Dynamic belief updating in anxiety disorders during a predictive inference task with functional magnetic resonance imaging

Adaptive behaviour in uncertain and changing environments depends on the formation and updating of internal beliefs about latent environmental states, a process referred to as dynamic belief updating. Alterations in this process have been proposed to contribute to anxiety disorders (ADs), which are characterised by biased learning, elevated threat expectations, and reduced flexibility in updating beliefs. However, empirical evidence in clinically diagnosed adult AD populations remains limited, particularly from studies combining computational modelling with neuroimaging.

The aim of the present study was to investigate behavioural and neural signatures of dynamic belief updating in adult patients with anxiety disorders compared to healthy controls using a predictive inference task. Thirty adult patients with ADs and forty-five healthy controls thus far completed the Predator Task, a validated paradigm in which participants estimate and update predictions of predator attack angles in a volatile virtual environment. Behavioural performance was first assessed using standard group comparisons and computationally derived learning parameters obtained from a reduced Bayesian model.

Behavioural data were analysed using independent-samples t-tests to assess group differences in task performance and model-derived learning rates. No significant differences were observed between patients and controls in mean estimation error or learning parameters, including fixed learning rate (p = 0.165) and adaptive learning rate (p = 0.125), indicating comparable behavioural performance at the level of both aggregate performance and model-derived summary measures.

In addition to these analyses, a reduced Bayesian framework was used to derive trial-wise estimates of relative uncertainty and change-point probability, capturing latent aspects of environmental volatility and belief updating. These variables were incorporated into a computational regression model of behavioural updates, in which prediction errors were expressed as a function of uncertainty, change-point probability, and task-related modulators including noise condition, visibility, and outcome-related factors. The model further included parameters capturing motor noise and learning-rate-dependent variability to account for individual differences in behavioural stochasticity. To ensure robust estimation, regression parameters were fitted using multiple random initialisations per participant. This model-based regression analysis, together with functional magnetic resonance imaging analyses using the same computational regressors within a general linear modelling framework, is currently ongoing.

Preliminary findings indicate that, at the level of summary behavioural metrics, dynamic belief updating appears largely preserved in adult anxiety disorder patients within the current sample. Model-based analyses suggest that behavioural updates are well captured by prediction-error driven learning signals across participants. Ongoing analyses will determine whether more subtle group differences emerge when decomposing trial-wise updating into computational components and examining their associated neural correlates.

These initial results suggest that overt behavioural performance during predictive inference may not fully reflect potential alterations in latent belief-updating mechanisms in anxiety disorders. The planned computational regression and functional magnetic resonance imaging analyses will further clarify whether preserved behavioural performance masks differences in uncertainty-driven updating at the algorithmic and neural levels. Overall, this work demonstrates the feasibility of combining computational modelling and neuroimaging to study adaptive learning processes in clinical anxiety populations while emphasising the importance of distinguishing between behavioural outcomes and latent computational mechanisms.

References

No References