Speaker: Yibei Chen (Ghosh Lab)

Title: Behavior and event history predict brain-state dynamics during continuous gameplay

Abstract: Spontaneous behavior and recent event history explain a substantial fraction of neural variance in animal systems neuroscience. Whether human brain dynamics under continuous, hours-long naturalistic behavior follow a similar organization has been hard to test, because most human paradigms rely on resting state or passive stimuli. We analyzed 87 hours of within-subject fMRI from five CNeuroMod participants who each completed 15–20 hours (18–30 sessions) of Super Mario Bros. gameplay with concurrent electrodermal activity and respiration. Per-subject Gaussian hidden Markov models on parcellated, PCA-reduced BOLD yielded 9–11 active brain states per individual. Despite independent per-subject fits, eight states recurred across all five participants (Hungarian-aligned mean cosine = 0.72; network-block permutation p < 0.001). In every subject, action rate and time since the last death ranked among the top three predictors of both state identity and transition destination, while slow stimulus context (level identity, world position) did not. Death and level-completion events preceded HRF-locked transition bursts in every subject (max-statistic permutation p < 0.001). The state architecture partially transferred to a second game (Shinobi); the event vocabulary did not — reward-rate features dominated where time-since-death had dominated. Autonomic signals indexed states reliably (breathing-rate ICC = 0.83) but did not determine them (arousal–epoch-duration \|ρ\| < 0.12 in every subject). Brain-state dynamics during interactive behavior track the action–outcome loop more than the slow stimulus context.