Modeling perception and behavior in individuals at clinical high risk for psychosis: Support for the predictive processing framework

Early intervention in psychotic spectrum disorders is critical for maximizing key clinical outcomes. While there is some evidence for the utility of intervention during the prodromal phase of the illness, efficacy of interventions is difficult to assess without appropriate risk stratification. This will require biomarkers that robustly help to identify risk level and are also relatively easy to obtain. Recent work highlights the utility of computer-based behavioral tasks for understanding the pathophysiology of psychotic symptoms. Computational modeling of performance on such tasks may be particularly useful because they explicitly and formally link performance and symptom expression. Several recent studies have successfully applied principles of Bayesian inference to understanding the computational underpinnings of hallucinations. Within this framework, hallucinations are seen as arising from an over-weighting of prior beliefs relative to sensory evidence. This view is supported by recently-published data from two tasks: the Conditioned Hallucinations (CH) task, which determines the degree to which participants use expectations in detecting a target tone; and a Sine-Vocoded Speech (SVS) task, in which participants can use prior exposure to speech samples to inform their understanding of degraded speech stimuli. We administered both of these tasks to two samples of participants at clinical high risk for psychosis (CHR; N = 19) and healthy controls (HC; N = 17). CHR participants reported both more conditioned hallucinations and more pre-training SVS detection. In addition, relationships were found between participants' performance on both tasks. On computational modeling of behavior on the CH task, CHR participants demonstrate significantly poorer recognition of task volatility as well as a trend toward higher weighting of priors. A relationship was found between this latter effect and performance on both tasks. Taken together, these results support the assertion that these two tasks may be driven by similar latent factors in perceptual inference, and highlight the potential utility of computationally-based tasks in identifying risk.