Patients with acute respiratory failure (ARF) and sepsis (the most common and lethal cause of the acute respiratory distress syndrome) are commonly admitted to intensive care units (ICUs) in the United States even when they do not require life support. However, ICU admission rates for these patients vary considerably among hospitals and even within hospitals over time due to bed availability. This practice variability suggests that some patients may be “under-triaged” to wards when their outcomes could be improved by sending them directly to the ICU (e.g., due to closer nurse or respiratory therapist monitoring). Other patients may be “over- triaged” to ICUs when their outcomes and costs of care would be as good or better by sending them directly to wards (e.g., due to less delirium or immobility). Furthermore, determining which of these “discretionary” ICU admissions could be treated as well or better on wards would improve timely access to ICUs for patients with more advanced forms of cardiorespiratory collapse and more obvious indications for ICU admission. To improve outcomes for patients with ARF and/or sepsis by optimizing their triage, this application proposes a mixed-methods study of patients admitted through the emergency departments of 26 diverse hospitals in the Kaiser Permanente Northern California health system and the University of Pennsylvania Health System. Our primary goals are to determine which patients with ARF and/or sepsis benefit from ICU admission, and which emergency department, ward, and ICU characteristics and processes of care contribute to such net ICU benefits. Several methodological innovations will enable us to achieve these goals, and to surmount key limitations to prior studies that have sought to determine which acutely ill patients benefit from ICU admission. First, whereas all prior studies used approaches to causal inference that were susceptible to confounding by unmeasured differences among hospitals, we will obtain unbiased estimates of which patients benefit from ICU or ward admission by modeling a within-hospital variable – ICU capacity at the time of triage – as an instrumental variable. Second, although prior studies have been unable to determine exactly which patients the results apply to, the granular electronic health record data available across our study hospitals will enable our results to apply directly to well-characterized individual patients, promoting personalized triage decisions. Third, our use of ethnography and semi-structured interviews in hospitals that obtain large and small net ICU benefits will identify mechanisms by which different hospitals achieve improved outcomes in ICUs and wards. Completing the aims of this study will improve public health by identifying ways in which emergency departments, ICUs, and wards can improve outcomes for the more than 4 million Americans hospitalized each year with ARF and/or sepsis. Such results will also provide the evidence needed to develop and...