Statistical Learning and Inference Is Impaired in the Nonclinical Continuum of Psychosis

Abstract

Our perceptions result from the brain's ability to make inferences, or predictive models, of sensory information. Recently, it has been proposed that psychotic traits may be linked to impaired predictive processes. Here, we examine the brain dynamics underlying statistical learning and inference in stable and volatile environments, in a population of healthy human individuals (N = 75; 36 males, 39 females) with a range of psychotic-like experiences. We measured prediction error responses to sound sequences with electroencephalography, gauged sensory inference explicitly by behaviorally recording sensory statistical learning errors, and used dynamic causal modeling to tap into the underlying neural circuitry. We discuss the findings that were robust to replication across the two experiments (Discovery dataset, N = 31; Validation dataset, N = 44). First, we found that during stable conditions, participants demonstrated greater precision in their predictive model, reflected in a larger prediction error response to unexpected sounds, and decreased statistical learning errors. Moreover, individuals with attenuated prediction errors in stable conditions were found to make greater incorrect predictions about sensory information. Critically, we show that greater errors in statistical learning and inference are related to increased psychotic-like experiences. These findings link neurophysiology to behavior during statistical learning and prediction formation, as well as providing further evidence for the idea of a continuum of psychosis in the healthy, nonclinical population.

Auditory mismatch responses are differentially sensitive to changes in muscarinic acetylcholine versus dopamine receptor function

“…In study 1, we found a significant difference between stable and volatile mismatch responses, such that mismatch was stronger in stable than in volatile periods. This mirrored previous reports of volatility effects on mismatch signals (Todd et al, 2014;Dzafic et al, 2020). However, in our study, this was mainly due to the altered mismatch response in the biperiden group, which was particularly affected during stable mismatch.…”

“…It should be noted that previous reports have presented volatility effects on mismatch processing in single-channel analyses (focusing only on Fz, (Todd et al, 2014;Dzafic et al, 2020), and that the whole-volume corrected effect presented in (Dzafic et al, 2020) did not replicate in a validation data set, suggesting that the effects of volatility on mismatch might be relatively subtle compared to the size of the mismatch effect itself.…”

Cross-modal auditory priors drive the perception of bistable visual stimuli with reliable differences between individuals [preprint]

“…The formation of prior expectations and prediction errors is also altered along the psychosis-spectrum, as indicated by impaired mismatch negativity, a brain response to violation of sensory regularities (Farkas et al, 2015;Fitzgerald & Todd, 2020;Koshiyama et al, 2020;Michie et al, 2002;Näätänen et al, 2016;Stefanics et al, 2014), that may reflect precision-weighted prediction errors (Stefanics et al, 2018). Relatedly, nonclinical individuals reporting psychotic-like experiences demonstrate deficient learning of reguralities and inference on a sensory task, and this deficit was associated with impaired mismatch negativity (Dzafic et al, 2020). We conducted exploratory analyses to investigate the association between task performance and schizotypal traits that may facilitate hypothesis generation.…”

Predicting subclinical psychotic-like experiences on a continuum using machine learning

“…Confidence was reported on an ordinal three-point scale (1 = 'not confident', 2 = 'moderately confident', 3 = 'very confident'). For more detail regarding the task and stimulus paradigm, refer to Dzafic et al (2020). and pre-processing 2.3.1…”

“…A similar approach, referred to as generative embedding has been used previously (Brodersen et al, 2011), however in that paper, model optimization was done using the pooled data (both training and testing). For more detail regarding DCM fitting and parameter estimation, refer to Dzafic et al (2020).…”