Santos Ferreira, Fábio Daniel; (2021) Multi-view machine learning methods to uncover brain-behaviour associations. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PhDthesis_FabioFerreira.pdf - Accepted Version Download (41MB) | Preview

Abstract

The heterogeneity of neurological and mental disorders has been a key confound in disease understanding and treatment outcome prediction, as the study of patient populations typically includes multiple subgroups that do not align with the diagnostic categories. The aim of this thesis is to investigate and extend classical multivariate methods, such as Canonical Correlation Analysis (CCA), and latent variable models, e.g., Group Factor Analysis (GFA), to uncover associations between brain and behaviour that may characterize patient populations and subgroups of patients. In the first contribution of this thesis, we applied CCA to investigate brain-behaviour associations in a sample of healthy and depressed adolescents and young adults. We found two positive-negative brain-behaviour modes of covariation, capturing externalisation/ internalisation symptoms and well-being/distress. In the second contribution of the thesis, I applied sparse CCA to the same dataset to present a regularised approach to investigate brain-behaviour associations in high dimensional datasets. Here, I compared two approaches to optimise the regularisation parameters of sparse CCA and showed that the choice of the optimisation strategy might have an impact on the results. In the third contribution, I extended the GFA model to mitigate some limitations of CCA, such as handling missing data. I applied the extended GFA model to investigate links between high dimensional brain imaging and non-imaging data from the Human Connectome Project, and predict non-imaging measures from brain functional connectivity. The results were consistent between complete and incomplete data, and replicated previously reported findings. In the final contribution of this thesis, I proposed two extensions of GFA to uncover brain behaviour associations that characterize subgroups of subjects in an unsupervised and supervised way, as well as explore within-group variability at the individual level. These extensions were demonstrated using a dataset of patients with genetic frontotemporal dementia. In summary, this thesis presents multi-view methods that can be used to deepen our understanding about the latent dimensions of disease in mental/neurological disorders and potentially enable patient stratification.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item