Naturalistic affective behaviors decoded from spectro-spatial features of multi-day human intracranial recordings

Task-based studies have uncovered distributed neural networks that support emotions, but little is known about how these networks produce affective behaviors in non-laboratory, ecological settings. We obtained continuous intracranial electroencephalography (iEEG) recordings from the emotion network in 11 patients with epilepsy during multi-day hospitalizations. We coded naturalistic affective behaviors (spontaneous expressions of positive or negative affect) from 116 hours of time-locked video recordings obtained over multiple days from subjects hospital rooms and utilized data driven classifiers to determine whether we could decode naturalistic affective behaviors from the neural data. Results indicated that binary within-subject random forest models could decode positive and negative affective behaviors from affectless behaviors (behaviors lacking valence) with up to 93% accuracy. Across the emotion network, positive and negative affective behaviors were associated with increased high frequency activity and decreased lower frequency activity. The anterior insula, amygdala, hippocampus, and anterior cingulate cortex (ACC) made strong contributions to affective behaviors in general. In a subset of subjects, three-state decoders distinguished among the positive, negative, and affectless behaviors using the spectro-spatial features from the emotion network. This study demonstrates that multi-day, highly resolved iEEG recordings in cortical and deep brain structures can reveal the circuit-level physiology of affective behaviors. By measuring behavior in an ecologically valid setting, our findings provide novel insights into the spatially distributed dynamics of local neural populations underlying naturalistic affective behaviors.