disc spacing, were the most influential shape factors that contributed to both magnitude and spatial variation of ¹⁸F-NaF uptake. Study (ii) revealed that shank girth, gender and age are significant factors that can be used to predict subject-specific fat-free muscle volume. Study (iii) found that activity as measured by IMUs was often in contrast to the patient self-reported oxford knee questionnaire, and that the peak tibial acceleration from the IMU was linearly correlated to the knee joint reaction force on a patient-specific level. This wearable technology can assist clinicians identify and monitor high risk patients. The outcomes of the three thesis studies demonstrate the power and applicability of population-based machine learning and how they can play a supporting role in the health sector.