Computational and neuroimaging approaches to major depressive disorder

Major depression is a severely debilitating psychiatric condition with high prevalence and substantial economic impact. However, its aetiology is largely unknown, mechanistic understanding remains limited, and treatment outcomes are hard to predict. Recently, a “computational psychiatry” approach has emerged which embraces the idea of using computational models to link brain function, behaviour and psychiatric illness. This thesis describes the use of computational psychiatry tools and techniques to advance understanding of abnormalities in decision-making and neuronal activity associated with depressive illness. Behaviour during novel reward learning tasks was analysed from patients diagnosed with major depressive disorder and healthy controls. Formal computational modelling was used to show behavioural impairments associated with depression during both learning and decision-making phases. Depressed participants displayed lower memory of rewards and decreased ability to use internal value estimations during decision-making. Functional MRI results showed decreased reward signals in areas including the striatum were associated with depression symptoms. Computational models were used to generate latent variable time-series of internal value estimations which were used for model-based fMRI analyses. Reward value encoding in hippocampus and rostral anterior cingulate was abnormal in depression and anterior mid-cingulate (aMCC) activity was altered during decision-making. A signal encoding the difference between the values of the two options was also found in the aMCC, linking the behavioural model to localised brain function. Depressed patients showed decreased event-related connectivity between aMCC and rostral cingulate regions, implying impaired communication between value estimation and decision-making regions. A large community-based sample of participants reporting a range of depressive symptoms performed a different probabilistic reward learning task. Mood symptoms were associated with blunted striatal reward signals. Event-related directed medial prefrontal cortex to ventral striatum effective connectivity was abnormally decreased related to the severity of depression symptoms. A generative-embedding machine learning approach was used to classify never-depressed healthy controls from participants with current or past major depression. A support vector machine classifier achieved 72% diagnostic accuracy using estimated connectivity parameters as features. The thesis replicates previous reports of abnormal depression-related neural activity in areas including the striatum, hippocampus and prefrontal cortex using novel reward learning tasks. Findings support the theory about abnormal neural reward valuation in major depression being a core pathophysiological process which could be a target for treatment. The thesis also provides important novel evidence for decreased connectivity between prefrontal and limbic brain regions, and within different prefrontal areas in depression. It shows how abnormalities in reward value based decision-making may be related to abnormal reward activation and connectivity in the brain, supporting glutamatergic and cortical-limbic related theories of depression