van Walraven C, Forster AJ, Parish DC, Dane FC, Chandra KMD, Durham MD, Whaley C, Stiell I. Validation of a Clinical Decision Aid to Discontinue In-Hospital Cardiac Arrest Resuscitations. JAMA. 2001;285(12):1602-1606. doi:10.1001/jama.285.12.1602
Author Affiliations: Department of Medicine, University of Ottawa, Ottawa, Ontario (Drs van Walraven, Forster, and Stiell); Department of Internal Medicine, Mercer University School of Medicine, Macon, Ga (Drs Parish, Dane, Chandra, Mr Durham, and Ms Whaley); Clinical Epidemiology Unit, Loeb Health Research Institute, Ottawa Hospital, Ottawa, Ontario (Drs van Walraven and Stiell); and Institute for Clinical Evaluative Sciences, Toronto, Ontario (Dr van Walraven).
Context Most patients undergoing in-hospital cardiac resuscitation do not survive
to hospital discharge. In a previous study, we developed a clinical decision
aid for identifying all patients undergoing resuscitation who survived to
Objective To validate our previously derived clinical decision aid.
Design, Setting, and Participants Data from a large registry of in-hospital resuscitations at a community
teaching hospital in Georgia were analyzed to determine whether patients would
be predicted to survive to hospital discharge (ie, whether their arrest was
witnessed or their initial cardiac rhythm was either
ventricular tachycardia or ventricular fibrillation or
they regained a pulse during the first 10 minutes of chest compressions).
Data from 2181 in-hospital cardiac resuscitation attempts in 1987-1996 involving
1884 pulseless patients were analyzed.
Main Outcome Measure Comparison of predictions based on the decision aid with whether patients
were actually discharged alive from the hospital.
Results For 327 resuscitations (15.0%), the patient survived to hospital discharge.
For 324 of these resuscitations, the patients were predicted to survive to
hospital discharge (sensitivity = 99.1%, 95% confidence interval, 97.1%-99.8%).
In 269 resuscitations, patients did not satisfy the decision aid and were
predicted to have no chance of being discharged from the hospital. Only 3
of these patients (1.1%) were discharged from the hospital (negative predictive
value = 98.9%), none of whom were able to live independently following discharge
from the hospital.
Conclusion This decision aid can be used to help physicians identify patients who
are extremely unlikely to benefit from continued resuscitative efforts.
Cardiopulmonary resuscitation and advanced cardiac life support interventions
are used to resuscitate patients having cardiac or respiratory arrest. Physicians
and other health care workers follow algorithmic advanced cardiac life support
guidelines from the American Heart Association1
in an attempt to restore a spontaneous circulation in these patients. Despite
these guidelines, decision making during resuscitation is difficult because
cardiac arrest is a complex common pathway of a diverse collection of diseases.2
The decision to stop resuscitative efforts can be particularly difficult
for physicians. In these situations, decision making must balance a "respect
for human dignity" and clinical judgment.3
Prognostic factors that portend a poor outcome following arrest may help the
latter. Death following resuscitation has been associated with both prearrest
factors (including hypotension,4 renal failure,5,6 pneumonia,4
low functional status,4,7 and
metastatic cancer4,6) and intra-arrest
factors (including prolonged duration of resuscitation4,7,8
and initial cardiac rhythms other than pulseless ventricular tachycardia or
ventricular fibrillation7,9- 11).
However, decision aids that use these and other factors to determine a patient's
prognosis following resuscitation are complicated and difficult to use at
the bedside. They have operated poorly when validated in distinct populations12 and have been unable to definitively identify patients
with no hope of survival.13 A simple decision
aid that reliably identifies patients who—regardless of the cause of
their arrest—have a poor outcome would be helpful. This could help avoid
the "tendency to try prolonged, excessive resuscitative efforts."3
For this reason, we previously derived a simple clinical decision aid
to identify all patients who were eventually discharged from hospital after
their arrest.14 Our goal, when deriving this
decision aid, was to maximize sensitivity for identifying these patients.
We studied 1077 adults undergoing in-hospital resuscitation who participated
in 2 randomized clinical trials15,16
involving 5 teaching hospitals. Using recursive partitioning, we found that
all resuscitated patients who were eventually discharged from the hospital
had a witnessed arrest, an initial cardiac rhythm of either ventricular tachycardia
or ventricular fibrillation, or a pulse within the first 10 minutes of chest
compressions. We proposed that physicians might safely withdraw resuscitative
efforts on patients who did not satisfy the decision aid since none of these
patients were discharged from the hospital.
In that report we stressed 2 issues. First, we emphasized that while
100% of the patients not satisfying the decision aid eventually died in hospital,
the 95% confidence interval (CI) of this point estimate extended down to 97.1%.
Second, we cautioned that validation of the decision aid in a separate patient
population was required before the decision aid should be used clinically.14 This was the objective of our current study.
This validation study was a secondary analysis of a resuscitation registry
at the Medical Center of Central Georgia (MCCG) in Macon. The MCCG is a 550-bed
community teaching hospital affiliated with the Mercer University School of
Medicine. It is the major hospital for the metropolitan area as well as the
tertiary center for the surrounding rural areas and has approximately 25 000
admissions per year.
Multidisciplinary teams consisting of nurses, residents, staff physicians,
respiratory therapists, and pharmacists conducted the resuscitations. Staff
physicians or senior medical residents directed the resuscitations using standard
protocols published by the American Heart Association.3
Team members were all trained in basic life support and most were trained
in advanced cardiac life support. All members underwent regular updates. During
the study period, resuscitation was attempted for approximately 1.5% of admissions.
With the exception of arrests occurring in the neonatal intensive care
unit or operating room, all in-hospital arrests occurring between 1987 and
1996 were entered into the registry. The resuscitation registry was reviewed
and approved by the Institutional Review Board of Mercer University School
of Medicine and MCCG. Resuscitation was defined, as has been suggested by
the Utstein Conference,17 as "any effort to
reverse a clinical death in progress." Code sheets were completed after each
resuscitation attempt and were used to identify events. To ensure complete
capture of resuscitations, the records of patients without a code sheet but
whose hospital charges included cardioversion, defibrillation, or the administration
of epinephrine were also reviewed to determine if a resuscitation attempt
occurred. Data were entered into the registry by MCCG project staff and were
checked prior to analysis. All data were cross-checked for accuracy through
an extensive review of hospital records and death logs. Clinical categories,
such as whether or not the arrest was witnessed and initial cardiac rhythm,
used standardized definitions that had interrater κ values exceeding
0.9 before final classification of the data set was complete. Questionable
cases were reviewed by 2 researchers (D.C.P., K.M.D.C.) for consensus definition
in the final review. This was done before this study was conceived. A detailed
description of the registry was previously published.11
All resuscitations in the registry were eligible for this study. Resuscitations
were excluded if they were performed on patients younger than 16 years old
or were performed on patients in the operating room at the time of the arrest.
Our decision aid applies to patients who were pulseless at the start of the
resuscitation. Therefore, resuscitations were excluded if the initial rhythm
was any other than pulseless ventricular tachycardia, ventricular fibrillation,
pulseless electrical activity, or asystole. Resuscitations were also excluded
if patients received no chest compressions, if information required by the
decision aid was missing, or if time to initial chest compressions exceeded
15 minutes. Each of these exclusion criteria was used for patient selection
in the decision aid derivation.14
To apply the decision aid, we determined the time from the start of
chest compressions to the end of the resuscitation. This was the return of
any spontaneous circulation lasting 2 or more minutes or the end of resuscitative
efforts. The former criterion addressed a concern of our decision aid,2 namely, that patients who regained a pulse for most
of the resuscitation only to lose it before the 10-minute mark might have
resuscitation stopped if the aid was applied. Patients who were directly seen
to lose spontaneous circulation were classified as "witnessed." Also, all
patients who had an arrest while on a cardiac monitor, in the intensive or
coronary care unit or in the cardiac catheterization laboratory, were classified
as "witnessed arrests." This was regardless of whether or not the patient
was directly visualized to lose spontaneous circulation. Patients in the emergency
department were included in the study only if they actually had an arrest
after they arrived in the department. Finally, we applied the decision aid
in this study in a more clinically intuitive order than that presented in
the derivation study. This altered order does not affect the model's statistical
significance or operating characteristics.
A 2 × 2 contingency table comparing actual discharge status to
that predicted by the decision aid was used to calculate the classification
performance of the clinical decision aid with 95% CIs. We reviewed the medical
record of all patients who did not satisfy the decision aid but survived more
than 24 hours to determine their course in hospital. Analyses are by resuscitation.
When this study was conducted, the registry recorded 3960 resuscitations.
We excluded 1779 events for the following reasons: no chest compressions were
given during the resuscitation (n = 1147), the patient was not pulseless at
the start of the resuscitation (n = 254), the patient was younger than 16
years (n = 85), the patient was in the operating room when the arrest was
called (n = 1), the time from arrest being called to the first chest compression
exceeded 15 minutes (n = 9), or information required to apply the decision
aid was missing (n = 283).
This left 2181 attempted resuscitations comprising 1884 patients. These
patients had a mean age of 65 years (95% CI, 64.3-65.7) and 47.3% were female.
Most arrests occurred off ward, had a cardiorespiratory cause, and were witnessed
(Table 1). The predominant initial
rhythms were asystole and pulseless electrical activity, and for 2094 (96.0%)
resuscitations, chest compressions were delivered within 5 minutes of arrest.
A spontaneous circulation was attained in almost half of resuscitations, and
for 15.0% of resuscitations, patients survived to discharge from hospital.
Figure 1 illustrates how the
decision aid identified patients who would be discharged from the hospital.
Of the 327 resuscitations for which patients were discharged from hospital,
287 satisfied the first component of the decision aid ("arrest was witnessed").
Of the remaining 40 resuscitations, 10 satisfied the second component of the
aid ("initial rhythm was ventricular tachycardia or ventricular fibrillation").
Twenty-seven of the remaining 30 resuscitations satisfied the final component
of the decision aid ("pulse regained during the first 10 minutes of chest
Table 2 presents the classification
performance of the decision aid. Of the 327 resuscitations for which patients
were discharged from hospital, all but 3 satisfied the decision aid, resulting
in a sensitivity of 99.1% (95% CI, 97.1%-99.8%). That is, the decision aid
correctly identified all but 0.9% of those who were discharged from the hospital.
The decision aid had a negative predictive value of 98.9% (95% CI, 96.5%-99.7%).
That is, 1.1% of arrests that the decision aid predicted had no chance of
survival were actually discharged from the hospital.
Likelihood ratios allow clinicians to measure the quantitative importance
of test results.18 The negative likelihood
ratio of the decision aid was 0.064. To put this into perspective, assume
that we could accurately determine the probability that hospitalized patients
will survive to discharge if they required resuscitation. Also, assume that
we have 3 patients whose probability of surviving to hospital discharge, in
the event of an arrest, is 30%, 15%, and 5%. If these patients did not satisfy
the decision aid during their resuscitations, their probabilities of surviving
to discharge would decrease to 2.7%, 1.1%, and 0.3%, respectively.
We determined the outcome of the 3 people whom the aid predicted had
no chance of being discharged from the hospital. The first patient was a 76-year-old
man with dementia, hypertension, and chronic obstructive pulmonary disease
who was transferred to another hospital following resuscitation to continue
inpatient medical therapy. When he was discharged he was in a very poor condition
and required tracheostomy, gastrostomy, foley catheter, and rectal tube. These,
however, were at least partially required for an obstructive oropharyngeal
carcinoma as opposed to ischemic cerebral damage. He died 2 months following
discharge from the hospital. The second patient was a 43-year-old man with
chronic obstructive pulmonary disease and alcoholic cardiomyopathy. Although
he had minimal ischemic damage from the arrest, he was discharged to a nursing
home residence because of problems caring for himself. The final patient was
a 65-year-old previously well woman who had an arrest following back surgery.
She had no ischemic injury but required nursing home placement because of
complications of her back surgery.
Of the 269 resuscitations in which patients were predicted to have no
chance of surviving to hospital discharge, the mean resuscitation duration
was 22.6 minutes (SD, 11.1 minutes; range, 10-72 minutes). In 53 of these
resuscitations (19.7%), patients achieved a spontaneous circulation and were
transferred to the intensive care unit. Twenty-six of these patients remained
alive for at least 24 hours but died later during the hospitalization. These
26 patients survived a mean of 8.5 days following the resuscitation (range,
1-29 days; total, 213 days). Of the 20 patients whose chart was available,
15 (75%) never regained consciousness. For 9 (45%) of these 20 patients, a
decision was made to withdraw active care.
Using one of the world's largest continuous registries of hospital resuscitations,
we found that a simple clinical decision aid performed well to identify patients
with any chance of discharge from hospital following resuscitation. All but
3 patients (1.1%) who did not meet criteria for the decision aid died following
their arrest. We believe that this decision aid can be used with other clinical
factors to help physicians identify patients who are extremely unlikely to
benefit from ongoing resuscitation efforts.
Our aid meets the most important methodological standards for decision
aids.19- 21 The
outcome predicted by the aid is important and objective. The aid is clinically
sensible since each component (ie, witnessed arrest,8
initial ventricular tachycardia or ventricular fibrillation,7
and duration of resuscitative efforts4,7,8,22)
has been associated with survival in other studies. The potential effects
of using the decision aid—in terms of avoided intensive care unit days—have
been estimated in both the derivation14 as
well as this study. Most important, the aid has been validated on a distinct
population, a standard met by fewer than half of all decision aids published
between 1991 and 1994 in 4 major medical journals.19
We hope that this aid will be tested further in other patient populations,
preferably in a prospective fashion, to further ensure its validity.
Two factors regarding the decision aid require more detailed discussion.
First, this study showed that the decision aid is very robust since it was
valid in a separate patient group despite large differences between the derivation
and validation populations. Compared with the derivation group,14
patients in this study were significantly younger (mean age, 65 vs 68 years, P<.001), and were less likely to have have had an arrest
on the ward (43.9% vs 55.2%). Outcomes for the validation group were much
better with a significantly greater number of patients surviving to 1 hour
(48.8% vs 33%, P = .0001) and discharge (15.0% vs
9.6%, P<.001). Most important, the patients used
for the derivation and validation studies were from very different health
care systems. The observation that our decision aid performed so well in such
a different patient population should give physicians confidence to apply
it to their own patients.
Second, although each component of the aid is very objective, it must
be used with care since difficulties in measuring time, classifying cardiac
rhythms, and determining the presence of a pulse during the resuscitation
have been well documented in the medical literature. All patients who had
an arrest in the intensive care unit or while on a monitor are considered
witnessed arrests when applying this aid, even if they are not actually visualized
to become unstable. Similarly, physicians must be confident about the classification
of the initial rhythm, and special care must be exercised to ensure that perfusing
rhythms with hypotension are not classified as pulseless electrical activity.
Also, physicians must ensure that resuscitative efforts have truly proceeded
for a complete 10 minutes, without the return of a pulse that persisted for
2 or more minutes, before the last component of the aid is determined. Finally,
this decision aid cannot be applied to out-of-hospital resuscitations without
Although the resuscitations in this study spanned over 9 years, we do
not believe that changes in resuscitation significantly affected our results.
Between 1987 and 1996, the time period of the study, no pharmacological or
mechanical intervention was introduced that reliably improved the patient
outcomes following resuscitation. Although monitoring technology, such as
telemetry, might have become more prevalent during the study, the decision
aid would account for this since all such patients would be classified as
having a witnessed arrest.
This study addresses several criticisms of our decision aid.2 There was concern that patients who regained a pulse
within the first 10 minutes of the resuscitation, only to lose it again at
10 minutes, would not satisfy the decision aid and would have resuscitative
efforts withdrawn. Since the registry used for this study recorded when patients
had any return of spontaneous rhythm, we ensured that patients who regained
a pulse for longer than 2 minutes within the first 10 minutes of resuscitation
satisfied the decision aid. Second, the decision aid only identified 12% of
the study group as patients with no chance of survival. However, this is comparable
to other decision aids used to identify patients with poor outcomes postarrest
such as the Pre-Arrest Morbidity score,12,23,24
the Prognosis After Resuscitation score,12,23
and the APACHE (Acute Physiology and Chronic Health Evaluation) III index,12 each of which applied to between 5% and 20% of all
resuscitations. Third, we were criticized for both ignoring prearrest patient
factors that are associated with outcomes and attempting to derive a simple
decision aid to be applied to such a diverse population of patients having
a broad range of survival probabilities. As we demonstrated in this article,
physicians could combine the negative likelihood ratio of the decision aid
with estimated prearrest survival probabilities to improve prognostication.18 However, we believe that freeing physicians from
having to calculate a pretest probability of survival based on various factors
is a strength of the aid. Since many patients with a good prognosis have some
return of spontaneous circulation within the first 10 minutes of the resuscitation,
our aid innately accounts for these patients. Finally, we agree that any time
patients are interactive with their environment after resuscitation—even
if they end up dying later—is valuable. However, we do not feel that
this advocates adopting a "never give up" mind-set during resuscitative efforts.
Our data show that for each person who did not satisfy the decision aid but
regained consciousness, approximately 15 people did not. In many of these
cases, family members are put in the unenviable position of having to decide
whether to withdraw care from a loved one. Given the stress that this causes,25,26 it is arguably a high price to pay.
We believe that the decision aid validated in this study can be used
by physicians to identify patients who are extremely unlikely to benefit from
ongoing resuscitation efforts. Further validation of the aid in multiple sites
with prospectively collected data would be welcome. In addition, this decision
aid might also help patients be more directive regarding their resuscitation
attempts. Our data could allow patients and physicians to precisely quantify
when resuscitative efforts would be stopped, such as when our decision aid
is not satisfied. Since patients are often afraid of being resuscitated only
to remain on life support, our decision aid could be used to help patient
decision making that might avoid such a situation.