Brain-Pupil Coupling Revealed Through Deep Learning of Intracranial Recordings

Pupillary responses are windows into human cognition, but their neural substrates are poorly understood. We studied brain-pupil coupling through intracranial recordings and pupillometry in 13 children and youth with epilepsy (ages 9-18) during an attentional set-shifting task. Time-resolved mixed-effects modelling identified associations between pupil diameter, neural activity and cognitive performance. We first showed that pupillary dynamics are closely linked to cognitive performance, with task-stage dependencies. Larger pupil sizes prior to stimulus onset were associated with faster reaction times, whereas smaller pupil sizes during and after stimulus presentation were linked to better performance. Next, linear models identified associations between band-limited power in task-relevant neural networks and pupil size changes during the task. Finally, deep learning models based on intracranial neural activity captured patterns predictive of changes in pupil size in five of seven participants that generalised to recordings from a separate day. Using salience-based gradient mapping, we identified a network of task-relevant cortical and subcortical regions whose engagement was consistently associated with higher model performance in predicting pupil dynamics during attentional set-shifting. Our findings suggest pupillary responses are coordinated with goal-oriented cognitive processing, providing a basis for modelling cognitive functions through pupillary dynamics.