Factors Associated with Do-Not-Resuscitate Orders: Patients' Preferences, Prognoses, and Physicians' Judgments

Methods

Phase I of SUPPORT (1989 to 1991) evaluated the decision-making process and outcomes for 4301 severely ill hospitalized patients. Phase II (1992 to 1994) was a controlled clinical trial of an intervention that gave physicians information on patient prognoses and preferences for end-of-life care. Nurse clinicians were trained to assist with and facilitate communication to the 2652 patients who received the intervention. Another 2152 persons served as controls and received usual medical care. Phase II enrolled a total of 4804 patients [18]. A major hypothesis of this trial was that accurate information and better communication would increase the frequency of DNR orders and decrease the time taken to write an order. The phase II intervention neither increased the frequency nor accelerated the timing of DNR orders [18]. Because no secular trends in the timing of DNR orders were seen during the entire study period (1989 to 1994), we could combine data from phase I and phase II for this analysis.

Inclusion Criteria

The SUPPORT trial enrolled patients who met entry criteria at five medical centers: Beth Israel Hospital, Boston, Massachusetts; MetroHealth Medical Center, Cleveland, Ohio; Duke University Medical Center, Durham, North Carolina; Marshfield Clinic-St. Joseph's Hospital, Marshfield, Wisconsin; and the University of California Medical Center, Los Angeles, California. To be eligible, patients had to have an anticipated aggregate 6-month mortality rate of approximately 50% and had to have one or more of the following diagnoses: acute respiratory failure, multiple organ-system failure with sepsis or malignant condition treated in an intensive care unit, coma, chronic obstructive lung disease, congestive heart failure, cirrhosis, metastatic colon cancer, or non-small-cell lung cancer (stage III or IV). The entry requirements of the study have been described elsewhere [17].

The study sample included all patients enrolled in SUPPORT (or their surrogates) who answered the question about patient preferences for resuscitation in the first interview. We excluded the following patients: those for whom a DNR order had been written before study entry, those who did not survive or were discharged during the first 48 hours of the study; those admitted to the hospital with a scheduled discharge within 72 hours; those younger than 18 years of age; those with the acquired immunodeficiency syndrome (AIDS); those who did not speak English; those who were admitted to the psychiatric service; those who were pregnant; and those who had sustained an acute burn, head trauma, or other trauma (unless they later developed acute respiratory or multiple organ-system failure). We selected disease categories on the basis of their prevalence among dying patients in the acute care hospital. One of the goals of the SUPPORT study was to develop prognostic models; thus, at the time of the conceptualization and pilot testing of the study (1986 to 1988), we decided that the treatment and prognoses of patients with AIDS were changing too rapidly to allow us to include such patients. We remained in direct contact with surviving patients for 6 months, and we used the National Death Index to follow patients thereafter.

Data Collection

We interviewed patients and their surrogates (a surrogate was defined as the person who would make decisions if the patient was unable to do so) on the third day after study enrollment (95% of interviews took place between the second and seventh study days). Patients and surrogates were interviewed again during the second study week and at 2 and 6 months. For our analysis, we used only the data obtained from the first interview, during which patients were asked about their socioeconomic status, functional status before hospitalization, self-assessed quality of life, and preferences for cardiopulmonary resuscitation [19]. The question on preferences for cardiopulmonary resuscitation was worded “As you probably know, there are a number of things doctors can do to try revive someone whose heart has stopped beating, which usually includes a machine to help breathing. Thinking of your current condition, what would you want your doctors to do if your heart stops beating? Would you want your doctors to revive you, or would you want your doctors not to try to revive you?” Responses were coded as wanted resuscitation, wanted resuscitation but no ventilator, wanted no resuscitation, or did not know. In all analyses, a surrogate interview was substituted when no patient interview was possible. Functional status was measured by using a slightly modified version of the Katz Activities of Daily Living scale [20, 21]. The scale proposed by Katz measured impairment in bathing, dressing, eating, continence, transferring, and toileting; we added a question about impairment in walking. We summed the number of dependencies, which could range from 0 to 7 (the latter indicated dependency in all seven functions).

For study days 1, 3, 7, 14, and 25, patients' medical records were abstracted concurrently for physiologic variables that were known to be predictors in the Acute Physiology and Chronic Health Evaluation (APACHE) II and APACHE III [22, 23] prognostic systems (as used in the SUPPORT prognostic model [17]) that predicted survival probability for as long as 180 days after enrollment. We report the acute physiology scores of APACHE III; these scores range from 0 to 299 (a higher number indicates greater acuity of medical illness). In this analysis, we used only the 2-month SUPPORT survival estimate, which was based on data collected on the first study day [17]. After discharge or death, medical records were abstracted for discussions and treatment decisions about cardiopulmonary resuscitation.

We considered the attending physician to be the physician of record at the end of the second study day (patients had to survive 48 hours to qualify for the study). On the basis of an interview during which demographic information was collected, we grouped physician specialties into the following categories: oncology, pulmonary or intensive care medicine, cardiology, surgery, and general medicine (which included other medical subspecialties, such as rheumatology and gastroenterology). Each study institution had policies stating that a DNR order had to be discussed and that the attending physician had to sign or co-sign the order.

Statistical Analysis

We did univariate analyses to determine the unadjusted incidence of DNR orders by patient and institutional characteristics. We did a multivariable analysis using a log-normal regression model that contained the following predictor variables: patient preferences about cardiopulmonary resuscitation; age, sociodemographic factors, and diagnosis; scores on the Katz Activities of Daily Living scale [20, 21] and a 5-point scale that rated the patient's quality of life from excellent to poor; comorbid conditions; the Glasgow coma score [24]; the probability of surviving for 2 months based on the SUPPORT model; and the hospital and the attending physicians' specialty and the relation between the two.

The number of days between study enrollment and a DNR order was the dependent variable. To avoid taking the logarithm of zero, we added one half-day to the time the DNR order was written. We used cubic spline functions for all continuous variables to avoid assumptions of linearity between these variables and the timing of DNR orders [17]. The log-normal model was chosen from among the log-logistic, log-normal, and Cox models (including the Weibull model). For some of the most important predictors (such as patient prognosis), we stratified data on the predictors and computed Kaplan-Meier estimates of the probability of having no DNR order. We then transformed the y-axis (using the inverse distribution) and the x-axis (by taking the logs). The log-normal model provided the most linear and parallel associations; thus, the fit of the log-normal model was deemed adequate for the distribution of data.

We derived adjusted ratios of the predicted median number of days (time ratios) to documentation of a DNR order from model coefficients. A time ratio less than one indicates less time to a DNR order; a ratio greater than one indicates more time. Patients without a DNR order were censored at the time of hospital discharge or death. We used boot-strapping to obtain variance estimates, which were the basis for CIs and hypothesis tests [25].

For a typical SUPPORT patient, the probability of having a DNR order by day 30 was plotted as a function of the patient's age and probability of surviving for 2 months; because we used cubic spline functions, both variables could vary [17]. To examine the effect of using only the less severely ill interviewed patients [19], we did a regression analysis of the entire SUPPORT cohort that did not include patient preferences. To examine the effects of substituting the surrogates' reports of patients' preferences when the latter were unavailable, we also did an analysis that included only the patients' preferences.

Results

In phases I and II of SUPPORT, 9105 participants were enrolled. A total of 8836 patients remained after we had excluded 30 patients who lacked DNR documentation and 239 patients whose DNR orders had been written before study enrollment. Of these remaining patients, 3286 (37%) were interviewed and answered the question about their preferences for cardiopulmonary resuscitation. We also substituted 3516 surrogate responses when patients could not be interviewed. The final sample consisted of 6802 patients (Figure 1).

Patient Characteristics

The median patient age was 65 years; the average APACHE III score was 49; and the median probability of surviving 2 months (estimated by the SUPPORT prognostic model) was 69%. The average number of functional impairments present before hospital admission (based on the Activities of Daily Living scale) was 1.4. The hospital mortality rate was 22%, and the 6-month mortality rate was 39%. Most patients were white (88%) and male (63%); 18% of the patients were black. The distribution of incomes was wide (57% of patients earned less than $11 000 per year, and 29% earned $25 000 or more). Acute respiratory failure or multiple organ-system failure was the most common diagnostic category, and intensive care medicine-pulmonary medicine and internal medicine were the most common clinical specialties (Table 1).

Frequency, Timing, and Discussions of Do-Not-Resuscitate Orders

Do-not-resuscitate orders were written for 2138 (31%) of all 6802 patients and for 1238 (81%) of the 1523 patients who died during the index admission. For patients with a DNR order, the median time between study enrollment and a DNR order was 5 days (interquartile range, 2 to 13 days). For patients who died in the hospital, the median time between DNR orders and death was 3 days (interquartile range, 1 to 7 days).

Discussions about resuscitation were documented in the records of 2644 patients (39%); 1945 (74%) of these had a DNR order written. For the remaining 699 (10% of the total sample) patients, a discussion was documented in the medical record but no DNR order was written. For 260 of these 699 patients (37%), the medical record contained no information about a DNR order or decision; for 252 (36%), the record contained documentation that cardiopulmonary resuscitation was to be fully used; for 106 (15%), documentation showed that cardiopulmonary resuscitation was to be partially used; for 71 (10%), there was written documentation to forego cardiopulmonary resuscitation but no written order. Twelve of the 699 patients (2%) died with no resuscitation attempted and with no order or decision documented in the chart.

Patient Preferences for Cardiopulmonary Resuscitation

Of the patients (or surrogates) who responded to the interview question about resuscitation, 3723 (55%) preferred to have resuscitation in the event of cardiopulmonary arrest and 1937 (29%) preferred to forego cardiopulmonary resuscitation (Table 1). Of patients who preferred to forego cardiopulmonary resuscitation, 1010 (52%) subsequently had a DNR order written. Ninety (28%) of the 325 patients who wanted cardiopulmonary resuscitation but did not want to be placed on a ventilator had DNR orders, as did 270 (33%) of the 817 patients who had no preference at the time of the interview. Of patients for whom a DNR order was written and who voiced a preference for cardiopulmonary resuscitation, 93% had documentation in the medical record stating that the DNR order had been discussed with the patient (29%) or the patient's surrogate (64%). Do-not-resuscitate orders were written sooner for patients who preferred to forego cardiopulmonary resuscitation (adjusted time ratio, 0.27). The timing of the DNR order was intermediate for those who wanted resuscitation but not ventilation (adjusted time ratio, 0.68) and for those who did not know their preferences (adjusted time ratio, 0.67). In the multivariable regression model, the variable representing patient preferences was the most important predictor of the timing of DNR orders (P < 0.001) (Figure 2).

Patient Age

The frequency of DNR orders increased with age, from 22% for patients younger than 55 years of age to 56% for patients 85 years of age or older (Table 1). After adjustment for all factors listed in Table 1, patients 85 years of age or older had the shortest time to a DNR order (adjusted time ratio, 0.32). Orders were written more rapidly for patients 75 to 84 years of age (adjusted time ratio, 0.57) and for patients 55 to 74 years of age (adjusted time ratio, 0.75) (Table 1). This resulted in a nonlinear association between age and the timing of DNR orders: Patients younger than 75 years of age had less than a 15% estimated probability of having a DNR order by day 30 after study entry; that probability doubled for patients older than 84 years of age (Figure 3, top).

Prognosis

The unadjusted frequency of DNR orders increased as the probability of surviving for 2 months decreased. For example, without adjustment, we found that DNR orders existed for 14% of patients whose probabilities of surviving for 2 months were greater than 75% and for 67% of patients whose probabilities of surviving for 2 months were less than 25% (Table 1). After adjustment, however, we found a nonlinear association between the timing of the DNR order and objective prognosis (Table 1). The bottom panel of Figure 3 shows that the overall probability that a typical SUPPORT patient would have a DNR order within 30 days of study entry was less than 30%, regardless of the patient's prognosis. For patients with a survival probability greater than 50%, the likelihood of having a DNR order decreased to almost zero; at survival probabilities less than 50%, the association was flat. Overall, the estimated probability of surviving for 2 months was the second most important predictor of the timing of DNR orders (P < 0.001) (Figure 2).

Diagnosis

The unadjusted frequency of DNR orders was highest for patients who were in a coma (64%) and lowest for patients with congestive heart failure (14%) (Table 1). After adjustment for the factors listed in Table 1, DNR orders were written most slowly for patients with acute respiratory failure or multiple organ-system failure than for patients with congestive heart failure (adjusted time ratio, 1.46) and most quickly for patients who were in a coma or had lung cancer than for patients with congestive heart failure (adjusted time ratios, 0.71 and 0.69, respectively). The timing of DNR orders for patients in the other diagnostic categories was similar to the timing for patients with congestive heart failure.

Quality-of-Life Assessment and Functional Status

Low patient- or surrogate-assessed quality of life and poor functional status before hospitalization were related to an increased frequency of DNR orders and, after adjustment for the factors listed in Table 1, with faster timing of DNR orders (P < 0.001) (Table 1 and Figure 2). For example, 3776 patients (56%) reported having fair quality of life; 1540 (41%) of these had a DNR order. Orders were written faster for these patients (adjusted time ratio, 0.67) than for patients who reported excellent to good quality of life (Table 1). The frequency of DNR orders increased and the timing of the orders decreased as the number of functional dependencies increased to two or more (Table 1). Socioeconomic factors, the presence of dementia, and the number of comorbid conditions were not associated with the timing of DNR orders (Figure 2). The number of days of hospitalization before study enrollment, the Glasgow coma score, and other components of the 2-month SUPPORT prognostic model were slightly associated with the timing of DNR orders (Figure 2).

Physician and Institutional Factors

Physician specialty was the third most important predictor of the timing of DNR orders (Figure 2). Patients whose attending physicians were cardiologists had the fewest DNR orders (22%); those whose attending physicians were pulmonologists or intensive care specialists had the most (36%) (Table 2). Surgeons took the longest time to write DNR orders (adjusted time ratio, 1.80) compared with medical specialists (Table 2 and Figure 4). Patients hospitalized at study site E were more than 50% more likely to have a DNR order than were patients at institution A (Table 2), and DNR orders were written more quickly at sites B and E (adjusted time ratios, 0.75 and 0.63, respectively). An interaction between hospital site and specialty groups was statistically significant (P < 0.001).

Sensitivity Testing

The associations between the timing of DNR orders and patient and institutional factors were similar in the regression model that was run for the entire SUPPORT cohort of 8836 patients and did not include patient preferences (Figure 1). This indicates that the relation between other predictors and the timing of DNR orders was consistent when preferences about cardiopulmonary resuscitation were not considered. Findings were also similar in the model that was run only on the subgroup (3286 patients) and that included patient preferences, indicating that surrogate substitution did not change the associations between predictor variables and the timing of DNR orders.

Discussion

Our study, in which we estimated the relative association between patient preferences and other important characteristics and the frequency and timing of DNR orders, is a relatively new undertaking for clinical medicine. It is, however, supported by work in other fields [26, 27]. These studies have emphasized how difficult it is to incorporate specific factors, such as patient age, into decision making in a manner consistent with the biological influences of the factors. It is also difficult for individual physicians to accurately estimate the relative weight or influence of these factors when several factors, such as diagnosis, laboratory results, and medical history, are considered simultaneously. Therefore, just as with other complex tasks, medical decision making can be distorted. For example, thresholds for important therapies vary substantially among individual practitioners [28].

Our results indicate that patients' preferences and short-term prognoses are strongly associated with the frequency and timing of DNR orders. This finding is reassuring because both preferences and prognoses are ethically and medically important. However, our findings also suggest that there is room for improvement. Compared with stated patient preferences, DNR orders were infrequent and late. In addition, physicians appear to have relied on simple interpretations of complex patient risk factors. Do-not-resuscitate orders were not written for almost half of the patients who wanted resuscitation withheld or for many patients whose probability of surviving for 2 months was less than 5%. When deciding to write a DNR order, physicians may have relied heavily on an age threshold of 75 years and a 2-month survival estimate of 50% or less. Finally, even after we controlled for patients' preferences for cardiopulmonary resuscitation and other important clinical characteristics, the frequency and timing of DNR orders varied widely across institutions and clinical specialty groups (Table 1 and Figure 4).

One of the major factors associated with the underuse of DNR orders in our study was an apparent lack of communication among physicians, patients, and families. We previously reported that physicians often misunderstand patient preferences to forego cardiopulmonary resuscitation and that this misunderstanding is associated with increased use of resources [19]. Most orders were written within 3 days of the patient's death; this suggests, as have findings in other reports, that clinicians wait until patients are close to death before making plans [9, 11, 29]. At that point, it is too late for much direct patient involvement. Discussions about resuscitation were documented in only 39% of the medical records. Another reason for underuse of DNR orders may have been that for patients who had a 2-month survival estimate of 50% or less at the time of study admission, the timing of DNR orders and prognosis were not related. For these patients, the overall estimated probability of having a DNR order within 30 days of study enrollment was less than 30%, even for patients whose 2-month survival probability was less than 25% (Figure 3, bottom). These findings suggest that physicians used a survival probability of approximately 50% as a threshold before considering DNR orders; they apparently did not differentiate among estimates below this level. Thus, patients with poor prognoses were exposed to the risk for an unsuccessful resuscitation attempt [15, 30-32].

Our finding that age was a strong independent predictor of the timing of DNR orders is consistent with findings in previous studies [10, 14, 15]. However, DNR orders were written most rapidly for persons older than a threshold of 75 years of age. Age has consistently been shown to be an independent and linear risk factor for short-term death among seriously ill hospitalized patients, but without the specific age thresholds shown in the top panel of Figure 3[17, 20, 21, 33, 34]. Other studies [35, 36] have also documented the minor role of age as a predictor of outcome from resuscitation.

Controversy has surrounded age-based rationing of health care, and proponents have argued that care for the elderly be rationed at a specific age [37-40]. Our results suggest that physicians in our study relied on age when deciding whether or not to use cardiopulmonary resuscitation; they used an age of 75 years or greater as an important threshold for limiting attempts at resuscitation. This finding is compatible with the rationale for the claim [38] that advanced age should, by itself, be a reason for less intense medical treatment.

Wachter and colleagues [16] showed a ninefold difference in the frequency of DNR orders between patients with AIDS and those with congestive heart failure. In a similar finding, we noted a fivefold difference in the frequency of DNR orders between patients who were in a coma and those who had congestive heart failure (Table 1). The patient's primary disease was the fourth most important predictor of the timing of DNR orders, suggesting that physicians were influenced by disease when considering the appropriateness of a DNR order. Given these findings, it is not surprising that the timing of DNR orders varied substantially among different clinical specialty groups and among the five participating hospitals [14, 15]. Together, the relative association of clinical specialty group and hospital was as important as the patient's probability of surviving for 2 months (Figure 2). The significant interaction between specialty group and hospital also indicated that similar clinical specialties also varied across medical institutions (Figure 4). These findings suggest that much of the variability in end-of-life decision making in our study was related to physician and institutional characteristics rather than to patient preferences or patients' abilities to benefit from resuscitation and other aggressive therapies. Such findings warrant future research on the origins of this variation from multiple perspectives, including the psychology of decision making and the organizational culture of the health care institution.

Limitations

The DNR decision-making process that we documented may not be representative of national patterns. The five participating hospitals were part of a clinical trial that attempted but failed to improve communication with and decision making for these ill patients [18]. This effort was intended, however, to increase the attention paid to the resuscitation decision-making process, and any potential bias should have been toward improvement. Furthermore, we studied patients in only nine disease categories. These categories, however, accounted for almost 40% of the deaths in a defined geographic population [41].

We relied on medical record review and interviews to collect information about the decision-making process. However, interviews may not accurately reflect patient participation in making decisions about cardiopulmonary resuscitation, which is part of a complex and dynamic process. Finally, although it would have been ideal to explicitly account for the day of death in the analysis that predicted the timing of a DNR order by a particular day, methods with which to do this are not currently available.

Conclusions

Do-not-resuscitate orders are strongly associated with patient preferences. In relation to the stated wishes of the patient, however, physicians may underestimate the desire for such orders. When orders are written, the patient's age and short-term prognosis are important predictors, but physicians appear to rely on these two variables in ways not consistent with the variables' influence on outcomes. Practice patterns vary substantially across specialties and institutions. The decision-making process leading to consideration and writing of DNR orders needs to be improved if it is to meet well-known standards of objectivity and patient involvement [1-428, 38-40].

Appendix: SUPPORT Investigators

George Washington University, Washington, D.C.: Rose Baker, MSHyg, Rosemarie Hakim, PhD, William A. Knaus, MD, Barbara Kreling, BA, Detra K. Robinson, MA, and Douglas P. Wagner, PhD; Dartmouth Medical School, Hanover, New Hampshire: Jennie Dulac, BSN, RN, Joanne Lynn, MD, MA, MS, Joan Teno, MD, MS, and Beth Virnig, PhD; Johns Hopkins University, Baltimore, Maryland: Marilyn Bergner, PhD (deceased), Albert W. Wu, MD, MPH, and Yutaka Yasui, PhD; Beth Israel Hospital, Boston, Massachusetts: Lee Goldman, MD, MPH, E. Francis Cook, ScD, Mary Beth Hamel, MD, Lynn Peterson, MD, Russell S. Phillips, MD, Joel Tsevat, MD, Lachlan Forrow, MD, Linda Lesky, MD, and Roger Davis, ScD; Cleveland MetroHealth Medical Center, Cleveland, Ohio: Alfred F. Connors Jr., MD, Neal V. Dawson, MD, Claudia Coulton, PhD, C. Seth Landefeld, MD, Theodore Speroff, PhD, and Stuart Youngner, MD; Duke University Medical Center, Durham, North Carolina: William J. Fulkerson Jr., MD, Robert M. Califf, MD, Anthony N. Galanos, MD, Peter Kussin, MD, Lawrence H. Muhlbaier, PhD, Maria Winchell, MS, Carlos Alzola, MS, and Frank E. Harrell Jr., PhD; Marshfield Medical Research Foundation, Marshfield, Wisconsin: Norman A. Desbiens, MD, Steven Broste, MS, Michael Kryda, MD, Douglas J. Reding, MD, Humberto J. Vidaillet Jr., MD, and Marilyn Follen, RN, MSN; University of California, Los Angeles, School of Medicine, Los Angeles, California: Robert K. Oye, MD, Paul E. Bellamy, MD, Gill Cryer, MD, James W. Davis, MD, Jonathan Hiatt, MD, Neil S. Wenger, MD, MPH, Honghu Liu, PhD, and Margaret Leal-Sotelo, MSW; Medical College of Wisconsin, Milwaukee, Wisconsin: Peter M. Layde, MD, MSc; Ohio University, Athens, Ohio: Hal Arkes, PhD; Presbyterian-St. Luke's Medical Center, Denver, Colorado: Donald J. Murphy, MD.

From George Washington University Medical Center, Washington, D.C.; Duke University Medical Center, Durham, North Carolina; University of California, Los Angeles, School of Medicine, Los Angeles, California; Beth Israel Hospital, Boston, Massachusetts; Medical College of Wisconsin, Milwaukee, Wisconsin; Marshfield Clinic, Marshfield, Wisconsin; and MetroHealth Medical Center, Cleveland, Ohio.

Drs. Harrell and Knaus: University of Virginia Health Sciences Center, Box 600, Charlottesville, VA 22908.

Dr. Wenger: Department of Medicine, University of California, Los Angeles, School of Medicine, B-564 Factor Building, Los Angeles, CA 90095-1706.

Dr. Phillips: Beth Israel Hospital, 330 Brookline Avenue, Boston, MA 02215.

Dr. Layde: Medical College of Wisconsin, DFCM, 8701 Watertown Plank Road, Milwaukee, WI 53226.

Dr. Califf: 2024 West Main Street, Bay A-108, Durham, NC 27707.

Dr. Connors: MetroHealth Medical Center, 2500 Metrohealth Drive, Cleveland, OH 44109-1998.