Mixture of Input-Output Hidden Markov Models for Heterogeneous Disease Progression Modeling

Learning the progression patterns of treatments using a probabilistic generative model

“…The main contribution of this paper is the development of a novel probabilistic generative model, which characterizes the progression of the treatment trajectories of a disease. State-of-the-art disease progression approaches [5,6,7,8,11,12,13,14,15,16,17,18,19,20,21] partially adopt the main properties of our model, which we consider essential in order to describe and understand the behavior of the treatment trajectories. In particular, our model simultaneously classifies the heterogeneous sequences of actions based on their treatment evolution over time, segments the sequences of actions in different progression stages of the disease, and captures the sequential dependence between medical actions.…”

“…Most existing HMMs [10,11,12,13,14,15,16] assume that all patients evolve through the same latent state transition dynamics, thus ignoring the heterogeneity of different subtypes of disease progression. Other probabilistic approaches that simultaneously address disease state progression and treatment subtyping [17,18,19] are limited to model the evolution of observed data through a latent process and do not handle the sequential dependence within medical actions. These methods actually model the number of each type of action that occurs in each stage, rather than being generative models of the sequence of actions.…”

A probabilistic generative model to discover the treatments of coexisting diseases with missing data

“…Probabilistic models are a practical solution to face this challenge, not only because they can handle missing data, but also because they account for temporal relationships in data and are interpretable models that can extract clinically meaningful representations from the inferred latent variables. In the literature, most probabilistic models developed for disease progression are based on variants of Hidden Markov models [2,7,8,9,10,11,12] or are extensions of latent Dirichlet allocation [13,14] that capture the evolution of disease trajectories through latent states. While medical events are time-dependent variables, these models generally ignore the direct stochastic dependence between such observations and are limited to modeling sequential correlations of data only through latent states [11].…”