Human claustrum neurons encode uncertainty and prediction errors during aversive learning

Abstract

Flexible behavior depends on continuous updating of internal models, yet the neural circuits coordinating this process remain poorly understood [1]. The claustrum -- reciprocally connected to nearly the entire neocortex -- is uniquely positioned to influence cortical processing. Here we report single-neuron recordings from the human claustrum during aversive learning [2], with anterior cingulate cortex and amygdala recordings for comparison. Claustrum and anterior cingulate neurons displayed structured, task-related responses. Distinct subpopulations encoded stimulus onset and action-contingent outcomes, with outcome representations diverging between regions. Critically, both regions encoded model-derived latent variables -- uncertainty and prediction error -- but with different temporal profiles: only the anterior cingulate carried uncertainty signals during the intertrial period, while both regions encoded uncertainty and prediction error during the active-avoidance period. The amygdala, by contrast, showed minimal latent-variable modulation. These findings provide evidence that human claustrum neurons track higher-order cognitive variables not directly observable from sensory input, and reveal dissociable roles for the claustrum and anterior cingulate cortex in tracking latent task states.