UCSF

Assessing extubation readiness and determining the earliest safe time to extubate patients in the Pediatric Intensive Care Unit (PICU) are challenges clinicians face daily. Until recently, no consensus guidelines existed for managing pediatric extubation, and studies have concluded that the decision to extubate ultimately relies on clinician judgment. This variability in care leads to increased morbidity, mortality, and costs, arising from both unnecessary ventilator days due to delayed extubation and from re-intubation following extubation failure. The power of artificial intelligence could be harnessed to optimize identification of extubation readiness in the PICU. Deploying prediction models in the electronic health record (EHR) as clinical decision support tools could safely shorten extubation times by decreasing variation in care and identifying subsets of patients for earlier, safe extubation.

The objective of this dissertation is to predict extubation readiness using methods that could be implemented as real-time clinical decision support in a health system. Chapter 1 explores how to handle missing longitudinal data for clinical prediction models. Chapter 2 compares machine learning models built with EHR data to predict extubation readiness. Chapter 3 demonstrates a novel method to integrate expert knowledge directly into machine learning models for this prediction problem. This project will advance extubation practices for critically ill children, yielding a predictive tool ready for prospective testing in the EHR that moves toward delivering high reliability healthcare for patients with respiratory failure.