Bayesian Nonparametric Clustering of Continuous-Time Hidden Markov Models for Health Trajectories

We develop clustering procedures for healthcare trajectories based on a continuous-time hidden Markov model and a generalized linear observation model. Specifically, we carry out Bayesian nonparametric inference for a Dirichlet process mixture model, and utilize restricted Gibbs sampling split-merge proposals to achieve inference using Markov chain Monte Carlo. In our analysis on a large Canadian cohort of subjects suffering from chronic obstructive pulmonary disease, a three-cluster model is chosen, and the inferred Markov transition rate matrix in each cluster suggests that each cluster has its own transition characteristics and observation process, and that patients are more likely to stay in more severe disease states.