Modelling of Behaviour and Neuroimaging in Decision Making and Memory Through Application of Variational Bayesian Methods, In Health and Disease

Abstract

Computational Psychiatry is a rapidly emerging field, which combines traditional neuroscience with formal computational methods to investigate the transfer of information in neural circuits.
In this thesis, I aimed to combine analyses of behavioural and neuroimaging data with computational models of cognition and biological brain circuits, by applying Bayesian computational modelling techniques: specifically Active Inference and Dynamic Causal Modelling, in both health and disease.
I optimised a behavioural study with the aim of inverting a probabilistic Markov decision task for individual phenotyping. I observed a range of behavioural profiles across cohorts of healthy volunteers, and revealed distinct exploratory and exploitative behaviours.
I then applied human behavioural data to an Active Inference modelling framework, in which I inverted generative models to estimate subject-specific parameters encoding key mechanisms underlying behaviour and reward. I found that model inversion was successful in the accurate retrieval of these parameters within-subject, and that these parameters could predict coarse behavioural metrics on the group level.
By combining these techniques, I conducted a functional MRI study in which healthy participants performed the optimised behavioural task, then underwent a drug manipulation to induce selective noradrenaline reuptake inhibition. By constructing generative models for these participants, I found significant associations between neural activity in the locus coeruleus and anterior cingulate cortex, and model parameters estimated through inversion.
Finally, to examine biological circuits in neurogenerative disease, I analysed electroencephalography data collected from patients with Alzheimer’s disease and healthy older controls. This revealed left-lateralized memory circuit dropout in deeper memory tasks, with potential right-hemisphere compensation in simpler visual memory recall.
Taken together, these studies demonstrate the application of computational modelling in the study of problems in psychiatry and neuroscience to link mechanism to behaviour. The studies provide evidence that emerging Bayesian frameworks in computational psychiatry provide robust and mechanistically interpretable phenotypes.

Date of Award	2 Dec 2021
Original language	English
Awarding Institution	University of Bristol
Supervisor	Rosalyn Moran (Supervisor) & Iain D Gilchrist (Supervisor)