Using Big Data to Improve Bacterial Sepsis Risk Stratification Among Immunocompromised Cancer Patients

Thesis (Ph.D.)--University of Washington, 2021Allogeneic hematopoietic stem cell transplant (HCT) recipients are an immunocompromised population that is disproportionately affected by sepsis, a life-threatening dysregulated immunologic response to an infection. While it is well established that early detection and treatment of sepsis with fluids and broad-spectrum antibiotics reduce the risk of mortality, recent data suggests early broad-spectrum antibiotic use in allogeneic HCT recipients may have microbiota-mediated detrimental effects on morbidity and mortality. Because of the risks associated with both missed and falsely identified sepsis events among allogeneic HCT recipients, early and accurate sepsis diagnosis is crucial. However, sepsis is generally challenging to diagnose and is made more complicated in allogeneic HCT recipients by the fact that sepsis presents differently following transplantation and common complications of the transplant procedure present like sepsis. In previous work, we demonstrated that current sepsis clinical prognostic tools have limited predictive validity among allogeneic HCT recipients and concluded that population specific prediction tools are needed. In the following papers, we outline the development and validation of two automated population-specific prognostic tools for bacterial sepsis (a full predictor and a clinical-factor specific tool) and two bedside tools (a tree and a score) specifically for use in data or computationally limited scenarios (ie. outside of intensive care units - ICUs). Each tool was developed to provide bacterial sepsis risk stratification at a crucial time in patient management (blood culture collection) in both inpatient and outpatient settings. Our performance-based validation suggests that our full predictor automated tool (super HCT bacterial sepsis learner - SHBSL) had superior predictive validity among allogeneic HCT recipients and that, compared with existing tools, our bedside tools (HCT Bacteremia Sepsis Score - HBS2 and HCT Bacteremia Sepsis Tree - HBST) had superior predictive validity among patients with blood cultures collected outside of ICUs. Through our novel health outcomes-based tool validation, we showed that SHBSL led to meaningfully fewer deaths and little additional antibiotic use compared with existing tools among a simulated cohort of 2000 allogeneic HCT recipients followed for 100 days post transplantation and that HBST and HBS2 provided the best alternative to SHBSL in scenarios for which the use of an automated tool may be improbable. Taken together, these findings support the use of SHBSL when possible and HBST or HBS2 in situations for which computationally intensive tools are impractical