1 Abstract
Sepsis is the leading cause of hospital mortality in the US and imposes immense health and economic burdens.
Optimizing clinical outcomes of sepsis hinges upon early detection, accurate classification of patient phenotype,
and prompt treatment. However, the understanding of sepsis at a pathophysiologic level is limited, making it
challenging for clinicians to effectively diagnose and treat sepsis patients. With technological advancement, mas-
sive electronic health records (EHRs) and registries have been generated by the routine collection of data from
medical and daily activities. These data provide exceptional opportunities for generating insights into improved
clinical decision-making, prognosis, and personalized medicine treatment strategy for sepsis among many other
diseases. On the other hand, EHR data are often high-dimensional, sparse, longitudinal, and include abundant
temporal information. Such complex structures pose great barriers to the use of traditional statistical and machine
learning methods. Although EHR data are often stored in relational databases, they can be represented by an
order-3 tensor, i.e., an array with three directions representing different subjects, features, and time points that
often vary greatly across subjects/features. This tensor-based representation presents a fresh perspective for
analyzing EHR data, offering a new opportunity to improve sepsis care.
In this proposal, we aim to improve the progress assessment, patient phenotyping, and treatment planning
of sepsis through an integrated tensor-based modeling framework to tackle the challenges associated with high-
dimensional EHR data and their complex temporal structure. The specific tasks include: a) developing an iterative
timeline registration algorithm to enhance the assessment of the progress of sepsis by leveraging the tensor
structure of EHR data; b) devising a tensor-based method for phenotyping using longitudinal high-dimensional
EHR data; c) creating a novel method to estimate the potential outcomes and treatment effects of time-varying
treatment regimes through low-rank tensor completion. The developed tools will become Sepsis Pulse, a module
to be incorporated into an existing digital workflow known as Sepsis Watch within Duke Hospital’s healthcare
system.