To determine how well triage physicians judge the probability of death or severe complications that require treatment only available in an ICU to maintain life for patients with acute congestive heart failure (CHF). Prospective cohort study. An urban university hospital, a Veteran's Administration hospital, and a community hospital. Patients were those visiting the emergency department (ED) with acute CHF, excluding those who already required a treatment only available in an ICU to maintain life, and those with possible or definite myocardial infarction. Physician participants were those caring for the patients in the ED. We performed chart reviews to ascertain whether each patient died or had severe complications develop by 4 days. We collected judgments of the probability of this outcome from the physicians taking care of the study patients in the ED. The prevalence of death or severe complications was 43 per 1,032 patients (4.2%). The mean +/- SD of physicians' judgments of the probability of this outcome was 32.1 +/- 28.4%. A calibration curve that stratified these judgments by decile demonstrated that physicians consistently overestimated this probability (p < 0.01). Physicians' judgments were only moderately good at discriminating which patients would have the outcome (receiver operating characteristic curve area, 0.715). Patients admitted to an ICU received the highest average predicted probability (56.4%), followed by those admitted to a telemetry unit (34.1%), to a regular hospital ward (29.8%), and those sent home (17.9%.) Physicians drastically overestimated the probability of a severe complication that would require critical care for patients with acute CHF who were candidates for ICU admission. Their judgments of this probability were associated with their triage decisions, as they should be according to several guidelines for ICU triage. Overestimation of the probability of severe complications may have lead to overutilization of scarce critical care resources. Current critical care triage guidelines should be revised to take this difficulty into account, and better predictive models for patients potentially requiring critical care should be developed.