Customize your JAMA Network experience by selecting one or more topics from the list below.
Carson JL, Duff A, Berlin JA, et al. Perioperative Blood Transfusion and Postoperative Mortality. JAMA. 1998;279(3):199–205. doi:10.1001/jama.279.3.199
Context.— The risks of blood transfusion have been studied extensively but the
benefits and the hemoglobin concentration at which patients should receive
a transfusion have not.
Objective.— To determine the effect of perioperative transfusion on 30- and 90-day
Design.— Retrospective cohort study.
Setting.— A total of 20 US hospitals between 1983 and 1993.
Participants.— A total of 8787 consecutive hip fracture patients, aged 60 years or
older, who underwent surgical repair.
Main Outcome Measures.— Primary outcome was 30-day postoperative mortality; secondary outcome
was 90-day postoperative mortality. The "trigger" hemoglobin level was defined
as the lowest hemoglobin level prior to the first transfusion during the time
period or, for patients in the nontranfused group, as the lowest hemoglobin
level during the time period.
Results.— Overall 30-day mortality was 4.6% (n=402; 95% confidence interval [CI],
4.1%-5.0%); overall 90-day mortality was 9.0% (n=788; 95% CI, 8.4%-9.6%).
A total of 42% of patients (n=3699) received a postoperative transfusion.
Among patients with trigger hemoglobin levels between 80 and 100 g/L (8.0
and 10.0 g/dL), 55.6% received a transfusion, while 90.5% of patients with
hemoglobin levels less than 80 g/L (8.0 g/dL) received postoperative transfusions.
Postoperative transfusion did not influence 30- or 90-day mortality after
adjusting for trigger hemoglobin level, cardiovascular disease, and other
risk factors for death: for 30-day mortality, the adjusted odds ratio (OR)
was 0.96 (95% CI, 0.74-1.26); for 90-day mortality, the adjusted hazard ratio
was 1.08 (95% CI, 0.90-1.29). Similarly, 30-day mortality after surgery did
not differ between those who received a preoperative transfusion and those
who did not (adjusted OR, 1.23; 95% CI, 0.81-1.89).
Conclusions.— Perioperative transfusion in patients with hemoglobin levels 80 g/L
(8.0 g/dL) or higher did not appear to influence the risk of 30- or 90-day
mortality in this elderly population. At hemoglobin concentrations of less
than 80 g/L (8.0 g/dL), 90.5% of patients received a transfusion, precluding
further analysis of the association of transfusion and mortality.
BLOOD TRANSFUSIONS, like all other medical treatments, should be prescribed
only after consideration of the risks vs the benefits of the therapy. While
the potential risks associated with blood transfusions are well described,1,2 much less is known about the benefits
of blood transfusion. The practice of giving a patient a transfusion for a
hemoglobin level less than 100 g/L (10 g/dL) or a hematocrit less than 0.30
is no longer uniformly accepted.3
We know of only 5 randomized clinical trials, including a total of 207
patients, that have contrasted transfusion triggers.4-8
No differences in mortality or morbidity were found between high and low transfusion
thresholds. However, even collectively, these trials followed too few patients
to evaluate the effect of lower transfusion triggers on clinically important
outcomes. Published observational studies did not adequately control for risk
factors for death or explore transfusion at hemoglobin levels other than 100
g/L (10.0 g/dL).9,10 The lack of
adequate studies on the efficacy of red blood cell transfusion may help explain
the great variation in transfusion practices.11
We evaluated 30- and 90-day postoperative mortality in patients with
hip fracture, comparing those who received preoperative and postoperative
transfusions with similar patients who did not receive transfusions. The goals
of this study were to determine whether red blood cell transfusion influences
mortality at different preoperative and postoperative hemoglobin levels.
The study population included consecutive hip fracture patients, aged
60 years or older, who underwent surgical repair at one of the study hospitals
between 1983 and 1993. Patients were excluded if they refused blood transfusion,
had metastatic cancer, or underwent a surgical procedure involving a site
other than the hip. The 20 participating hospitals were drawn from 4 metropolitan
areas: New Brunswick, NJ; San Antonio, Tex; Philadelphia, Pa; and Richmond,
Va. These hospitals included university, community, and Veterans Affairs medical
centers and were selected based on their willingness to have their medical
We performed a retrospective cohort study. The primary study outcome
was defined as death within 30 days of the operative procedure. The secondary
outcome was death within 90 days of the operative procedure. A National Death
Index search was used to identify deaths that occurred after discharge but
within 90 days of the operation.
A retrospective chart review was conducted using standardized, pretested
forms and an explicit abstraction process. Quality assurance was performed
by reviewing a random sample of medical records. We collected information
on demographic characteristics (age, sex, race, insurance, preadmission residence,
hospital admission year), comorbid conditions (see below), habits (smoking,
alcohol use), medications used prior to admission and during the preoperative
and postoperative time periods, preoperative physical examination (vital signs,
cardiac examination, mental status, motor strength, whether the patient was
malnourished or cachectic, or presence of decubitus ulcer), laboratory results
(electrocardiogram, chest radiograph, arterial blood gas values, echocardiogram,
blood glucose, creatinine, and liver enzyme levels, coagulation tests), cointerventions
(preoperative admission to the intensive care unit, thromboembolism prophylaxis,
antibiotic prophylaxis, physical therapy, respiratory therapy, preoperative
consultations), hip fracture treatment (type of hip fracture, surgical procedure,
use of cement, physical and occupational therapy), operative events (type
and duration of anesthesia, hypotension, tachycardia, arrhythmias, use of
pressors, Swan-Ganz catheter pressures, operative blood loss and intraoperative
transfusions), postoperative complications, and deaths. Detailed information
on hemoglobin levels and timing of transfusion was also collected. The cause
of death is not reported because it was not available in the charts.
Perioperative transfusion was defined as a transfusion occurring within
7 days before or after the surgical repair of the hip fracture. The patient's
perioperative transfusion status was defined separately for the preoperative
and postoperative time periods. In the preoperative transfusion analysis,
patients who received 1 or more transfusions within 7 days prior to the operative
procedure were compared with patients who did not receive a transfusion during
the corresponding time period. Similarly, for the postoperative analysis,
patients who received 1 or more transfusions within 7 days after the operative
procedure were compared with patients who did not receive a transfusion during
the corresponding time period.
For patients receiving a preoperative transfusion, the preoperative
trigger hemoglobin level was defined as the lowest hemoglobin level prior
to the first transfusion and within 7 days prior to surgery. For those who
did not receive a preoperative transfusion, the preoperative trigger was the
lowest hemoglobin level within 7 days prior to surgery. For patients who received
a postoperative transfusion, the postoperative trigger hemoglobin level was
defined as the lowest hemoglobin level prior to the first transfusion. For
those who did not receive a postoperative transfusion, the postoperative trigger
was the lowest hemoglobin level within 7 days after surgery. We did not study
intraoperative transfusions because hemoglobin levels are not routinely measured
during the intraoperative time period.
Information on the presence of many comorbid conditions was collected.
Cardiovascular disease was defined as history of any of the following: myocardial
infarction, angina or ischemic chest pain, coronary artery disease, coronary
artery bypass surgery, percutaneous transluminal coronary angioplasty, history
of congestive heart failure, or history of peripheral vascular disease. Data
were collected on the following other comorbid conditions: history of valvular
heart disease, arrhythmia, hypertension, diabetes mellitus, dementia, stroke
or transient ischemic attack, thromboembolism, chronic lung disease, malignancy,
gastrointestinal bleeding, swallowing disorder, liver disease, arthritis,
hospitalizations within the preceding month, and prior hip fracture.
We also collected information to calculate the Charlson comorbidity
index score,12 the acute physiology score (APS)
of the Acute Physiology and Chronic Health Evaluation (APACHE) II index,13 and the 30-day sickness at admission scale score.14 The APACHE II index is predictive of in-hospital mortality
for critically ill patients.15 The APS was calculated
without the hemoglobin points because hemoglobin is an important a priori
confounding variable and therefore was included as a separate variable in
all models. The Charlson index incorporates many common, serious comorbid
conditions and is a predictor of mortality for medical inpatients.13 We did not include cardiovascular disease points in
the total Charlson score so that we could evaluate the independent effect
of cardiovascular disease. We also used the 30-day sickness at admission scale
that was developed specifically to predict mortality in hip fracture patients
and the American Society of Anesthesiologists (ASA) physical status classification
system that predicts postoperative mortality.16,17
For the postoperative transfusion analysis, the APS was analyzed as a continuous
variable and the Charlson index score as a dichotomous variable (no points
vs any points), and the ASA classifications were grouped into 3 categories
(1 or 2; 3; 4 or 5).
We performed 3 separate analyses: (1) postoperative transfusion and
30-day mortality, (2) postoperative transfusion and 90-day mortality, and
(3) preoperative transfusion and 30-day mortality. Preoperative transfusion
by 90-day mortality was not examined because there was no a priori reason
that this would relate to long-term mortality. Detailed descriptive information
in the tables is only presented for postoperative transfusion and 30-day mortality
for the purposes of brevity and clarity. The results for the other analyses
were very similar.
For each of the separate analyses described above, the unadjusted relationships
between outcome and transfusion status and potential confounders (including
trigger hemoglobin levels, cardiovascular disease, and other patient characteristics)
were assessed using an independent sample t test
or a χ2 test.18 We calculated
the unadjusted odds ratio (OR) for the effect of transfusion instead of the
relative risk, so it could be compared with the adjusted OR generated by a
logistic regression model.
Logistic regression (30-day mortality) and Cox proportional hazards
models (90-day mortality) were used to describe the effect of transfusion
on mortality after adjusting for potential confounders. The trigger hemoglobin
and cardiovascular disease variables were included in all models because of
strong a priori hypotheses about their relationships with transfusion and
In the postoperative transfusion analyses, the trigger hemoglobin level was
included in the model as a continuous variable with values greater than 110
g/L (11.0 g/dL) set to 110 g/L (11.0 g/dL), because we assumed a priori that
there would be no further decrease in risk with increases in hemoglobin levels
above 110 g/L (11.0 g/dL) and to reduce any undue influence of the small number
of very high hemoglobin values. Repeating the analyses without grouping hemoglobin
levels greater than 110 g/L (11.0 g/dL) resulted in similar results. In the
preoperative analysis, the relationship between trigger hemoglobin and 30-day
mortality was not linear, and, therefore, hemoglobin was grouped as follows:
less than 80 g/L (8.0 g/dL), 80 through 89 g/L (8.0-8.9 g/dL), 90 through
99 g/L (9.0-9.9 g/dL), and 100 g/L or more (≥10.0 g/dL). The assumption
that the hazard ratios are proportional could not be rejected.
Potential confounding variables included characteristics that met all
the following criteria: (1) had a statistically significant univariate relationship
with death (P≤.05), (2) were present in at least
5% of the population, and (3) had no expected value less than 5 in the contingency
table analysis. Any variable meeting these criteria was added individually
to a model with transfusion status, trigger hemoglobin, and cardiovascular
disease. All variables maintaining a P value of .10
or less were included in the final model. The potential confounding variables
considered are listed in the data collection portion of the "Methods" section.
Logistic regression (30-day mortality) and Cox proportional hazards
models (90-day mortality) were also used to determine whether the effect of
transfusion differed according to trigger hemoglobin level and cardiovascular
disease status. We constructed regression models with a term for the 3-way
interaction among transfusion status, trigger hemoglobin level, and cardiovascular
disease status along with all second-order interactions and main effect variables.
In the postoperative transfusion analysis of 30-day mortality, the results
suggested a possible 3-way interaction. Therefore, subanalyses using procedures
identical to the overall analysis were performed within hemoglobin stratum
and within cardiovascular disease stratum.
In addition to controlling for confounding in the logistic regression
model as described above, we also stratified patients based on their predicted
probability of receiving a transfusion. First, we constructed a predictive
model for postoperative transfusion. Candidate variables for the model included
those that were clinically plausible, occurred in more than 5% of patients,
and had a significant univariate relationship with transfusion. Stepwise regression
was then used to identify variables for the final model. The probability of
transfusion for each patient was generated using the final predictive model.
We then divided patients into quintiles based on their predicted probability
of receiving a transfusion. We used the Breslow-Day test across quintiles
of predicted transfusion probability to assess homogeneity of the ORs23 and then calculated the common OR using the Mantel-Haenszel
All analyses were performed using SAS version 6.11.25
Of the 9598 patients eligible for the study, 806 were excluded because
a postoperative hemoglobin trigger could not be defined. Of these excluded
patients, 591 received a transfusion, and a hemoglobin level was not recorded
prior to the first postoperative transfusion; 215 of these patients did not
receive a transfusion, and no hemoglobin level was recorded within 7 days
after surgery (30-day mortality was 6.3% and 8.8%, respectively). Five patients
had a postoperative transfusion during a subsequent operative procedure and
were also excluded from the postoperative analysis. The 8787 patients included
in this analysis are described in Table 1. The mean age was 80.3 years (SD, 8.7; range, 60-106). The overall
30-day mortality was 4.6% (n=402; 95% confidence interval [CI], 4.1%-5.0%).
A total of 3699 (42%) of the 8787 patients received a transfusion within
7 days of the surgical repair. The greatest variability in transfusion practice
occurred in patients who received a postoperative transfusion with trigger
hemoglobin levels between 80 and 99 g/L (8.0-9.9 g/dL), in which 2474 (55.6%)
of 4452 patients received a transfusion (Table 2). The vast majority of patients (1014 [90.5%] of 1120) with
a trigger hemoglobin level less than 80 g/L (8.0 g/dL) received a transfusion.
In contrast, only 211 (6.6%) of 3195 patients with a trigger hemoglobin level
greater than 100 g/L (10.0 g/dL) received a transfusion. The clinical characteristics
of the transfused and nontransfused patients are shown in Table 3.
The death rate at 30 days in the postoperative transfused group was
5.3%, compared with 4.0% in the nontransfused group (OR, 1.34; 95% CI, 1.10-1.64).
After adjusting for trigger hemoglobin level using logistic regression analysis,
the OR for 30-day postoperative death after postoperative transfusion vs no
transfusion decreased to 1.15 (95% CI, 0.89-1.48). Adjusting for cardiovascular
disease did not substantially change the OR (OR, 1.11; 95% CI, 0.86-1.44).
The following variables maintained a P value of .10
or less and were included in the final model with transfusion, trigger hemoglobin
level, and cardiovascular disease: APS (continuous), Charlson comorbidity
index (any points vs no points), sickness at admission scale (quartiles),
age (grouped as ages 60-69 years, 70-79 years, 80-89 years, and ≥90 years),
sex, history of atrial fibrillation, history of anemia within 1 year of admission,
whether the patient was malnourished or cachectic on physical examination,
presence of decubiti at admission, abnormal preoperative chest radiograph,
intraoperative tachycardia, preoperative transfusion, and hospital (categorical).
The OR for postoperative transfusion controlling for all these variables further
declined to 0.96 (95% CI, 0.74-1.26). Operative blood loss and intraoperative
transfusion had a P value greater than .10 and were
not included in the final model.
We then evaluated whether the effect of transfusion on mortality was
different for patients with different hemoglobin levels and cardiovascular
disease status. We found a borderline significant result for the 3-way interaction
term among trigger hemoglobin level, cardiovascular disease, and transfusion
status in the unadjusted model (P=.07), with little
change in a fully adjusted model (P=.06). Therefore,
we performed subgroup analyses within 3 hemoglobin strata: 100 g/L or more
(≥10.0 g/dL), 80 to 99 g/L (8.0-9.9 g/dL), and 70 to 79 g/L (7.0-7.9 g/dL),
and by cardiovascular disease status. The results for each of the strata were
similar to one another and to the results of the overall analysis (Table 4). The cardiovascular disease by
transfusion interaction was not significant within the hemoglobin strata subgroups
We also stratified patients by their likelihood of receiving a transfusion.
The variables included in the final logistic regression model to predict postoperative
transfusion were trigger hemoglobin level (divided into 5 groups: ≥110
g/L [≥11.0 g/dL], 100-109 g/L [10.0-10.9 g/dL], 90-99 g/L [9.0-9.9 g/dL],
80-89 g/L [8.0-8.9 g/dL], and <80 g/L [<8.0 g/dL]); hospital (added
as 18 indicator variables); surgical procedure (divided into 4 groups: internal
fixation with pinning, internal fixation other than pinning, hemiarthroplasty,
and total arthroplasty); age (as a continuous variable); anesthesia time (divided
into quartiles); cardiovascular disease; Parkinson disease; anemia within
1 year of admission; tachycardia in the operating room; blood loss (divided
into 3 groups: ≤500 mL or missing, 501-1000 mL, and >1000 mL); femoral
neck fracture; subtrochanteric fracture; and admission year (entered as a
continuous variable). Model discrimination to predict who would receive a
transfusion was excellent (C-statistic=0.90).
The results from the analysis stratified by quintiles defined by probability
of transfusion are presented in Table 5.
The common OR was 1.02 (95% CI, 0.78-1.34). We repeated this analysis excluding
patients with missing values for blood loss, and the results were essentially
The overall 90-day mortality was 9.0% (n=788; 95% CI, 8.4%-9.6%). The
hazard ratio for transfusion was 1.34 (95% CI, 1.16-1.54), adjusted for hemoglobin
level was 1.18 (95% CI, 0.98-1.41), and adjusted for hemoglobin level and
cardiovascular disease was 1.15 (95% CI, 0.96-1.37). The fully adjusted hazards
ratio (using the same procedure as for 30-day mortality) was 1.08 (95% CI,
0.90-1.29). The 3-way interaction among postoperative trigger hemoglobin level,
cardiovascular disease, and transfusion status was not significant (P=.40) nor were any of the 2-way interactions. The adjusted
mortality curves for the transfused and nontransfused groups are displayed
in Figure 1.
A total of 9474 patients were included in the preoperative transfusion
analysis. Of the 9598 patients eligible for the study, 2 were excluded because
their preoperative transfusion status could not be determined from the available
data. An additional 122 were excluded because a preoperative trigger hemoglobin
level could not be defined. A total of 682 (7.2%) patients underwent a preoperative
The results of the preoperative transfusion analysis were similar to
those of the postoperative transfusion analysis. The unadjusted OR for preoperative
transfusion was 2.49 (95% CI, 1.90-3.26). After adjusting for preoperative
transfusion trigger hemoglobin, cardiovascular disease, and other confounding
variables, the adjusted OR was 1.24 (95% CI, 0.81-1.90). The 3-way interaction
among preoperative trigger hemoglobin level, cardiovascular disease, and transfusion
status was not significant (P=.11) nor were any of
the 2-way interactions.
We studied the effect of transfusion in a large, high-risk, elderly
population with extensive comorbidity and were unable to demonstrate that
transfusion was associated with a reduced 30- or 90-day postoperative mortality.
These results suggest that patients who had hemoglobin levels as low as 80
g/L (8.0 g/dL) and did not receive a transfusion were no more likely to die
than those with similar hemoglobin levels who received a transfusion. With
a hemoglobin level less than 80 g/L (8.0 g/dL), nearly all patients received
a transfusion, the effect of which could not be calculated.
We did not confirm prior studies that suggested there might be differences
in the effect of anemia in patients with and without cardiovascular disease.
In a study of 1958 patients who refused blood transfusion for religious reasons,
patients with cardiovascular disease had a greater risk of death than patients
without cardiovascular disease at hemoglobin levels less than 100 g/L (10.0
g/dL).23 Similarly, anemic dogs showed ischemic
ST segment changes and locally depressed cardiac function at higher hemoglobin
levels (70-100 g/L [7.0-10.0 g/dL]) with experimentally created coronary stenoses
varying from 50% to 80%21,26,27
more than normal animals (30-50 g/L [3.0-5.0 g/dL]).20-22
The most important potential limitation of an observational study evaluating
the effect of transfusion on mortality is that transfused patients may systematically
differ from nontransfused patients in ways that cannot be ascertained or controlled
for by a retrospective chart review. Great effort was made to identify and
control for factors that might be associated with transfusion status and death.
Information was collected on a vast array of comorbid conditions. In addition
to information on individual comorbid conditions, we calculated 4 widely used
indexes predictive of short-term mortality (ASA classification, Charlson comorbidity
index, the APACHE II APS, and sickness at admission scale). Intraoperative
events, surgical procedure, and type of anesthesia were assessed. Despite
the extensive and careful assessment of potential confounding variables, it
is still possible that important factors were not captured, and there may
be residual confounding by indication.
Several other limitations should be considered when interpreting the
results of this study. First, the data were collected by medical record review;
however, retrospective data collection should not substantially bias ascertainment
of the 3 primary study variables: transfusion status, hemoglobin level, and
postoperative death. Second, despite the large sample size (this is the largest
study to date to examine this question), inadequate power may still explain
our inability to detect a reduction in mortality related to transfusion; we
may have missed up to a 25% reduction in mortality. However, we estimate that
the study would need to be about 10 times larger to detect a 10% difference
in 30-day mortality with 80% power. Third, this study evaluated the effect
of transfusion on mortality, and it is possible that transfusion may affect
other outcomes such as morbidity, readmission to the hospital, speed of recovery,
and functional status. Fourth, we were unable to analyze the hemoglobin level
that a physician would aim for when giving a patient a transfusion because
of lack of relevant hemoglobin levels. Fifth, the data for the study were
collected over an 11-year period from 20 different hospitals in 4 geographic
regions. Data from earlier admissions may not be entirely comparable with
data from more recent admissions since surgical procedures, anesthetic technique,
physical therapy, and length of hospital stay may have changed. However, year
of admission was not related to either 30- or 90-day mortality, and adjustment
for hospital did not substantively change the results. Sixth, the results
may not generalize to other populations of patients or surgical procedures.
Seventh, the hemoglobin levels during the preoperative and immediate postoperative
periods may not always be accurate because of either dehydration or overhydration.
However, these hemoglobin levels are used by the surgeon to make transfusion
We have found no evidence that transfusion improves survival in these
elderly hip fracture patients with a high burden of chronic disease and hemoglobin
levels greater than 80 g/L (8.0 g/dL). Since there is good evidence about
the risks of transfusion, physicians should reconsider whether transfusion
is warranted for such patients, keeping in mind the lack of evidence about
the possible benefits, such as improving physical function and accelerating
recovery after surgery. A randomized clinical trial is needed to establish
definitively if and when transfusion is indicated in surgical patients with
Create a personal account or sign in to: