Computational and neural mechanisms of statistical pain learning

Abstract

Pain invariably changes over time. These fluctuations contain statistical regularities which, in theory, could be learned by the brain to generate expectations and control responses. We demonstrate that humans learn to extract these regularities and explicitly predict the likelihood of forthcoming pain intensities in a manner consistent with optimal Bayesian inference with dynamic update of beliefs. Healthy participants received probabilistic, volatile sequences of low and high-intensity electrical stimuli to the hand during brain fMRI. The inferred frequency of pain correlated with activity in sensorimotor cortical regions and dorsal striatum, whereas the uncertainty of these inferences was encoded in the right superior parietal cortex. Unexpected changes in stimulus frequencies drove the update of internal models by engaging premotor, prefrontal and posterior parietal regions. This study extends our understanding of sensory processing of pain to include the generation of Bayesian internal models of the temporal statistics of pain.

Fig. 1. Behavioural task and model explanation.

a Example trials from a representative participant, showing: the true probability of high (H) and low (L) stimuli given current stimuli, trial stimulation given, and participant rated probabilities. The arrows point to jump points of true probabilities, where a sudden change happens. b Rating screens. Occasionally, the sequence was paused and participants were asked to estimate the likelihood of the upcoming stimulus given the current one. For example, after a low stimulus participants would be asked to rate the probability of the upcoming stimulus being low (L − > L) or high (L − > H). c Graphical representation of the Markovian generative process of the sequence of low and high-intensity stimuli. The transition probability matrix was resampled at change points, determined by a fixed probability of a jump.

Fig. 2. Behavioural results.

a Relation between generative and rated frequencies, p(H), in each participant (one regression line per participant, n = 35). b Relation between generative and rated transition probabilities, p(H∣L), in each participant (one regression line per participant, n = 35). c Estimation accuracy, as measured by the correlation coefficient between generative and rated probabilities: frequency p(H) and transition probability p(H∣L); each circle represents one participant (n = 35). The boxes show the quartile of the data and the whiskers illustrate the rest of the distribution, except for points that are classified as "outliers'' based on the inter-quartile range. The two sets of correlations were not significantly different, based on a two-sided z-test on Fisher-transformed correlation coefficients (z = 0.376, p = 0.707). Source data are provided as a Source Data file.

Fig. 3. Model comparison results.

a Bayesian model comparison based on model fitting evidence. Subjects' predictive ratings of next trial’s pain intensity were fitted with posterior means from Bayesian models, values from Rescorla–Wagner (reinforcement learning) model, and random fixed probabilities. The winning model was the Bayesian jump frequency model, which assumes jumps in the sequence and infers the stimulus frequency. In our model comparison, the model frequency indicates how often a given model is used by participants; the model exceedance probability measures how likely it is that any given model is more frequent than the other models, and the protected exceedance probability is the corrected exceedance probability for observations due to chance. b Individual subject model evidence (each row represents a subject; colorbar indicates the model probability ranging from 0 to 1). Source data are provided as a Source Data file.

Fig. 4. Brain responses to noxious stimuli.

Red: high > low pain stimuli, blue: high < low pain stimuli. Two-sided statistics corrected for the false discovery rate (FDR) at level p < 0.001; colorbar shows Z scores > 3.3.

Fig. 5. Brain activity associated with the temporal statistical inference of pain intensity.

Neural correlates of the mean posterior probability of low pain (green) and high pain (pink) in the Bayesian jump frequency model (two-sided statistics, FDR corrected p < 0.001, colorbar shows Z scores > 3.3).

Fig. 6. Uncertainty and learning signals.

a Uncertainty (SD) of the posterior probability of high pain in the Bayesian jump frequency model was associated with activations in the right superior parietal cortex (FDR corrected p < 0.001, colorbar shows Z scores > 3.3). b Neural activity associated with the model update, i.e. the Kullback–Leibler (KL) divergence between posteriors from successive trials (positive contrast, FDR corrected p < 0.001, colorbar shows Z scores > 3.3).

Fig. 7. Statistical inference and model update activate adjacent sensorimotor and premotor regions.

Overlay of the temporal prediction (mean posterior probability) of low (green) and high pain (pink), their uncertainty (SD posterior probability, blue) and the model update (KL divergence between successive posterior distributions, red-yellow); FDR corrected p < 0.001, colorbar shows Z scores > 3.3.