OUP user menu

Transfusion and mortality in patients with ST-segment elevation myocardial infarction treated with primary percutaneous coronary intervention

E. Marc Jolicœur, William W. O'Neill, Anne Hellkamp, Christian W. Hamm, David R. Holmes Jr., Hussein R. Al-Khalidi, Manesh R. Patel, Frans J. Van de Werf, Karen Pieper, Paul W. Armstrong, Christopher B. Granger
DOI: http://dx.doi.org/10.1093/eurheartj/ehp279 2575-2583 First published online: 11 July 2009


Aims Red blood cell transfusion is associated with increased mortality among patients with acute coronary syndromes, but little is known about the consequences of transfusion in a contemporary setting of ST-segment elevation myocardial infarction. We describe the association between transfusion and 90-day mortality among patients with acute myocardial infarction treated with primary percutaneous coronary intervention.

Methods and results Analyses were performed on 5532 patients with ST-elevation myocardial infarction from the Assessment of Pexelizumab in Acute Myocardial Infarction trial. The primary objective of this analysis was to ascertain the relation between red blood cell transfusion and 90-day mortality in patients with recent myocardial infarction. We initially determined the baseline and in-hospital predictors of transfusion (multivariable logistic regressions) and subsequently assessed the association between transfusion and mortality using a series of Cox proportional hazards regression combined to a landmark analyses. A total of 213 patients (3.9%) received a transfusion. Transfusion remained significantly associated with mortality [hazards ratio = 2.16 (1.20–3.88)], despite adjustment for baseline characteristics, in-hospital co-interventions, and for propensity of receiving a transfusion. Among patients who survived to hospital discharge, however, the hazard of death was not different in patients treated with transfusion.

Conclusion Transfusion is associated with 90-day mortality in acute myocardial infarction treated with primary percutaneous coronary intervention. Although transfusion may be causally related to mortality, it is likely that at least part of the association is due to confounding. This association illustrates the complex relationship between transfusion, bleeding, and mortality and underscores the need for further research to understand the relationship between transfusion and clinical outcomes.

  • Red blood cell transfusion
  • Bleeding
  • Acute myocardial infarction
  • Primary percutaneous coronary intervention


Red blood cell transfusion has historically been assumed beneficial in critically ill and anaemic cardiac patients. This assumption has been challenged recently by observational analyses linking transfusion to mortality in patients with non-ST-elevation acute coronary syndromes (ACS).1,2 To this date, however, little information is available on the consequence of transfusion during acute ST-elevation myocardial infarction (STEMI), particularly when treated with primary percutaneous coronary intervention (PCI).

Several analyses have suggested that the observed association between transfusion and mortality is confounded by anaemia and bleeding, which are markers of disease severity.3,4 As a result, controversy exists concerning the risk benefit of transfusion in this cardiac populations.5,6 Likewise, previous analyses linking transfusion to mortality in patients with ACS have been largely confounded by the performance of invasive revascularization procedures. These procedures, usually performed in association with antithrombotic and antiplatelet therapy may cause bleeding, and are more frequently used in unstable or higher risk patients with worse prognosis.7 We assessed whether red blood cell transfusion was associated with increased mortality at 90 days in a contemporary population of patients with myocardial infarction. Using the database of a large STEMI trial where all patients were treated with primary PCI,8 we describe the use of blood transfusion, the baseline and in-hospital predictors of transfusion, and the adjusted relationship of transfusion and 90-day mortality.


Patients population

The rationale and design of the APEX AMI trial have been published previously.8,9 Briefly, APEX AMI was a randomized placebo-controlled trial comparing the effect of pexelizumab (an inhibitor of complement) to placebo on all-cause 30-day mortality in 5745 patients with STEMI treated with primary PCI in 17 countries and 296 sites from 2004 to 2006. Patients were deemed eligible for the study if presenting for primary PCI with symptoms for more than 20 min but for <6 h, and showing high-risk myocardial infarction electrocardiographic characteristics. Key exclusion criteria included use of fibrinolytic therapy, isolated inferior MI, active serious infection, or other serious medical conditions likely to alter recovery. All institutional review boards of participating medical centres approved the APEX AMI protocol and all enrolled patients gave written informed consent. Patients who underwent coronary artery bypass grafting (CABG) during their hospitalization (n = 204) were excluded from the current analysis because of the high use of transfusion related to specific issues of surgery.10

Clinical events definitions

Transfusion and bleeding

All red blood cell transfusions given after enrolment until hospital discharge or day 14 (whichever came first) were systematically and prospectively recorded on the electronic case report form. The occurrence of transfusion after discharge from the index hospitalization was not recorded. Timing of transfusion was not specifically collected, and thus was estimated from the date and time of bleeding or the date and time of nadir haemoglobin (when no bleeding was experienced).1,2,11 Bleeding events were individually recorded and classified according to the Global Use of Strategies to Open Occluded Coronary Arteries (GUSTO) classification, at their highest level of severity.12 Transfusions and bleeding were ascertained by site investigators.

Other clinical events

The clinical events of major focus (death, cardiogenic shock, and congestive heart failure) were collected until 90 days in all patients and centrally adjudicated by a blinded clinical events committee.8 Other events such as reinfarction, arrhythmia, strokes, invasive procedures, rehospitalization, and serious infection were collected as well.

Statistical analysis

Continuous variables are presented as median (interquartile range) and compared using Wilcoxon rank sum tests. Discrete variables are presented as number (percentages), and compared using either chi-square (χ2) or Fisher's exact tests. Kaplan–Meier methods were used to illustrate mortality rates at 90 days for patients who did and did not undergo transfusion. Statistical significance was determined at the two-sided α = 0.05 level. Data analyses were performed with SAS® software package (version 8.2, SAS Institute, Cary, NC). The authors had full access to the data and take responsibility for its integrity. All authors have read and agreed to the manuscript as written.

Modelling of outcomes

The primary objective of this analysis was to ascertain the relation between red blood cell transfusion and 90-day mortality in patients with recent myocardial infarction. To adjust for known or expected confounding variables, a series of multivariable models was developed. We initially established the baseline and in-hospital (post-enrolment) predictors of transfusion (Model 1 and Model 2, respectively) so that we could adjust for propensity of receiving a transfusion in the 90-day mortality analysis (Model 3; primary analysis). Finally, we explored the remote consequence of transfusion by excluding in-hospital death (Model 4). For each model, candidate variables were prespecified, using either clinical experience or previously reported multivariable analyses.11,12

The first model identified baseline independent predictors of transfusion using multiple logistic regression analysis with both stepwise and backwards variable-reduction techniques. Candidate variables included demographics (age, gender, weight, and African ancestry), baseline characteristics (heart rate, diastolic blood pressure, haemoglobin, creatinine clearance by MDRD formula), prior medical history, and medication (diabetes mellitus, hyperlipidaemia, bypass surgery, beta-blockers, nitrate, aspirin, and thienopyridine), pre-randomization heparin or glycoprotein IIb/IIIa inhibitor use, and country of enrolment. In previous studies, pexelizumab was not associated with bleeding or transfusion.9 For all models, candidate variables that either showed marginal associations to outcome on univariate testing (P < 0.20) or were judged clinically important were included in the multivariable analysis. The linearity assumption was evaluated by applying restricted cubic spline transformations to the continuous measures. The likelihood ratio χ2 for models with these transformations were compared with those in the linear models to assess the need for nonlinear terms. In case of nonlinearity, the plot of the restricted cubic spline transformation vs. the logit of the outcome was used to assist in identifying the appropriate transformations to apply. The baseline model's discrimination was evaluated by the c-index. The significance of multivariable regression coefficients was assessed with the Wald χ2. Regression coefficients were not adjusted for multiple comparisons.13 Only patients with complete information were included in the analysis and no data imputation was performed.

In the second model, we identified post-randomization events that remained independent predictors of transfusion after taking into account baseline characteristics. To this end, we used a landmark analysis. Landmarks as we are defining them are time-points in the follow-up period from which survival analyses are started.1,14 Before analysis, landmarks were prespecified at 24, 48, 72, and 96 h after randomization on the assumption that a majority of clinical events (bleeding, transfusion, and co-interventions) would occur by 96 h. Baseline characteristics and in-hospital events that occurred before a given landmark were modelled to predict the occurrence of transfusion after the landmark. For instance, events occurring after randomization but before the 24 h landmark were modelled to predict use of transfusion after 24 h. Likewise, events occurring after randomization but before the 48 h landmark were used to predict transfusion after the 48 h, and so forth. In other words, each model is conditional on having survived without a transfusion to the next landmark period. The independent variables used in the second model combined baseline pre-randomization variables (first model) with key post-randomization co-interventions: repeated cardiac catheterization, repeated PCI, intra-aortic balloon pump (IABP), Swan-Ganz catheter, permanent pace-maker, automatic implantable cardio defibrillator, and dialysis. Only co-interventions occurring prior to the first transfusion were considered into the analysis.

The third model ascertained the relation between transfusion and 90-day mortality. We used a landmark analysis to minimize the survival bias whereby patients who die soon after their myocardial infarctions are less likely to receive blood transfusion.14 Survival was compared between patients who did and who did not undergo transfusion by Cox proportional hazards regression, at each prespecified landmarks (24, 48, 72, and 96 h). Only patients alive at the given landmark time point are included in the analysis. This approach provides a trend of the adjusted association between transfusion and the 90-day mortality over time.1 The model was adjusted for baseline variables previously shown to predict 30-day mortality in the APEX-AMI trial (age, gender, weight, history of diabetes, systolic blood pressure and heart rate, total ST-segment change, qualifying Killip class),15 in-hospital interventions (recatherization, additional PCI, IABP). The stability of the model was tested after controlling for the propensity of receiving a transfusion. A transfusion propensity score, i.e. the predicted probability of receiving a transfusion, was derived for each patient using all candidate predictors from the baseline model as well as all two-way interactions. The candidate predictors used to derive the transfusion propensity score included: enrolment in North America, sex, race, age, weight, baseline haemoglobin, creatinine clearance, diastolic blood pressure, heart rate, prior CABG, diabetes, use of beta-blockers, nitrates, and GP IIb/IIIa inhibitors. Adequacy of the propensity score for balancing confounding variables was assessed by confirming that no candidate transfusion predictors were significantly different between transfused and non-transfused patients after adjustment for the propensity score.16 To explore the possibility of a dose–response relationship between transfusions and mortality risk, the number of units transfused was added to the model. Non-transfused patients were assigned a value of 0.

The fourth model assessed the remote consequences of blood transfusion on mortality and is considered an exploratory analysis. The survival analysis was restricted to patients who survived to hospital discharge. The Cox proportional hazard model used the date of discharge as time 0 for each patient. In addition to the candidate predictors specified in the third model, we adjusted for discharge medication (aspirin, thienopyridines, beta-blockers, ACE inhibitors, statins, and oral anti-coagulants).

Patients who were missing predictor or outcome data for a given model were omitted from that model. Of the 5532, non-CABG patients with known transfusion status, 4652 were used in the models predicting transfusion, 5347 were used in the adjusted mortality model, and 4500 were used in the mortality model containing both covariates and transfusion propensity score.

The multiple logistic regression models were internally validated using bootstrapping and the baseline model c-index was corrected for optimism. Bootstrapping was performed by taking a large number of samples from the entire data set and provides relatively unbiased estimates of predictive accuracy.17 Two hundred and fifty samples were randomly drawn from the original cohort with replacement. For every sample, a multiple regression model was fitted, coefficients were estimated, and a prediction equation was generated, after which a c-statistic (Cboot) was calculated. The prediction equations generated in each sample were run in the original cohort and c-statistic (Corig) was again calculated. Optimism was obtained by the subtracting Corig from Cboot. The bootstrap-corrected performance on the original transfusion model was obtained by subtracting the averaged optimism (obtained from each bootstrap sample) from the overall model c statistic (Capparent).


Baseline characteristics and medical management

The study included 5532 patients with STEMI treated with primary PCI. A total of 213 patients (3.9%) received a blood transfusion during their hospitalization (median of 2 units per patient). Patients treated with blood transfusion were more likely to be women, older, had lower median body weight (Table 1), and a higher prevalence of co-morbidities, including diabetes, hypertension, and history of congestive heart failure. At hospital presentation, patients treated with transfusion had lower systolic and diastolic blood pressure, lower baseline haemoglobin and creatinine clearance, and worse Killip class. During hospitalization, IABP and dialysis were more frequently performed among patients initially treated with transfusion. Transfusions were predominantly given when moderate or major bleeding occurred (81.7% of cases). Patients treated with transfusion during their index hospitalization were more likely to die in the following 90 days [24.9 vs. 3.9%, unadjusted hazards ratio (HR) = 6.30, 95% confidence interval (CI); 4.14–9.59]. Table 2 compares the unadjusted rates of major adverse cardiac events at 90 days between patients who did or did not undergo transfusion.

View this table:
Table 1

Baseline characteristics, in-hospital interventions, and medications among patients who did and did not undergo blood transfusion

CharacteristicsTransfusion (n = 213)No transfusion (n = 5319)P-value
 Age, median (IQR), years 71 (58, 79)  61 (52, 70)0.001
  Age ≥ 75 87 (41) 850 (16)0.001
 Female, n (%)112 (53)1170 (22)0.001
 Weight, median (IQR), kg 74 (64, 85)  80 (70, 91)0.001
 Race, n (%)0.23
  Caucasian197 (92)5024 (94)
  African American  6 (3) 124 (2)
  Hispanic  6 (3)  57 (1)
  Others  4 (2) 114 (2)
Medical history, n (%)
 Diabetes mellitus 48 (23) 816 (15)0.007
 Hypertension130 (61)2596 (49)0.001
 Prior myocardial infarction 40 (19) 625 (12)0.004
 Active smoking 56 (27)2346 (44)0.001
 Prior heart failure 17 (8) 184 (3)0.003
 Prior stroke 14 (7) 189 (4)0.037
 Prior PCI 33 (15) 511 (10)0.008
 Prior CABG  5 (2) 119 (2)0.92
Presenting characteristics
 Heart rate, median (IQR), b.p.m. 80 (65, 96)  75 (65, 86)0.001
 Systolic blood pressure, median (IQR), mmHg129 (110, 150) 133 (118, 150)0.012
 Diastolic blood pressure, median (IQR), mmHg 74 (62, 86)  80 (70, 90)0.001
 Infarct location, n (%)0.66
 High-risk inferior 84 (39)2178 (41)
 Anterior129 (61)3141 (59)
 Killip class, n (%)0.001
 I–II198 (93)5222 (98)
 III–IV 15 (7)  95 (2)
 Serum creatinine, median (IQR), µmol/L 97 (80, 115)  88 (80, 106)0.003
 Creatinine clearance, median (IQR), mL/min/1.73 m2 64 (47, 77)  75 (63, 89)0.001
 Baseline haemoglobin, median (IQR), g/dL 12.9 (11.2, 14.4)  14.8 (13.7, 15.8)0.001
Baseline medications, n (%)
 Aspirin153 (90)3713 (83)0.016
 Low molecular weight heparin 67 (39)2045 (46)0.093
 Thienopyridine 62 (36)1376 (31)0.13
 Glycoprotein IIb/IIIa inhibitors 32 (15) 806 (15)0.96
In-hospital interventions, n (%)
 Drug eluting stent, infarct-related artery 89 (44)2106 (42)0.45
 Repeat cardiac catheterization 22 (10) 303 (6)0.01
 Repeat PCI 15 (7) 375 (7)0.99
 Intra-aortic balloon pump 64 (30) 273 (5)0.001
 Left-ventricular assist device  2 (1)  22 (<1)0.32
 Implantable cardiac defibrillator  3 (1)  24 (<1)0.11
 Dialysis 10 (5)  11 (<1)0.001
Nadir haemoglobin, median (IQR), g/dL  8.7 (8.1, 9.3)  12.9 (11.8, 13.9)0.001
Medication at discharge, n (%)
 Aspirin169 (79)5079 (95)0.001
 Thienopyridines159 (75)4819 (91)0.001
 Beta-blocker142 (67)4729 (89)0.001
 ACE inhibitor124 (58)4210 (79)0.001
 Statin163 (77)4914 (92)0.001
 Oral anticoagulant 16 (8) 329 (6)0.45
  • CABG, coronary artery bypass graft surgery; b.p.m., beat per minute; Hb, haemoglobin; IABP, intra-aortic balloon pump; ICD, implantable cardiac defibrillator; IQR, interquartile range; LVAD, left ventricular assist device; PCI, percutaneous coronary intervention.

  • Categorical variables are shown as number (percent) with P-values from likelihood ratio χ2 tests. Continuous variables are shown as median (25th, 75th percentiles) with P-values from Wilcoxon rank sum tests. Creatinine clearance calculated using MDRD formula.

View this table:
Table 2

Unadjusted incidence of major cardiac events among patients who did and who did not undergo blood transfusion

Events by 90 days, n (%)Transfusion (n = 204)No transfusion (n = 4984)
Death53 (26.0)203 (4.1)
  Sudden cardiac death14 (6.9) 86 (1.7)
  Non-sudden cardiac death20 (9.8) 87 (1.8)
 Non-cardiac15 (7.4) 24 (0.5)
 Unknown cause 4 (2.0)  6 (0.1)
Re-infarction16 (7.8)128 (2.6)
CHF31 (15.2)217 (4.4)
Stroke10 (4.9) 50 (1.0)
  • This table includes only patients whose 90-day vital status is known. For non-fatal endpoints, occurrences both before and after transfusion are included.

At discharge, patients requiring blood transfusion were less likely to be treated with medications recommended by national guidelines.18 Aspirin and thienopyridine uses were significantly lower in comparison to patients who did not undergo blood transfusion. Beta-blockers, ACE-inhibitors, and statins use was also lower among transfused patients, a finding likely related to the higher rates of co-morbidities in this group of patients.19

Independent predictors of transfusion

In this study, two multivariable models were initially built to identify the baseline and the post-randomization predictors of blood transfusion (Table 3). In the model of baseline characteristics, haemoglobin and enrolment in a North American hospital were the strongest predictors of blood transfusion. The model performed well in discriminating patients with and without a transfusion (c-index = 0.84). After internal validation, the c-index remained at 0.83 (over-optimism = 0.04). In the second model where in-hospital variables where added to baseline predictors, the use of an IABP was the only additional independent predictor of transfusion. The association between this variable and the use of blood transfusion remained consistent during the pre-specified 96 h landmark time-points.

View this table:
Table 3

Baseline and in-hospital independent predictors of blood transfusion for patients with acute myocardial infarction

Transfusion odds ratio (95% confidence interval)
VariableModel 1. Baseline variablesaModel 2. Baseline and in-hospital variables
Baseline haemoglobin (per 10 g/L decrease)1.51 (1.39–1.64)1.44 (1.31, 1.59)
Enrolment in North America2.40 (1.74–3.31)2.08 (1.51, 2.85)
Creatinine clearance < 85 mL/min/1.73 m2 (per 10 mL/min/1.73 m2 decrease)1.24 (1.08–1.40)1.24 (1.13, 1.36)
Heart rate > 65 b.p.m. (per 10 b.p.m. increase)1.18 (1.09–1.28)1.17 (1.10, 1.26)
Female sex1.86 (1.33–2.60)1.89 (1.33, 2.68)
Diastolic blood pressure > 80 mmHg (per 10 mmHg decrease)1.23 (1.08–1.40)1.20 (1.06, 1.36)
Age (per 10 years increase)1.18 (1.03–1.36)1.18 (1.02, 1.36)
Intra-aortic balloon pump3.18 (2.05, 4.95)
  • b.p.m., beats per minute; mmHg, millimeter of mercury.

  • Predictors are presented in decreasing order of importance as determined by the χ2 for individual coefficients.

  • aCorrected c-index = 0.83.

Association between blood transfusion and mortality

After multivariable adjustment, blood transfusion independently predicted mortality at 90 days (HR = 2.21, 95% CI; 1.32–2.99) (Figure 1A and Table 4). The association between transfusion and mortality remained stable after further adjustment for the propensity of receiving a blood transfusion (HR = 2.16, 95% CI; 1.20–3.88) (primary analysis). The propensity score for receiving a transfusion adequately balanced the covariates that were confounded with transfusion (tests of differences in these variables between transfused and non-transfused patients after adjustment for the propensity score all had P > = 0.20)16 (see Supplementary material online, Table S5). Throughout the prespecified landmark time-points, the association between transfusion and 90-day mortality remained relatively stable (Figure 2). This suggests that survival bias has a trivial influence on the association between transfusion and mortality.14 The relationship of excess risk and transfusion did not differ according to North American vs. other regions (P interaction = 0.16). When added to the model, the number of units transfused had a non-significant effect overall (HR = 0.94, 95% CI 0.62–1.44 for each additional unit transfused, P = 0.79).

Figure 1

Kaplan–Meier estimates of 90-day mortality among patients who did or did not undergo blood transfusion. (A) All patients. (B) Patients who survived to hospital discharge.

Figure 2

Transfusion as an independent predictor of 90-days mortality: landmark analysis. The hazards ratios are adjusted for baseline variables that independently predicted mortality in the APEX-AMI study (age, history of diabetes, systolic blood pressure and heart rate, total ST-segment change, Killip class at presentation, anterior MI) and in-hospital interventions (recatherization, IABP).

View this table:
Table 4

Hazards ratios of 90-day mortality for patients treated with transfusion compared with patients not treated with transfusion

ModelsHazards ratio (95% confidence interval)
Unadjusted analysis6.30 (4.14–9.59)
Multivariable analyses
 1. Baseline characteristics and in-hospital co-interventionsa2.21 (1.32–2.99)
 2. Baseline characteristics and in-hospital co-interventionsa adjusted for transfusion propensity scoreb2.16 (1.20–3.88)
 3. Patients who survived to hospital dischargea with medications at dischargec2.18 (0.73–6.54)
  • aIndependent predictors: age, history of diabetes, systolic blood pressure and heart rate, total ST-segment change, qualifying Killip class, anterior myocardial infarction, recatherization, and intra-aortic balloon pump.

  • bPrimary analysis: transfusion propensity score is the predicted probability of transfusion from a logistic regression model including enrolment in North America, sex, race, age, weight, baseline haemoglobin, creatinine clearance, diastolic blood pressure, heart rate, prior CABG, diabetes, use of beta-blockers, nitrate, or GP IIb/IIIa inhibitors.

  • cA total of 48 deaths were recorded between discharge and Day 90. Medications collected at discharge: aspirin, thienopyridines, beta-blockers, ACE inhibitors, statins, and oral anti-coagulants.

Among the 53 in-hospital deaths in the 204 patients treated with transfusion, two were due to bleeding. Among the 203 in-hospital deaths in the 4894 patients not treated with transfusion, 1 patient (0.8%) was due to bleeding. Likewise infection was reported as the principal cause of in-hospital death in two patients treated with transfusion vs. three patients not treated with transfusion.

Among patients who survived to their hospital discharge, the hazards ratio for transfusion and 90-day mortality was consistent with higher mortality but no longer statistically significant (HR = 2.18, 95% CI; 0.73–6.54), possibly related to the smaller number of events in this subset (Figure 1B).


Red blood cell transfusion is associated with increased mortality in a contemporary population of patients with STEMI treated with primary PCI. After adjusting for the propensity of receiving transfusion and other key mortality predictors and biases, patients who received transfusion were still about twice as likely to die after their myocardial infarctions. Although transfusion may be causally related to 90-day mortality, it is also possible that the association is accounted for the effects of anaemia and bleeding, which could be directly resulting in death and/or indirectly related to risk of death as markers of disease severity.3,19,20

There has been in the recent years a growing body of evidence that the relationship between transfusion and death may be, at least in part, cause and effect. That evidence has been based on a series of observational studies linking transfusion to mortality,1,2,5,20,21 combined with several plausible biological models for how transfusion might be harmful.22,23 As is true for any claim of causation, however, the available evidence linking transfusion to mortality needs to be weighed according to its consistency, strength, and biological plausibility.24

In terms of consistency, the association between transfusion and mortality has not been generally constant among various ACS populations and studies. Our findings parallel those of previous publications showing an adverse association between transfusion and mortality in patients with ischaemic heart disease.1,2 Rao et al.1 showed a strong association between transfusion and death at 30 days among patients from three large international non-ST-elevation acute coronary syndrome trials (adjusted HR = 3.94; 95% CI 3.26–4.75). In a randomized controlled trial testing thresholds of transfusion, Hebert et al.5,25 initially showed equivalence between restrictive and liberal transfusion strategies in critically ill patients, but could not confirm their finding in the subgroup of patients with acute coronary syndromes. In contrast, Wu et al.6 conducted a retrospective analysis on 78 974 Medicare beneficiaries hospitalized for ACS and found transfusion to be protective in patients with admission haematocrit of <30%, although lack of timing of transfusion prevented adjustment for time-dependent confounding. Sabatine et al.4 found transfusion to be associated with decreased risk of cardiovascular death among patients with STEMI when their baseline haemoglobin was <12 g/dL (OR = 0.42, 95% CI; 0.20–0.89). Interestingly, the same group reported transfusion to be associated with an increased risk of the composite of death, reinfarction, and recurrent ischaemia among patients with NSTE ACS. Overall, studies that included patients with major bleeding in their analyses consistently showed a direct association between transfusion and mortality,1,2 whereas studies that excluded patients with bleeding from their analyses were either neutral or inconclusive.46 This raises the possibility that bleeding, not transfusion, is the major determinant of mortality. The recent finding of the HORIZONS-AMI trial further support this hypothesis,26 as anticoagulation strategy which reduced bleeding are associated with better outcomes. The linkage of transfusion and bleeding illustrates the limitation of multivariable analyses to resolve this type of complex relationship (see Supplementary material online, Table S6). In our study for instance, 82% of the patients who received blood transfusion were also noted to have experienced a moderate or severe bleeding episode. The absence of a dose–response effect of transfusion in our analysis could reflect insufficient power of a low signal-to-noise ratio (too few patients and events). It could also mean that indeed there is no dose–response relationship and that the first transfusion may carry the bulk of the risk, with little additional risk from subsequent transfusions. The failure to find a dose–response relationship suggests that there may not be a direct link between transfusion and mortality.

In terms of strength of association, our primary analysis shows a hazards ratio of borderline significance for transfusion (HR = 2.16, lower bound of the 95% CI = 1.20). It is likely that unknown confounders accounted for a portion of the association, as multivariable analyses can only adjust for confounders that have been considered and measured.27 Bleeding is a well-established predictor of adverse events among patients with acute coronary syndrome.3,19,20 Several dimensions of the decisional process for transfusion cannot be adequately captured by in a retrospective analysis, such as the haemodynamic stability and the overall clinical status at the moment of transfusion. This study illustrates the challenges of sorting out complex interactions between transfusion, bleeding, co-morbidities, and mortality. When we restricted the survival analysis to patients who survived their hospitalization, the association between transfusion and mortality was no longer significant (Table 4). Since a majority of deaths occurred during the index hospitalization, it is possible that this model was underpowered to detect the association between transfusion and mortality, after removal of a significant proportion of the susceptible patients.

Some biological hypotheses have been proposed to explain how transfusion could cause harm in patients with recent myocardial ischaemic injury. Storage of blood cells leads to cellular depletion of 2,3 disphosphoglycerate and nitric oxide,23,28 which in turn increases the affinity of haemoglobin for oxygen and decreases oxygen carriage in the blood. It has been hypothesized that when transfused to patients with acute myocardial damage, stored red blood cells could induce microvascular obstruction and adverse inflammatory reaction.3 However, clinical trials testing the effect of stored red blood cells on organ perfusion and microcirculation have reached discordant conclusions.29,30 In a small randomized controlled trial, Suttner et al.31 could not quantify any deleterious change in the skeletal muscle oxygen tension (measured by polarographic microprobes) after transfusion in a population of patients undergoing elective CABG. It has also been suggested that transfusion may cause immunosuppression and increase the likelihood of infection.32,33 In the APEX-AMI trial, similar rates of major infections were recorded in the 180 days following enrolment among transfused patients and non-transfused patients (2.4 vs. 1.0%, P = 0.11). Another plausible explanation for higher mortality among patients treated with transfusion is lower use of both antithrombotic and other medical treatments known to improve outcome.

Our study also provides a contemporary assessment of predictors of transfusion in the setting of an acute myocardial infarction, where primary PCI is widely used.34 The rate of transfusion observed in this study (3.9%) may be significantly lower than one might expect in general clinical practice.2 Of note, the first model using baseline clinical variables had excellent discrimination ability, with a c-index of over 0.80. This suggests that the model might reliably identify groups of patients with STEMI at risk of transfusion and could be used to adjust certain medical therapy accordingly. The use of Gp IIb–IIIa inhibitors did not independently predict the use of transfusion.11

Except for baseline haemoglobin, enrolment in North America was the strongest predictor of transfusion. The fact that transfusion rates were higher in North America, even in a multivariable model accounting for other predictors, suggests heterogeneity in the practice of transfusion driven by physician behaviour and an opportunity to improve practice if one approach was shown to be preferable. The relationship of excess risk and transfusion appeared to be consistent across regions.

Study limitations

Our study has certain limitations. First, transfusion is a post-enrolment event and is thus confounded by multiple factors. We believe that our series of multivariable analyses complemented with a landmark analyses has limited this confounding as much as possible. We tested the robustness of our findings with multiple sensitivity analyses, including an adjustment for the propensity of receiving a blood transfusion. Because of the multicollinearity shared between transfusion and bleeding, attempts to adjust for the later in multivariable mortality analyses resulted in model instability. Therefore bleeding could not be used as a candidate variable in the mortality model. Bleeding therefore remains a factor confounding the association between transfusion and mortality. Randomized controlled trials testing the use of transfusion are needed to define the consequence of transfusion on mortality. In the absence of direct evidence from randomized controlled trial, our study in the context of prior studies suggests that transfusion may well contribute to increased risk of death. Second, the conclusions presented in this study may be limited to patients treated with primary PCI. However, because all APEX-AMI patients had a primary PCI, and because invasive procedures are strongly related to bleeding, our analysis is less susceptible to the interventional bias usually encountered in transfusion outcome studies.35 Third, patients who survived to their myocardial infarction were more likely to receive blood transfusion. There is no perfect approach to control for the survival bias.36 However, the consistency of the association throughout the time-points in the landmark analysis suggests that survival bias did not play a significant role.14 To support the landmark analysis and to further explore the temporal dimensions of transfusion and mortality, we use a statistical model restricted to patients who were discharged alive. Findings from this model should be interpreted with caution because of the smaller number of patients in this subgroup, and because of the possibility of data over-fitting. Fourth, we relied on the date of bleeding as a proxy for the date of transfusion. It is possible that some misclassification might have happened. This possibility is limited given the clinical overlap between bleeding and transfusion.1,11 In addition, the use of transfusion as a marker of a preceding bleeding event has been validated in a previous study.2 Finally, some key variables that may be associated with bleeding were not collected, such as past history of peptic ulcer disease, use of gastro-protective agents, the vascular access (radial vs. femoral), and the vascular sheath dwell time.


The association between transfusion and mortality is consistent and independent of measured confounding factors in a population in which all patients underwent invasive procedures. Nevertheless, this does not prove cause and effect. Our results highlight the need for prospective cohort studies specifically designed to carefully collect clinical characteristics and events along with key biomarkers might better define the relationships and pathophysiology, and randomized studies of transfusion strategies are needed to define best practice.


The APEX-AMI trial was jointly funded by Procter & Gamble Pharmaceuticals and Alexion Pharmaceuticals. E.M.J. is supported by the Cardiovascular Clinical Research Fellowship Award, Hoffman-Laroche, Canada, and the Montreal Heart Institute Foundation, Montreal, Canada.

Conflict of interest: none declared.


The authors would like to thank Amanda Stebbins (Duke Clinical Research Institute, Durham, NC, USA) who made an important contribution to the development of the statistical model.


View Abstract