Individual differences in policy precision: Links to suicidal risk and network dynamics

Abstract

The Behavioural modelling of decision-making processes has advanced our understanding of impairments associated with various psychiatric conditions. While many studies have focused on models that best fit behavioural data, the extent to which such models reflect biologically plausible mechanisms remains underexplored. To bridge this gap, we developed a probabilistic two-armed bandit task model grounded in the active inference framework and evaluated its performance against established reinforcement learning (RL) models. Our model not only matched but outperformed conventional RL models in explaining individual variability in choice behaviour. A central feature of our model is the optimisation of policy precision based on previous outcomes. This process captures the dynamic balance between model-based predictions derived from the internal generative model and the influence of immediate past observations. Importantly, incorporating the temporal dynamics of policy precision significantly improved the model's capacity to explain large-scale brain network activity and inter-subject variability. We found that increases in policy precision were positively associated with default mode network dominance and negatively associated with states dominated by dorsal attention and frontoparietal networks. These opposing associations suggest functional coordination between these systems, as supported by the correlations between brain state transitions and behavioural parameters. Furthermore, prolonged dominance of another brain state, characterised by elevated ventral attention network activity and stronger inter-network connectivity, appeared to disrupt this coordination. Finally, we found that heightened sensitivity to negative outcomes in a loss-related context was associated with high suicidal risk among individuals with major depressive disorder.