Study objectives: To describe the variation in frequency of do-not-resuscitate (DNR) orders in 42 US ICUs and to examine the relationship between published guidelines and qualitative observations about terminal care in 9 ICUs.
Design: Prospective inception cohort.
Setting: Forty-two ICUs in 40 US hospitals with more than 200 beds: 26 randomly selected and 14 large tertiary-care volunteers.
Participants: A consecutive sample of 17,440 ICU admissions during 1988 to 1990.
Measurements and results: We used age, race, comorbid conditions, disease, functional status, and acute physiology score on ICU day 1 to predict the likelihood of a DNR order for each patient. A cross-validated model was then used to predict variations in the risk of an ICU DNR order from 0 to 45% (area under receiver operating characteristic curve = 0.9). The model was then used to compare aggregate observed with predicted frequency of ICU DNR orders. Finally, we compared observations of DNR practices by a team of clinical and organizational researchers at 9 of the 42 ICUs with published guidelines and risk-adjusted DNR frequency: 1,577 admissions (9%) had DNR orders written in the ICU (range, 1.5 to 22%). The ICU site was a significant (p < 0.0001) predictor of variance in the patient level model. DNR orders were written significantly (p < 0.05) less frequently than predicted in 5 and more frequently than predicted in 3 of 42 ICUs. Nonwhite patients had significantly (p = 0.0001) fewer DNR orders after adjustment. The research team's implicit judgments following on-site analysis failed to distinguish ICUs with more or less DNR orders than predicted. Site-visited ICUs exhibited practices to emulate and practice to avoid.
Conclusions: The frequency of ICU DNR orders can be predicted based on individual risk factors for groups of ICU patients. After adjusting for differences in patient characteristics, there is significant variation in the frequency of DNR orders in a national sample of ICUs. These variations may be due to unmeasured differences in patient characteristics such as treatment preferences, religious affiliation, educational level, or physician practices. We found no relationship between risk-adjusted DNR order frequency and adherence to published guidelines.