Prediction modelling for neuroimaging analysis

Thesis

This thesis examines how neuroimaging models can be built, evaluated, and interpreted when predicting individual di erences in cognition and behaviour. Across resting-state functional MRI and structural MRI, it focuses on three linked challenges: extracting useful predictive information from brain data, determining what signal predictive models are capturing, and understanding how confounding and population heterogeneity a ect generalisability.

The first part of the thesis investigates whether variability across models of brain dynamics can be exploited rather than treated purely as nuisance. Using hidden Markov models of resting-state fMRI and a stacking framework, it shows that combining predictions across multiple model configurations can improve the robustness and accuracy of cognitive trait prediction. The second part examines what drives apparent dynamic functional connectivity and finds that much of the recoverable signal is consistent with mean activity and amplitude (i.e., variance), with more limited evidence for additional time-varying correlation. The third part considers the role of age in structural brain-cognition modelling and shows that brain-cognition relationships vary across age groups, with predictive models generalising asymmetrically across the lifespan. 

These findings highlight a recurring tension between model complexity, interpretability, and generalisability in predictive neuroimaging. The thesis therefore argues for a careful framework in which model performance is considered alongside what signal is being captured, how robustly findings transfer across populations and analyses, and what conclusions can reasonably be drawn from them.

Authors: Griffin, B

• DOI: 10.5287/ora-zby1ky0qm
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; functional magnetic resonance imaging (fmri); brain ageing; predictive modelling; hidden markov models; dynamic functional connectivity; brain–cognition relationships; neuroimaging
• Subjects: Machine learning; Neurosciences