Bayesian inference for the Markov-modulated Poisson process with an outcome process

Abstract

In medical research, understanding changes in outcome measurements is crucial for inferring shifts in a patient's underlying health condition. While data from clinical and administrative systems hold promise for advancing this understanding, traditional methods for modelling disease progression struggle with analyzing a large volume of longitudinal data collected irregularly and do not account for the phenomenon where the poorer an individual's health, the more frequently they interact with the healthcare system. In addition, data from the claim and health care system provide no information for terminating events, such as death. To address these challenges, we start from the continuous-time hidden Markov model to understand disease progression by modelling the observed data as an outcome whose distribution depends on the state of a latent Markov chain representing the underlying health state. However, we also allow the underlying health state to influence the timings of the observations via a point process. Furthermore, we create an addition "death" state and model the unobserved terminating event, a transition to this state, via an additional Poisson process whose rate depends on the latent state of the Markov chain. This extension allows us to model disease severity and death not only based on the types of care received but also on the temporal and frequency aspects of different observed events. We present an exact Gibbs sampler procedure that alternates sampling the complete path of the hidden chain (the latent health state throughout the observation window) conditional on the complete paths. When the unobserved, terminating event occurs early in the observation window, there are no more observed events, and naive use of a model with only "live" health states would lead to biases in parameter estimates; our inclusion of a "death" state mitigates against this.