Improving prediction of persistent hypoxemic respiratory failure to support clinical trial enrichment

About This Grant

Abstract: Over 30% of patients who develop acute hypoxemic respiratory failure (AHRF) and require ventilator support will die in the hospital. Treatment for AHRF remains extremely limited, as nearly every clinical trial in the last 20 years has failed to demonstrate improvements in mortality. A major obstacle for these trials is that most patients improve with existing care; this dramatically limits our ability to detect benefits for the remaining patients with persistent HRF, who are the ones most at-risk for death and in-need of new therapies. Currently, there are no accurate ways to distinguish patients with persistent HRF early, when trial enrollment and intervention is critical. The goal of this project is to develop robust models to identify patients at high risk for persistent HRF early, by using innovative opportunities in data science and machine learning to capture complex data sources (text and imaging) and accurately predict risk. The project will also allow Dr. Neha Sathe, an early career investigator and Pulmonary & Critical Care physician, to gain expertise in state-of-the- art methods to develop, evaluate, and improve multisource prediction models in real-world settings. In Aim 1, Dr. Sathe will develop and validate models that identify patients at high risk for persistent HRF, by integrating retrospective data from electronic health records, chest radiograph reports, and banked blood at two medical centers. In Aim 2, she will evaluate and explain these models in a new prospective cohort, to develop strategies and infrastructure for deploying and monitoring these models in future work. In Aim 3, she will develop novel methods to analyze sources of data with high potential to improve prediction of persistent HRF (chest radiograph images and tracheal aspirates, which are readily collected but under-leveraged). This work will yield models that improve the ability of trials to identify effective therapeutics for high-risk patients, minimize exposure to potentially costly or toxic therapeutics in patients likely to resolve, and provide significant insights to advance precision medicine. This work will also yield research infrastructure that can be adapted to rigorously develop and test predictive models for other clinical problems in critical care. To achieve these aims, Dr. Sathe will have complementary mentorship across the thriving ecosystem at the University of Washington for translational AHRF research, medical data science, and informatics. Altogether, this proposal aligns with NIH strategic objectives to leverage new opportunities in data science, and will support Dr. Sathe's long-term goal of understanding how to best use these opportunities to individualize and improve the care of patients with AHRF.