Modeling Bacterial Infection Risk for Data-Driven Antibiotic De-Escalation in Critically Ill  Adults

Eickelberg, Garrett

doi:https://doi.org/10.21985/n2-bb0s-bc62

Work

Modeling Bacterial Infection Risk for Data-Driven Antibiotic De-Escalation in Critically Ill Adults

Public

Download PDF

Download All Files (.zip)

Bacterial infections (BI) are a frequent, expensive, and life-threatening condition for critically ill patients. For patients with serious BI, minimizing the time between admission to the intensive care unit (ICU) and administration of appropriate antibiotic therapy is crucial to improve prognosis. However, the current gold-standard for identifying the appropriate antimicrobials to administer, microbiology cultures, have long resolution times and are rarely able to guide early antibiotic treatment choices in the ICU. Consequently, critical care providers are advised to broadly administer empirical antibiotics to all patients suspected of BI, then adjust the treatment regimen based on follow-up information. This approach presents challenges in cases where the infection status is uncertain and current methods of characterizing BI risk underperform. In this paradigm, populations of patients with low risk of BI are exposed to unnecessarily prolonged antibiotic regimens and experience iatrogenic harm as a result. In this thesis, we demonstrated how leveraging electronic health record data with statistical learning techniques and informatics tools can supplement existing BI detection methods to inform antibiotic de-escalation decisions in the ICU. First, we developed and optimized a modeling framework to predict patient-level BI risk using an open and de-identified ICU database. Next, we developed and validated an open-source python package (MicrobEx) to extract BI status concepts from free-text microbiology reports. Then we performed an external validation and transportability evaluation of our BI modeling architecture in two unaffiliated tertiary intensive care unit (ICU) settings and a community ICU setting. Finally, using these same data sources, we performed a retrospective impact study to estimate the treatment effect of prolonging antibiotic therapy past 96 hours in critically ill patients predicted to have low risk for BI and adjust for selection bias using propensity score matching. Our analyses showed that sensitive and transportable performance can be achieved by using longitudinal patient features, such as temperature and white blood cell count, to predict BI status with our modeling framework. Furthermore, we present compelling evidence that critically ill patients who are predicted to be at low risk for BI can experience improved outcomes when discontinued from antibiotic therapy prior to 96 hours. To our knowledge, these analyses are the first to utilize EHR-based clinical prediction modeling to help guide antibiotic de-escalation decisions in critically ill adults.

Creator