Grant: Travel Awards - Institute Community Support

Health system managers track the health of Canadians using health indicators; one example of an indicator is the number of people admitted to hospital with COPD (a severe lung condition). Although the COPD admission indicator can identify hospitals or regions with more admissions than average, it does not help us understand why or, more importantly, what can be done to prevent those admissions. Large amounts of data in the form of people's longitudinal records, such as electronic health records (EHRs) and healthcare administrative database, are now available to enable dynamical monitoring of unknown processes governing the observations. However, such latent progression manipulating the observations is not observed directly and so requires inferential methods to ascertain progression. The patient data is still, however, very complex; to identify patient trajectories automatically. I developed a continuous-time hidden Markov model (CTHMM) under the framework of GLM to analyze the issues of irregular visits, dierent types of observations and multiple time-dependent covariates. Both likelihood and Bayesian inferences for CTHMM with finite state space are investigated via the expectation-maximization (EM) algorithm and Gibbs sampling respectively. We apply this model to a COPD cohort from health administrative data with the number of drugs taken as observations and the types of healthcare utilization as a covariate. A three-state model is chosen based on AIC and BIC. The individual rate matrix suggests that the patients are more likely to progress to more severe state if they are old when they entered the cohort.