Mesoscopic cortical activities associated with pupil-linked perceptions inferred via explainable machine learning

Pupil dilation reflects arousal-related neural processes and is closely linked to sensory perception, attention, and cognitive state, but the mesoscopic cortical dynamics that accompany stimulus-evoked dilation remain unclear. Here, we combined simultaneous pupillometry and wide-field Ca{superscript 2} imaging in mice with explainable machine learning to identify cortical activity patterns predictive of pupil dilation. Cortical activity was recorded during hindpaw somatosensory stimulation, visual pattern change, and visual contextual stimulation using an visual oddball paradigm; spontaneous pupil dilation was additionally analyzed using a public resting-state dataset. To reduce multicollinearity in wide-field imaging signals, cortical activity was decomposed into independent components, which were used as inputs to recurrent neural network (RNN) decoders. Models reliably predicted sensory-evoked pupil dilation above shuffled-label controls when sessions with pre-existing pupil dilation were excluded, indicating that stimulus-evoked dilation was associated with distinct cortical activity patterns. Feature attribution using SHapley Additive exPlanations (SHAP) and spatial back-mapping revealed condition-dependent cortical signatures. Activity around the retrosplenial cortex and higher-order visual areas contributed to predictions during somatosensory and visual pattern stimulation, whereas secondary motor cortex activity was prominent during visual contextual stimulation. Spontaneous pupil dilation required lower-frequency cortical signals for decoding, suggesting partly distinct dynamics from sensory-evoked dilation. These findings demonstrate that explainable machine learning can extract candidate cortical activity patterns associated with pupil-linked arousal from high-dimensional wide-field Ca{superscript 2} imaging data. Because feature importance does not establish causality, the identified regions provide hypothesis-generating targets for future circuit manipulation studies.