Human single-neuron recordings reveal population coding of attentional dynamics during naturalistic movie viewing.

Our eyes move constantly. Where they land is shaped by internal cognitive state, making eye movements a rare, non-invasive window onto that state. However, how moment-to-moment eye-movements map onto human neural population activity during naturalistic viewing remains poorly characterized. We analyzed a public intracranial dataset in which 14 neurosurgical patients watched an eight-minute movie during simultaneous eye-tracking and neuronal activity recording. Single-neuron activity from 814 neurons across anterior cingulate cortex, pre-supplementary motor area, amygdala, hippocampus, and ventromedial prefrontal cortex was pooled into a single distributed population and related to a joint eye-movement state comprising pupil size, saccade rate and fixation duration. Using a cross-validated, multivariate analysis of population activity we found group-level coupling between the neural population and the joint eye-movement state (mean held-out canonical r = 0.145; one-sample t-test p = 0.0003; 9 of 14 subjects individually significant). The coupling was carried principally by saccade dynamics (r = 0.141) and pupil size (r = 0.094), with fixation duration contributing only marginally (r = 0.041). At the single-neuron level the picture differed: more neurons were eye-coupled than expected by chance, yet so weakly that almost none survived correction for multiple comparisons. Read out jointly, the same weakly coupled cells cohered into one reliable population dimension, and no single region's removal significantly reduced it. This pattern is more consistent with a distributed, redundant organization than with a small set of strongly coupled cells. In summary, eye-tracking offers a non-invasive window onto population-level neural states in humans.