Health Data Science CDT

Lead Research Organisation: University of Oxford
Department Name: Health Data Science CDT

Abstract

Recent advances in neuroimaging have transformed our understanding of the human brain, facilitating
research in many other disciplines such as medicine, neuroscience and psychology. While modern
studies collect data on 103-105 subjects, the dimensionality of the imaging and non-imaging data is
even greater. The challenges of the analysis of big data in neuroimaging encompasses large sample
size, high-dimensionality, complex dependence both across space and over multiple imaging
modalities. Variable screening is a prominent dimension-reduction approach to allow further, more
accurate, and more interpretable analysis. There has been a growing interest in selecting important
volume elements (voxels) in images to identify brain regions that are highly associated with certain
brain functions, behaviours or psychiatric disorders. However, most existing methods overlook useful
prior knowledge in specific applications. There is a lack of Bayesian variable screening methods to
systematically incorporate multi-type prior knowledge and structural information while selecting
important explanatory variables.To promote the computational tractability and interpretability of the methods and corresponding results,
we develop Bayesian scalar-on-image models for making predictions on clinical outcomes that
integrate variable screening. A motivating example is initiation of substance use as an outcome in the
Adolescent Brain Cognitive Development (ABCD) study, where we wish to identify associated features
in imaging biomarkers, and other covariates such as social, behavioral, clinical, and environmental
factors. Bayesian frameworks enable us not only to make use of the given data, but also to use prior
knowledge to guide the computation and account for uncertainty. In neuroimaging, despite the high
dimensionality, we have the benefit of pre-established atlases that partition the brain into distinct
regions based on functional and anatomy. These underlying structures can act as a priori information
and, thus, a Bayesian framework that can incorporate such prior knowledge on brain structure should
improve the interpretability and the accuracy of the screening method.This project will use data from the UK Biobank and the ABCD studies which can be accessed through
both of the project supervisors. Both studies collect brain images and other background information of
over ten-thousand subjects with follow-ups over a long period of time.Aim 1 will extend the use of the Posterior Mean Screening method (PMS) by thoroughly exploring the
potential of PMS covariance kernel construction. Expanding on my preliminary work with just imaging
data, I will develop computationally efficient approaches for the inclusion of large numbers of
non-imaging variables. Inclusion of non-imaging data at the same time as multiple imaging modalities
would make the PMS framework more generalisable.Aim 2 will focus on selecting imaging features for a Bayesian scalar-on-image neural network (SI-NN)
model for making predictions on clinical outcomes whilst promoting model interpretability. We will
develop fully connected NN's that are more flexible than a convolutional NN but are more stable due to
spatial smoothness priors and spatially-informed sparsity. This approach will account for complex
dependence among brain regions yet provide interpretable inferences on how the imaging features
influence the clinical outcome.Aim 3 will extend Aim 2 to account for multimodal imaging data by extending the SI-NN to a
scalar-on-multi-image neural network (SMI-NN) model. This model presents a new framework for
evaluating the interaction effects for multimodal imaging data within the same regions as well as across
different brain regions.
The proposed study will provide a better understanding of psychopathology by identifying important
imaging biomarkers.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S02428X/1 01/04/2019 30/09/2027
2432761 Studentship EP/S02428X/1 01/10/2020 30/09/2024 Kan Keeratimahat