Triage Judgments in the Emergency Department

Efforts to use technology to aid human decisions have typically tried to emulate average human decisions (i.e., "crowdsourcing"), but to improve outcomes, algorithms need to emulate good human decisions. We adopt an approach combining quasi-experimental variation and machine learning to study this question in the setting of emergency department triage.

We use rich data from the Veterans Health Administration (VHA) on approximately 11 million emergency visits in 130 VHA health care systems, including triage nurse identities and patient bed assignments, ESI scores, vital signs, comorbidities, and disposition and health outcomes. We show that patients are as good as randomly assigned to triage nurses based on the time of arrival to the emergency department, and we show that triage nurses have a causal effect on patient mortality. Among patients with predicted mortality greater than 0, one standard-deviation increase in triage nurse value added results in a 20% increase in patient mortality.

We next characterize the mechanisms by which triages may have a mortality effect, along two key dimensions of triage behavior: First, triage nurses control the amount of time patients spend waiting to receive care. Second, triage nurses communicate the severity of patients' conditions by assigning an Emergency Severity Index (ESI) level to each patient. We find significant discretion and variation in both triage activities across triage nurses. We use machine learning and empirical Bayes methods to project high-dimensional variables describing these behaviors onto smaller-dimensional space. We find that these mechanisms together explain up to 80% of triage nurse effects on mortality. These results suggest that, relative to crowdsourcing, algorithms that link behaviors with outcomes could potentially capture a large amount of the effect of triage on outcomes.