Machine learning to predict cure in severe pneumonia episodes

PROJECT ABSTRACT Pneumonia is the most common cause of death due to infection worldwide and in patients admitted to the intensive care unit (ICU). Some patients are successfully cured with a short course of antibiotics, but over a third are not readily cured despite appropriate therapy and have worse outcomes. Current ICU prediction models

focus on remote outcomes, such as hospital mortality, and are not built to predict short term outcomes of pneumonia episodes. These episodes evolve over days, which are the timeframe during which clinical interventions can shift patient trajectories toward favorable outcomes. I developed a machine learning approach

called CarpeDiem based on the practice of daily physician rounds to disentangle the contributions of unresolving pneumonia to poor outcomes in patients with respiratory failure. These data suggest identifying patients likely to fail pneumonia treatment early in their clinical course could impact ICU outcomes.

In this proposal, I will build upon my previous CarpeDiem methodology to develop and validate a machine learning model to predict whether a pneumonia episode will be cured on day 7 using features from days 1-3. I will leverage the data from the Successful Clinical Response in Pneumonia Therapy (SCRIPT) U19 Systems

Biology Center, a one-of-a-kind biorepository that includes over 700 patients and 1,400 bronchoalveolar lavages. I will use both clinical electronic health record data and single-cell RNA sequencing transcriptomic data in my models. My predictions and hypotheses are founded on causal biologic studies in murine models of lung injury

demonstrating the importance of regulatory T cells to promote pneumonia resolution. This proposal will give me training in deep learning methods, deployment of machine learning models, and computational skills to analyze high-dimensional single-cell transcriptomic data. My project addresses critically important gaps in current modeling and predictors of pneumonia outcome and

will overcome current limitations in machine learning approaches to ICU data. I am supported by my mentors Richard Wunderink, a world-renowned expert in pneumonia and critical care. My co-mentor Benjamin Singer provides lung immunology expertise, and my co-mentor Yuan Luo provides expertise in machine learning. My

Research Advisory Committee includes experienced scientists in computational transcriptomics, electronic health record modeling, database informatics, and lung immunology. My training plan leverages the unique research environment at Northwestern University to grow my substantial track record while expanding my skills

in electronic health record and genomic data to support my academic development as a future R01 funded investigator. Identifying early features of treatment failure provides targets for intervention, prompting additional diagnostics, such as repeat bronchoalveolar lavage or imaging, or treatment modifications. Recognition of early

treatment failure offers the opportunity to divert a patient’s downward trajectory toward cure in critically ill patients.