Detecting latent pain states in chronic pain using expectation and perception: a Bayesian multilevel Hidden Markov Model approach

OBJECTIVES: To examine whether subjects with chronic pain experienced discernible pain states that were also reflected in their expectations. MATERIALS: Questionnaires about pain perception at time t and expectations at time t+1 were administered 5 times a day for 15 days to 40 subjects. Additionally, values for different mental variables were collected at baseline. METHOD: using mHMMBayes in R, we modeled two/three pain states and their transition probabilities. We produced four models with different sensible starting values using uninformative priors. RESULTS: 2 states: Convergence of the mcmc algorithm across chains was acceptable in estimating both emission probabilities and transition probabilities (R_hat <= 1.20). No significant label switching problems arose. 3 states: not supported by data. Models converged poorly, with significant missingness when sampling values from the posterior for the emission values. DISCUSSION: Our analyses show that a two state model reflected well in our data, both at an inter-subject and between-subject level. However, the three state models performed poorly. Taken together, these results tepidly suggest that subjects with chronic pain experience moments of high and low pain, here defined as a combination of actual perception and expectation. Moreover, subjects tend to remain in one state or the other. CONCLUSIONS: While the results are encouraging, the study is severely limited by low sample size, low granularity of beeps, and low number of variables measured. Still, the good fit of our model should be encouraging in persisting in this direction, by collecting more data at a denser granularity. It should be noted that, while the three state model is not currently supported by our data, collecting more observations over longer time periods may render it more appropriate.