Hierarchical Bayesian Inference in Psychosis

Schizophrenia is a severe mental illness that affects millions of people worldwide and can have a drastic impact on a patient’s life. The illness is characterised by symptoms such as hallucinations and delusions. In recent years, a powerful theoretical framework has been developed to understand better how such symptoms emerge, the predictive coding account of psychosis. In this thesis, I cast different symptoms of psychosis as instances of hierarchical Bayesian inference in a series of studies. The first study examined the question of how persecutory delusions emerge in early psychosis. We derived hypotheses based on previous literature and simulations and tested them empirically in a sample of 18 first-episode psychosis patients, 19 individuals at clinical high risk for psychosis (CHR) and 19 matched healthy controls (HC). Our results suggest that emerging psychosis may be accompanied by an altered perception of environmental volatility. In a second study, this modelling approach was applied to delusions more broadly in a large dataset including 261 patients with psychotic disorders and 56 HC to examine the relationship between delusions and reasoning biases that were previously reported in psychosis. The results of this study suggest that beliefs of patients with psychotic disorders were characterised by increased belief instability, which explained increased belief updating in light of disconfirmatory evidence. We also assessed the clinical utility of this approach by testing its ability to predict treatment response to a psychotherapeutic intervention and found that the parameters of the computational model were able to predict treatment outcome in individual patients. Lastly, in a final study, we modelled brain activity during an implicit sensory learning task in a third independent sample of 38 CHR, 18 early-illness schizophrenia patients, and 44 HC to assess the biological plausibility of this approach. Our results suggest that hierarchical precision-weighted prediction errors derived from the model modulate electroencephalography (EEG) amplitudes. Moreover, we found not only differences in the expression of precision-weighted prediction errors between schizophrenia patients and HC, but also between CHR, who later converted to a psychotic disorder, and non-converters. Jointly, this work demonstrates that this computational approach may not only be conceptually useful to understand the computational mechanisms underlying psychosis, but also clinically relevant and biologically plausible.