Logo image
Back
The clinical epidemiology of sepsis: balancing sepsis program and antibiotic stewardship initiatives
Dissertation

The clinical epidemiology of sepsis: balancing sepsis program and antibiotic stewardship initiatives

Nicole Rafalko
Doctor of Philosophy (Ph.D.), Drexel University
Apr 2026
DOI:
https://doi.org/10.17918/00011333
pdf
Rafalko_Nicole_202626.78 MB
PDF Embargoed Access, Embargo ends: 30 Apr 2027

Abstract

Antimicrobial stewardship Clinical epidemiology Healthcare utilization Sepsis
Sepsis continues to be a major contributor of morbidity, mortality, and healthcare costs in the US. Sepsis patients require rapid clinical decision-making, yet current guidelines often apply a "one-size-fits-all" approach that may not account for patient heterogeneity and may contribute to unintended consequences of antibiotic use. This dissertation examines strategies to improve sepsis management and antibiotic stewardship in critically ill patients. The specific aims are to: 1) evaluate the effect of early antibiotic treatment on mortality among critically ill patients with suspected sepsis, 2) develop a predictive model using electronic health record data and machine learning to identify patients with suspected sepsis who are most likely to have positive cultures, and 3) assess the impact of antibiotic discontinuation among patients with culture-negative suspected sepsis on outcomes including Clostridioides difficile infection, hospital length of stay, and mortality. This dissertation used a retrospective cohort of critically ill patients with suspected sepsis identified from electronic health record data in the Medical Information Mart for Intensive Care IV database. In Aim 1, we emulated a target trial using the cloning-censoring-weighting approach, defining early treatment as antibiotic administration within 3 hours of sepsis onset and evaluating effects among patients with greater illness severity and those with confirmed bacterial infection. In Aim 2, we developed predictive models using a large set of candidate predictors from the electronic health record and compared the performance of a penalized regression approach (LASSO) with a machine learning algorithm (XGBoost). Finally, for Aim 3, we used inverse probability of treatment weighting to estimate effects of antibiotic discontinuation on Clostridioides difficile infection, hospital length of stay, and mortality under alternative definitions of antibiotic discontinuation. In Aim 1, early antibiotic administration within 3 hours did not significantly change overall 7- or 30-day mortality, though among patients with severe illness (on vasopressors), early treatment was associated with lower 7-day mortality. Aim 2 demonstrated that predictive models using high-dimensional electronic health record data could identify patients most likely to have culture-confirmed infection, with XGBoost outperforming LASSO regression. Key predictors included suspected infection site, organ dysfunction site, prior antibiotic use, and laboratory markers. In Aim 3, antibiotic discontinuation within 96 hours among patients with culture-negative sepsis was associated with lower odds of 30-day Clostridioides difficile infection and 30-day mortality, as well as shorter hospital length of stay. A dose-response pattern was observed, with earlier discontinuation associated with greater reductions, while later discontinuation still offered benefit. Together, these findings suggest that a one-size-fits-all approach to antibiotic administration may be ineffective and highlight the importance of targeting patients most likely to benefit while strengthening antibiotic stewardship to prevent unintended harms.

Metrics

1 Record Views

Details

Logo image