Emotional Context Modulates Neural Prediction-Error Encoding

Emotional experiences are closely intertwined with reinforcement learning, yet the extent to which emotional context influences the neural encoding of learning signals remains unclear. The present study examined whether emotional context modulates trialwise neural prediction error (PE) encoding during reinforcement learning. Fifty-six participants completed a drifting three-armed bandit task while electroencephalography (EEG) was recorded. Choice options were consistently paired with positive, neutral, or negative emotional images that were presented during feedback processing. Trial-by-trial PEs were estimated using a computational reinforcement-learning model and related to feedback-locked EEG activity using hierarchical linear mixed-effects modeling. Traditional condition-averaged analyses revealed robust feedback-related ERP responses to reward outcomes but minimal effects of emotional context on overt behavior or mean reward positivity (RewP) amplitude. In contrast, trialwise analyses demonstrated that emotional context significantly moderated the relationship between PE and feedback-locked neural activity. The relationship between trialwise PEs and EEG amplitude was strongest in neutral contexts, whereas it was attenuated in positive emotional contexts during the early RewP interval and in negative emotional contexts during a later posterior stage of feedback processing. Consistent with this temporal pattern, scalp topographies revealed a corresponding shift from early frontocentral to later posterior activity over the course of feedback processing. Together, these findings suggest that emotional context does not substantially alter overt learning performance or condition-averaged reward responses, but instead subtly shapes the neural encoding of PEs during reinforcement learning. More broadly, the results highlight the value of computationally informed trial-level analyses for identifying affective influences on learning signals that may be obscured by traditional condition-averaged approaches.