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

A practical solution to these problems has been using hidden Markov models (HMMs), which (i) can A particular challenge for disease progression be trained using small datasets, (ii) can handle missing data modeling is the heterogeneity of a disease and in a principled approach and (iii) are interpretable models, its manifestations in the patients. Existing approaches e.g., it is possible to relate inferred latent states to particular often assume the presence of a single symptoms. Most existing HMMs (Jackson et al., 2003; disease progression characteristics which is unlikely Sukkar et al., 2012; Guihenneuc-Jouyaux et al., 2000; Wang for neurodegenerative disorders such as et al., 2014; Sun et al., 2019; Severson et al., 2020; 2021), Parkinson' disease. In this paper, we propose however, assume that each patient follows the same latent a hierarchical time-series model that can discover state transition dynamics, ignoring the heterogeneity in the multiple disease progression dynamics. The proposed disease progression dynamics.