Identifying Sepsis Subtypes in Tanzania Using Prior Information

Recent studies have found that sepsis can be expressed as a spectrum of heterogeneous subtypes and that outcomes can be improved by treatments tailored to each subtype. However, sepsis subtypes have only been identified in cohorts in the USA and Europe. In this project, we derive subtypes within a group of patients from Tanzania, where patients tend to be younger, have been sicker longer, and have a higher prevalence of HIV infection than previous cohorts.

In “Bayesian Learning of Clinically Meaningful Sepsis Phenotypes in Northern Tanzania”, we propose CLustering Around Meaningful Regions (CLAMR), a novel prior for cluster analysis that incorporates medically interpretable cut-offs. We accomplish this by modeling the cluster center as a mixture over these meaningful regions (MRs). Keeping the components of these mixtures separated motivates default hyperparameter values, and we apply this method to the Investigating Febrile Deaths in Tanazania (INDITe) cohort.