130 Desoto Street, Pittsburgh, 15261

View map

"Machine Learning in Sepsis Clinical Management" Department of Biostatistic and Health Data Science, School of Public Health. 

Advisors and Committee Chairs

  • Lu Tang
  • Victor Talisa

Abstract

Sepsis is a serious and potentially life-threatening condition caused by the body's response to an infection. It affects roughly two million individuals in the US each year. Despite decades of research, it remains a leading cause of mortality in critical care. One of the biggest challenges in sepsis management is heterogeneity. Patients can look very different from one another – in terms of their underlying conditions, disease progression, and response to treatment. Because of this variability, clinical decision-making is often difficult. This dissertation focuses on addressing challenges posed by patient heterogeneity in sepsis care using machine learning methods.

In the first project, we develop Federated Learning of Robust Individualized Decision Rules (FLoRI), a flexible machine learning-based individualized treatment framework that can be deployed in complex clinical settings. Through a novel objective function and federated learning, FLoRI provides safe treatment recommendations, improves generalizability across patient populations, and mitigates infrastructure limitations across healthcare systems simultaneously. The performance of FLoRI is demonstrated through an application to University of Pittsburgh Medical Center (UPMC) sepsis patients. FLoRI effectively improves the survival by 2-3 percentage points among sepsis patients and by 10 percentage points among sepsis patients with higher risk of death.

In the second project, we develop Tilted Individualized Decision Rules (TIDE), an individualized treatment framework that is robust to data contamination. Through tilted empirical risk minimization, TIDE down-weights corrupted observations and emphasizes the signal from the majority of the data, leading to more reliable treatment recommendations. TIDE is then extended to TIDE+ for multistage settings, allowing treatment decisions to be dynamically adjusted over time. The performance of TIDE is demonstrated using UPMC sepsis data. TIDE effectively mitigates the impact of contaminated observations and yields treatment recommendations that improve patient outcome compared to standard approaches.
In the third project, we develop Registered Trajectory Clustering for Sepsis Phenotyping (ReTraC), a clustering framework that identifies sepsis subphenotypes using electronic health records (EHRs). ReTraC jointly aligns and clusters patient trajectories through iterative updates. Both simulation studies and real-world applications to UPMC data demonstrate improved clustering accuracy and the identification of clinically interpretable subphenotypes.

Public health significance: This dissertation advances public health by addressing key challenges in sepsis care using real-world EHR data. The proposed methods improve the reliability of individualized treatment decisions and enable more precise identification of patient subgroups, ultimately supporting more effective and personalized care in critical care settings.

Event Details

Please let us know if you require an accommodation in order to participate in this event. Accommodations may include live captioning, ASL interpreters, and/or captioned media and accessible documents from recorded events. At least 5 days in advance is recommended.

University of Pittsburgh Powered by the Localist Community Event Platform © All rights reserved