Title: Coupled Markov-switching regression models with application to electronic health record data
Authors: Jennifer Pohle - Bielefeld University (Germany) [presenting]
Roland Langrock - Bielefeld University (Germany)
Ruth King - University of Edinburgh (United Kingdom)
Mihaela van der Schaar - University of Oxford (United Kingdom)
Abstract: Hidden Markov models (HMMs) are time series models which assume the observations to depend on an underlying unobserved Markov chain with finitely many states. They have been applied in many different areas, for instance in speech recognition, finance, medicine, and ecology. In the case of multivariate time series, within a basic HMM formulation, the variables would be expected to evolve synchronously in the sense that they are driven by the same underlying state sequence. However, in some applications, e.g. in medicine, the observed variables do not necessarily evolve in lockstep, although they may be correlated. Coupled hidden Markov models overcome this limitation by assuming separate but correlated state sequences to underlie the different variables observed, hence coupling the state processes of multiple HMMs. However, the observations often depend not only on the underlying state, but also on external factors. Therefore, we extend coupled HMMs to allow for covariates in the observation processes, which leads to the flexible class of coupled Markov-switching regression models. We apply this method to electronic health record data collected for 842 patients within the medical intensive care unit at the University of California in Los Angeles.