Abstract
This thesis explores statistical methodology in the analysis of data from patients with Parkinson’s disease (PD), with research conducted in collaboration with the “Host Microbiome Interaction: Clinical Pharmacology and Therapeutics” group at King’s College London. A key focus is the quantification of tremor, which is a cardinal sign of Parkinson’s disease; a better understanding of tremor can lead to an improved understanding of the disease and present opportunities for earlier diagnosis. Unlike in existing work which use accelerometers to measure tremor, the use of numerical integration to estimate displacement due to tremor is explored. Acknowledging that, whilst theoretically simple, numerical integration is challenging in practice, the work explores the existing literature on numerical integration and compares a collection of methodologies that each propose different ways of minimising the numerical error in the estimated displacement. Ground truth data is mechanically simulated using a ‘wobbulator’ and used to find the best performing method. The work then presents a novel method using techniques from functional data and time series analysis to detect noise within the recordings. These segments can be omitted before numerical integration, providing guaranteed improvements to the estimated displacement and other downstream metrics, for example, overall displacement due to tremor and intermittency of tremor, both by frequency. An analysis of these tremor metrics is presented alongside a discussion of the clinical relevance to disease understanding, including the relationship of these metrics with Calprotectin, a biomarker for intestinal inflammation.Analyses of bradyphrenia (cognitive efficiency) in PD is presented, using linear (mixed) regression models to model the effect of drugs on bradyphrenia. Cross-sectional and longitudinal analyses are performed. Given the observational nature of the data, various robustness checks are applied which demonstrate the validity of the results. The cognitive efficiency metric is compared to the results of a bradyphrenia questionnaire to ascertain whether a simple questionnaire can be used to determine severity of bradyphrenia.
Finally, a review of the literature into the use of audio biomarkers to detect or predict the severity of PD is conducted. The purpose of this is to understand the extent to which researchers assess the potential effects of confounding variables in the design of the data collection and model training. This work is motivated by a case study into the use of machine learning models to detect COVID-19 from the sounds of spoken language or coughs, which found that after adjusting for confounders that were not identified in similar studies, COVID-19 detection from audio offered no improvement to classification performance over simple symptom checkers. Various recommendations are made to researchers conducting similar experiments in future.
Postural instability is a key feature of the disease that can be challenging to measure objectively. This work assesses the feasibility of calculating angles of lateral and anterior lean throughout the gait cycle by applying a 3D joint estimation machine learning model to videos of participants. This proof of concept shows promise and the proposed analysis of these data using functional linear models is discussed.
| Date of Award | 1 Nov 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Steve Gilmour (Supervisor) & Davide Pigoli (Supervisor) |