Dynamic interactions in human amygdala subregions support hierarchical processing of fearful faces

Fearful-expression recognition is vital for survival, relying on rapid fear detection and fine-grained face encoding. However, how human amygdala subregions differentially implement these distinct cognitive components remains unclear. Here, we recorded intracranial EEG from lateral and medial amygdala in epilepsy patients performing an emotional face-matching task, and combined multivariate decoding, time-frequency and directed-connectivity analyses with intracranial stimulation. We found that the lateral subregion exhibited early, fear-specific theta/alpha-band (4-12 Hz) activity with higher decoding accuracy and directed transfer of fear-specific information to the medial subregion, which showed delayed and sustained activation. By contrast, the medial subregion encoded face-specific information at a later stage in the 2-16 Hz band with superior decoding accuracy, and then relayed this information back to the lateral subregion. Critically, intracranial stimulation yielded double-dissociable behavioural effects, with lateral stimulation disrupting early fear detection while medial stimulation accelerated neutral-face recognition. Together, our study delineates a hierarchically organized, temporally structured division of labour within the human amygdala and show that dynamic interactions underlie hierarchical processing of fearful expressions.