Use this URL to cite or link to this record in EThOS:
Title: Personalised modelling of geographic movements in depression
Author: Palmius, Niclas
ISNI:       0000 0004 7971 5665
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:
It has been estimated that as many as 41% of individuals will experience depression at some point in their lives, a serious and debilitating mental illness. One way to improve the care that patients receive, and reduce the burden on the healthcare system, is to monitor the health of patients over time. One emerging area of longitudinal monitoring is using behavioural data, especially from smartphones, to assess the health of individuals in the community. This thesis explores the relationship between depression and the geographic movements of individuals. Geolocation data were analysed from a longitudinal study of over 130 participants, including patients with bipolar disorder and borderline personality disorder, one of the largest datasets from these patient groups in the community. After suitable preprocessing, several features hypothesised to be related to mental health were extracted from the raw geographic coordinates. A new method for clustering stationary places is described, which is shown to provide more consistent and reliable results than alternative methods. The features used were extended from previous work to provide better characterisation of individuals. To reduce noise in the extracted features and improve predictive performance, a Kalman filter-based technique is presented. This is demonstrated to provide improved performance over the unfiltered features, demonstrating the importance of considering noise in the feature space as well as in the data space. In patients diagnosed with bipolar disorder, statistical differences are shown between weeks where participants self-reported clinical-level depression and weeks where they did not. Classification of depression from the geolocation-derived features is demonstrated in these patients, achieving an accuracy of over 75%. Regression of the level of depression, however, proved challenging, and it is demonstrated that one of the reasons for this is that individuals naturally have different baseline values of features, but also fundamentally different behavioural traits under different levels of depression. A new technique is therefore presented that finds groups of individuals that have similar regression models linking their geolocation-derived features with their self-reported depression levels. This is applied to 59 participants who provided more than 5 weeks of data for analysis, and finds that there are indeed several groups of individuals that appeared to behave similarly. Regression models trained on these grouped individuals significantly improve on the regression results from models trained over all participants and models trained on limited calibration data. Overall, this thesis demonstrates the utility of geolocation as a marker of depression, but also highlights the importance of appropriate handling of data to maximise performance. Personalisation of models is also demonstrated to be critical for maximising performance.
Supervisor: Vos, Maarten De Sponsor: Wellcome Trust ; RCUK Digital Economy Programme
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available