Individual heartbeats track distinct prediction processes during human probabilistic learning

Abstract

Heart rate continuously adjusts to accommodate perception and action, and these shifts are frequently explained through predictive processes. Yet direct evidence that interbeat intervals exhibit graded scaling with prediction remains limited. Here we investigate millisecond-resolved physiological signatures of prediction processing using a mechanistically constrained analysis of beat-to-beat cardiac dynamics. We analysed trial-by-trial electrocardiogram and electroencephalogram recordings from 34 participants performing a probabilistic learning task. We quantified stimulus-locked cardiac responses as changes between consecutive interbeat intervals and accounted for cardiac phase at feedback. This single-beat approach separated anticipatory slowing, stimulus-locked parasympathetic brake, and subsequent acceleration. Anticipatory deceleration and rebound acceleration scaled with model-derived expectations, whereas the second heartbeat after feedback tracked signed prediction errors and outcome valence, particularly when feedback occurred early in the cardiac cycle. Peak stimulus-locked cardiac deceleration covaried with parietal P3b rather than prediction features. Thus, individual cardiac cycles carry separable signatures of anticipation, orienting, and feedback-based updating. These findings demonstrate how predictive processing propagates into human autonomic physiology on a beat-to-beat timescale and provide an interpretable, mechanistically grounded framework for quantifying brain-body co-modulation during adaptive behaviour.