Inferring Binary Feature Profiles Underlying Patient Health, with Applications to Sepsis

About This Grant

Project Summary/Abstract Sepsis is an often life-threatening medical condition in which the body has a dysregulated response to in- fection. A number of studies have attempted to identify sepsis subtypes based on patient similarities across demographic, clinical, and biological/immunological data, with the hope of identifying subtypes responsive to personalized treatments. Although clustering is a promising approach, its significant limitations are not always well recognized. These include a lack of robustness to small differences in datasets and a lack of reproducibility across different study types and populations. This project develops new statistical methods designed to help clinicians in the search for subtypes of disease, while taking advantage of heterogeneous data both within and across different studies. With improved statistical tools, medical researchers will have a better chance of iden- tifying meaningful subtypes that can be further interrogated in the search for improved treatments that lead to better survival. Specifically, our project aims to (1) develop, validate, and apply Bayesian multiview pyramids as an alternative to standard clustering techniques; (2) develop, validate, and apply Bayesian multistudy pyramids to maximize reproducibility across different studies and populations; and (3) develop, validate, and apply robust frameworks for assessing how disease subtypes evolve dynamically over time. We will study properties of these new approaches both from a theoretical perspective and via extensive simulation studies with comparisons to other leading approaches, and we will use these approaches to explore sepsis subtyping in a number of im- portant sepsis cohorts. While the disease under study in this application is sepsis, the methods will be broadly applicable to other diseases and syndromes for which clustering could be used to facilitate endotype discovery.