B1850
Title: Bayesian inhomogeneous hidden Markov model with incomplete observations and its application to EHR modelling
Authors: Dongrong Li - The Chinese University of Hong Kong (Hong Kong) [presenting]
Wing Ki HUI - The Chinese University of Hong Kong (Hong Kong)
Xiaodan Fan - The Chinese University of Hong Kong (Hong Kong)
Abstract: A novel Bayesian inhomogeneous hidden Markov model is introduced, designed to accommodate missing features and observable states. The approach draws inspiration from the complexities inherent in medical data analysis, where the actual disease status is typically modelled as unobservable latent variables. Concurrently, numerical features and medical diagnoses, which are often riddled with missing entries, are regarded as inaccurate outcomes. The inhomogeneous hidden Markov model is expanded, the missing observations are modelled as latent variables and those are incorporated into the inhomogeneous transition and emission probabilities via multinomial logistic regression models. The work further proposes an innovative forward-filtering backward sampling (FFBS) algorithm, which is designed to sample from the conditional distribution of latent sequences based on incomplete observations and numerical features. Beyond this, the conditional distributions of other parameters and latent variables are extrapolated, leading to the derivation of a Gibbs sampler that efficiently samples from the full posterior. To empirically validate the efficacy of the proposed model, numerical experiments are executed on simulated and validation datasets.
