Learning the statistics of pain: computational and neural mechanisms

Pain invariably changes over time, and these temporal fluctuations are riddled with uncertainty about body safety. In theory, statistical regularities of pain through time contain useful information that can be learned, allowing the brain to generate expectations and inform behaviour. To investigate this, we exposed healthy participants to probabilistic sequences of low and high-intensity electrical stimuli to the left hand, containing sudden changes in stimulus frequencies. We demonstrate that humans can learn to extract these regularities, and explicitly predict the likelihood of forthcoming pain intensities in a manner consistent with optimal Bayesian models with dynamic update of beliefs. We studied brain activity using functional MRI whilst subjects performed the task, which allowed us to dissect the underlying neural correlates of these statistical inferences from their uncertainty and update. We found that the inferred frequency (posterior probability) of high intensity pain correlated with activity in bilateral sensorimotor cortex, secondary somatosensory cortex and right caudate. The uncertainty of statistical inferences of pain was encoded in the right superior parietal cortex. An intrinsic part of this hierarchical Bayesian model is the way that unexpected changes in frequency lead to shift beliefs and update the internal model. This is reflected by the KL divergence between consecutive posterior distributions and associated with brain responses in the premotor cortex, dorsolateral prefrontal cortex, and posterior parietal cortex. In conclusion, this study extends what is conventionally considered a sensory pain pathway dedicated to process pain intensity, to include the generation of Bayesian internal models of temporal statistics of pain intensity levels in sensorimotor regions, which are updated dynamically through the engagement of premotor, prefrontal and parietal regions.