UC San Diego

Sepsis remains a persistent and severe problem in healthcare, accounting for over 1.7 million adult cases and resulting in approximately 350,000 deaths annually in the United States alone. Predictive modeling and machine learning hold immense potential to assist with the early recognition of sepsis and other critical illnesses. Yet despite the rapid advancements in healthcare technology and data analytics, there remains a considerable disconnect between the development of predictive algorithms and their practical application at the patient bedside. The reasons are multifaceted and include (1) technical barriers to effective electronic health record (EHR) integration, (2) challenges in validating the clinical utility of algorithms, (3) ‘one-size-fits-all’ approaches that fail to capture the relevant context from a heterogeneous and complex patient population, (4) low EHR data quality with high rates of missingness, and (5) delayed feedback loops from reportable quality measures to assess sustained improvement to quality of care. The aim of this dissertation is to address the aforementioned challenges. First, I designed a scalable predictive analytics platform that enables modular deployment of deep learning models, including large foundation models, directly onto the EHR. I then clinically validated and implemented a model for the early prediction of sepsis onto this platform. Finally, I improved the clinical utility of this platform for the management of sepsis through digital phenotyping, active sensing, endotyping with novel bioassays, and automated assessment of the Center for Medicare and Medicaid Services (CMS) Severe Sepsis and Septic Shock (SEP-1) measure.