Confidence-weighted nociceptive learning: behavioral evidences and EEG correlates

Pain perception is a warning signal enabling animals to preserve their body integrity. When pain persists over time, the brain needs to learn to predict its temporal evolution in order to minimize harm. Whereas several studies have already shown the effects of experience and expectations on responses to isolated nociceptive stimuli using axiomatic approaches, it remains largely unknown how individuals can learn to predict a stream of nociceptive events and how these processes are implemented in the brain. To clarify these aspects, we exposed healthy human participants (n = 31) to probabilistic sequences of contact thermal stimuli of two clearly distinct intensities while recording their brain activity with electroencephalography (EEG). Each sequence was generated using two Markovian transition probabilities (TPs) and participants were instructed to try predicting the probability of forthcoming intensities. Participants were occasionally asked to report their probability estimates and confidence in these predictions on numerical scales. First, we found that the probability and confidence reports were well approximated by a Bayesian model learning the sequence TPs. At the opposite of non-Bayesian models, this suggests that participants employed a confidence-weighted learning strategy: when confidence gets higher, the effect of newly received stimulus intensities is reduced. Then, we observed a negative correlation between the model confidence and the amplitude of the Vertex Potential (VP): the higher the confidence in the TP estimates, the smaller the VP. Overall, these findings confirm key predictions of a Bayesian learning model for nociception and suggest that the VP is modulated by inferential confidence when participants learn the structure of sequences of nociceptive stimuli.