Predictive coding and hallucinations
Abstract: Psychotic symptoms represent a profound departure from consensual reality. They must involve the brain's mechanisms for perception and belief. Under predictive coding theory, the brain contains an internal model of the world, organized hierarchically, with top-down predictions and bottom-up prediction errors. Depending on their relative precision, prior beliefs dominate and errors are ignored; or prediction errors prevail, leading to new learning. In this framework, hallucinations arise from overly precise prior beliefs. Using a conditioned-hallucinations paradigm we found evidence in favor of this hypothesis. Learned expectations generate hallucinations in human volunteers and people who hear voices are significantly more susceptible to this effect. Using computational modeling we demonstrated that the effect was driven by string prior beliefs (weighting ones' past experiences over current sensory data). The effect engaged a circuit incorporating superior temporal sulcus (STS) and the insula. In patients with intractable voices, daily transcranial magnetic stimulation, bilaterally, over STS, significantly attenuates voices. The magnitude of this effect correlates with changes in functional connectivity between the STS and the insula. Thus, taking a computational psychiatry approach we can understand hallucinations at the levels of circuits and psychological mechanisms and perhaps better treat them.