Approved Research

Identifying sepsis risk factors by learning the causal structure using deep neural networks

Lay summary

Sepsis contributes to 1 of every 5 deaths. To improve clinical outcomes for sepsis, it is essential to identify causal relations among clinical features since the causal relations can be used to identify potential therapeutic agents that can control sepsis. For example, if an abnormally high level of a certain trait plays a causative role in sepsis' pathobiology, treating with an agent to control the trait may lead to an effective treatment. In this proposal, we will generate hypotheses on causal relations among the UKBB features that may be associated with sepsis risk by developing a deep-learning-based causal inference method and running it on multiple groups of UKBB subjects. Specifically, we will select subjects who have experienced sepsis (sepsis case), trauma (trauma control), and have not had major health problems (healthy control). By learning causal structure and comparing the structure between the sepsis case and the trauma control, we will identify risk factors that can be controlled in hospitalization. By learning causal structure and comparing the structure between the sepsis case and the healthy control, we will bring an understanding of environmental factors associated with sepsis. Due to the high prevalence of sepsis in deaths worldwide, learning the causal structure that may lead to sepsis has important value for public health. We expect this study to take 4~5 years to finish and publish multiple results in journals.