Use this URL to cite or link to this record in EThOS:
Title: Objective assessment of Parkinson's disease using machine learning
Author: Prince, John
Awarding Body: University of Oxford
Current Institution: University of Oxford
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Neurodegenerative disorders, such as Parkinson's disease (PD), are infamously heterogeneous in regards to onset age, symptom prevalence, and severity progression rate. The current 'gold standard' techniques currently in use to assess the highly debilitating motor and non-motor symptoms are subjective, infrequent, and often ineffective with up to 20% of cases going undiagnosed until post-mortem. As such, these traditional and well established clinical assessment techniques are now starting to be fused with data-driven approaches in a bid to improve diagnosis and severity monitoring through the identification of objective disease biomarkers. Digital sensors present the opportunity to extract quantitative measures that are representative of disease presence and severity. However, existing studies using digital sensors to perform disease quantification are restricted by small cohorts, inconsistent experimental protocols, and purely cross-sectional analyses. The objective of this thesis is to provide novel insights as to how digital sensors can be further leveraged alongside prior clinical knowledge in order to improve the way Parkinson's disease is assessed in both clinical and remote environments. This thesis focuses on two large datasets that are juxtaposed in data quality, collection environment, and data type. Firstly, a clinical based approach to disease quantification is performed wherein an extensive network of wearable sensors is introduced into routine clinical care. Using this clinical dataset, the ability of wearable sensors to detect digital biomarkers distinctive of PD is investigated including the use of these biomarkers to perform automatic disease classification and severity prediction. Due to the longitudinal nature of data collection, new insights are revealed pertaining to symptom progression and the benefit of including longitudinal symptom variation into classification tasks. Secondly, the efficacy of performing disease assessment entirely in a remote environment using smartphones is investigated. Data was collected relating to many areas of disease manifestation and was often contributed daily by participants on a longitudinal basis. However, the remotely collected data suffers from a large degree of missingness, poor participant retention rate, and variable environmental conditions presenting new challenges during its analysis. High-frequency longitudinal analyses are performed and identify previously unseen motor and non-motor symptom progression characteristics. In order to perform disease classification using this dataset, a novel methodology is presented that compensates for the large quantities of source-wise missing data which, when combined with a stateof- the-art convolutional neural network, subsequently improves classification accuracy from 73.1% to 82.0%. Finally, a consistent analysis protocol is implemented on both datasets whilst simulating source-wise missing data; enabling a comprehensive comparison of missing data strategies for the purpose of disease classification. This thesis presents findings that highly support the hypothesis that digital sensors can successfully perform objective disease assessment at both the individual and population level.
Supervisor: De Vos, Maarten ; Clifton, David Sponsor: Engineering and Physical Sciences Research Council ; RCUK Digital Economy Programme
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: Neuroscience ; Multi-Modality Data Fusion ; Deep Learning ; Machine Learning