Use this URL to cite or link to this record in EThOS:
Title: Time-varying brain connectivity with multiregression dynamic models
Author: Harbord, Ruth
ISNI:       0000 0004 7227 5169
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Functional magnetic resonance imaging (fMRI) is a non-invasive method for studying the human brain that is now widely used to study functional connectivity. Functional connectivity concerns how brain regions interact and how these interactions change over time, between subjects and in different experimental contexts and can provide deep insights into the underlying brain function. Multiregression Dynamic Models (MDMs) are dynamic Bayesian networks that describe contemporaneous, causal relationships between time series. They may therefore be applied to fMRI data to infer functional brain networks. This work focuses on the MDM Directed Graph Model (MDM-DGM) search algorithm for network discovery. The Log Predictive Likelihood (model evidence) factors by subject and by node, allowing a fast, parallelised model search. The estimated networks are directed and may contain the bidirectional edges and cycles that may be thought of as being representative of the true, reciprocal nature of brain connectivity. In Chapter 2, we use two datasets with 15 brain regions to demonstrate that the MDM-DGM can infer networks that are physiologically-interpretable. The estimated MDM-DGM networks are similar to networks estimated with the widely-used partial correlation method but advantageously also provide directional information. As the size of the model space prohibits an exhaustive search over networks with more than 20 nodes, in Chapter 3 we propose and evaluate stepwise model selection algorithms that reduce the number of models scored while optimising the networks. We show that computation time may be dramatically reduced for only a small trade-off in accuracy. In Chapter 4, we use non-local priors to derive new, closed-form expressions for the model evidence with a penalty on weaker, potentially spurious, edges. While the application of non-local priors poses a number of challenges, we argue that it has the potential to provide a flexible Bayesian framework to improve the robustness of the MDM-DGM networks.
Supervisor: Not available Sponsor: Engineering and Physical Sciences Research Council
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: QA Mathematics ; RC Internal medicine