Author + information
- Received January 6, 2014
- Revision received April 10, 2014
- Accepted May 1, 2014
- Published online August 26, 2014.
- Adam C. Salisbury, MD, MSc∗,†∗ (, )
- Kimberly J. Reid, MS∗,
- Steven P. Marso, MD∗,†,
- Amit P. Amin, MD, MSc‡,
- Karen P. Alexander, MD§,
- Tracy Y. Wang, MD, MHS, MSc§,
- John A. Spertus, MD, MPH∗,† and
- Mikhail Kosiborod, MD∗,†
- ∗Department of Cardiology, Saint Luke’s Mid America Heart Institute, Kansas City, Missouri
- †University of Missouri–Kansas City School of Medicine, Kansas City, Missouri
- ‡Division of Cardiology, Washington University School of Medicine, St. Louis, Missouri
- §Division of Cardiology, Duke Clinical Research Institute, Durham, North Carolina
- ↵∗Reprint requests and correspondence:
Dr. Adam C. Salisbury, Saint Luke’s Mid America Heart Institute, 4401 Wornall Road, Kansas City, Missouri 64111.
Background Blood transfusion is controversial for anemic patients with acute myocardial infarction (AMI), with some previous studies reporting increased risk of transfusion-associated mortality.
Objectives The goal of this study was to examine variability in blood transfusions across hospitals and the relationship between blood transfusion and in-hospital mortality in a large, contemporary cohort of consecutive AMI patients.
Methods Among 34,937 AMI hospitalizations from 57 centers, patients receiving at least 1 packed red blood cell transfusion were compared with those who were not transfused. Using 45 disease severity, comorbidity, laboratory, and in-hospital treatment variables, we propensity matched patients who did and did not receive a packed red blood cell transfusion. A conditional logistic regression model was used to identify the association between transfusion and in-hospital mortality.
Results A total of 1,778 patients (5.1%) had at least 1 transfusion. In unadjusted analyses, transfusion was associated with higher in-hospital mortality (odds ratio: 2.05 [95% confidence interval: 1.76 to 2.40]). The vast majority of patients (91.1%) with and without transfusion had nonoverlapping propensity scores, reflecting incomparable clinical profiles. Thus, they were excluded from the propensity-matched analyses. After propensity matching those with overlapping scores, blood transfusion was associated with a reduced risk of in-hospital death (odds ratio: 0.73 [95% confidence interval: 0.58 to 0.92]).
Conclusions The majority of patients undergoing blood transfusion in clinical practice cannot be matched with nontransfused patients due to their markedly different clinical profiles. Among comparable patients, blood transfusion was associated with a lower risk of in-hospital mortality. These findings suggest that previous observational reports of increased mortality with transfusion may have been influenced by selection bias, and they highlight the need for randomized trials to establish the role of transfusion during AMI.
Anemia is common at the time of acute myocardial infarction (AMI) and has been shown to portend a poor prognosis, including greater short-term and long-term mortality (1–4). In the setting of AMI, administration of packed red blood cells may augment hemoglobin levels and improve myocardial oxygen delivery, but it also carries risks, including volume overload, increased thrombogenicity, impaired oxygen delivery, and a risk of infection (5,6). Despite the widespread use of blood transfusions in clinical practice, the safety and efficacy of this method have not been evaluated in large, randomized clinical trials. Accordingly, the use of blood transfusion in AMI patients remains controversial, with some observational studies suggesting benefit in patients with low nadir hemoglobin values (1,7,8), whereas others have reported increased mortality (9–11).
A major challenge in the interpretation of transfusion and outcomes in observational studies is the impact of confounding. Clinicians select treatments, such as transfusion, after considering a broad array of factors, including the perceived benefits and risks for each individual patient. In such cases, observational studies may yield a relationship between treatment and outcome that primarily reflects the underlying high-risk characteristics of treated patients. Although confounding can be minimized by the use of instrumental variables or propensity matching (12), these methods have not been uniformly used in previous research.
To further illuminate the association between transfusion and survival in anemic AMI patients, we used the Cerner Health Facts database (13,14), which collects data through the electronic medical record on consecutive AMI patients at 57 U.S. hospitals. Given the large size of the patient population and the detailed collection of in-hospital laboratory, treatment, and complication data, we were able to conduct a propensity-matched analysis to specifically focus on the patients eligible for transfusion. Importantly, some patients have such life-threatening anemia that they would always be transfused, while other, “healthier” patients would rarely receive a transfusion; inclusion of such patients could lead to substantial selection bias. Finally, given the diverse collection of hospitals participating in Health Facts, we were able to examine the variability in blood transfusion practices across institutions in real-world practice.
Health Facts captured de-identified data from the Cerner electronic medical record for patients admitted to participating hospitals between January 1, 2000, and December 31, 2008. Data collected included patients’ demographic characteristics, medical history, and comorbidities (using the International Classification of Diseases-Ninth Revision-Clinical Modification [ICD-9-CM], codes), laboratory studies, medications, procedures, and complications. A total of 78 hospitals contributed data to Health Facts. The median number of AMI patients from each hospital was 219 (interquartile range [IQR]: 48 to 1,030), and the median duration of hospitals’ participation was 2.9 years (IQR: 1.2 to 5.3 years). All data were de-identified before being provided to the investigators, and the institutional review board of Saint Luke’s Hospital provided an exemption to review.
We included all patients hospitalized with a primary discharge diagnosis of AMI as determined by using ICD-9-CM diagnostic codes 410.xx, and AMI was further confirmed by requiring that patients have at least 1 elevated cardiac biomarker (troponin or creatine kinase-myocardial band). Patients known to be transferred from other hospitals (full laboratory testing data may not be available) or from hospice (goals of care differ from the overall population) were excluded. Inclusion and exclusion criteria are listed in detail in Figure 1. Important exclusions were patients admitted from hospitals contributing <20 patients to Health Facts, those with very long lengths of stay (>31 days), and patients who underwent coronary bypass grafting, valve replacement, or valve repair during hospitalization. Patients without a recorded hemoglobin assessment and any patient with a length of stay <1 day were also excluded. The final analytic cohort included 34,937 patients with AMI from 57 hospitals.
Blood transfusion was defined by using ICD-9 procedure codes for administration of packed red blood cells, and each patient with an ICD-9 code for packed cell administration recorded during their index hospitalization was considered to have received a red blood cell transfusion. In-hospital mortality was the primary outcome.
Baseline patient characteristics, laboratory values, in-hospital treatments, and complications of patients who received at least 1 packed red blood cell transfusion were compared with those who did not. For descriptive purposes, categorical data are presented as frequencies, and groups were compared by using chi-square tests. Continuous variables are reported as mean ± SD, and differences were compared by using Student t tests. We used the Wilcoxon rank sum test to compare variables with skewed distributions, and results are reported as median and IQR.
To assess the association between transfusion status and in-hospital mortality, we first used a non-parsimonious propensity model to calculate the likelihood of transfusion based on all available patient characteristics for each patient, including 45 variables (Table 1). Importantly, these variables included multiple in-hospital hemoglobin assessments, key treatments, and complications during hospitalization, including acute renal failure, in-hospital shock, septic shock, cardiogenic shock, and in-hospital mechanical ventilation. After calculating propensity scores for the likelihood of transfusion for each patient, propensity scores of patients with and without transfusion during AMI hospitalization were examined for overlap. Comparable patients in the overlapping region of the propensity to be transfused were then stratified according to transfusion status, and those who were transfused were matched in a 1-to-many fashion with those not transfused (because a larger proportion of AMI patients in this cohort did not receive a transfusion), using a caliper width of 0.5. This caliper width is 0.5 times the SD of the logit of the propensity score for transfusion and reflects the maximum difference in propensity score between treated and untreated patients that would still allow matching. The adequacy of propensity matching was then assessed by calculating post-match standardized differences and examining patient characteristics post-matching. A significant imbalance was considered to be present if a >10% standardized difference was present between the 2 groups after propensity matching (15).
Another clinically important question is whether the relationship between transfusion and mortality differs depending on nadir hemoglobin. Because propensity matching incorporates the use of nadir hemoglobin, it eliminates the ability to study this interaction and also limits sample size, precluding stratification on nadir hemoglobin. Accordingly, we conducted separate analyses in the total patient population (N = 34,937) to assess the interaction between nadir hemoglobin level and transfusion. Patients were stratified according to nadir hemoglobin level (<7 g/dl, 7 to 8.9 g/dl, 9 to 10.9 g/dl, and ≥11 g/dl), and we compared the in-hospital mortality of patients with and without transfusion within strata of nadir hemoglobin. This analysis was then repeated by using hierarchical multivariable logistic regression to account for clustering within hospital site and adjusted for confounding related to patients’ clinical characteristics.
To examine variability in transfusion practices across hospitals while generating conservative estimates of variation, we excluded patients from any hospital reporting no blood transfusions and hospitals in which the absolute observed transfusion rate varied from the expected transfusion rate by >10%. When comparing transfusion rates across hospitals, shrinkage estimates were generated by using a hierarchical model including site as a random effect (with no additional covariates) to account for lower enrollment from small hospitals. This approach pulls estimates from smaller hospitals toward the overall mean because the crude transfusion rates at sites with small enrollments are unduly influenced by only a few transfusion events. Estimates were generated by using a generalized linear model, regressing site as a random effect on transfusion rate as the dependent variable. To adjust for case-mix differences, median odds ratios (ORs) were then calculated to assess the variability in transfusion rates independent of patient characteristics (16). The median OR reflects the median value of the ORs for the risk of transfusion if 2 patients with identical characteristics presented to all possible pairs of Health Facts hospitals. Statistical analyses were conducted by using SAS version 9.2 (SAS Institute, Inc., Cary, North Carolina).
At least 1 packed red blood cell transfusion was administered to 1,778 patients (5.1%). Patients who were transfused differed markedly from those who were not with respect to a host of prognostically important variables (Table 1). They were older, had much lower hemoglobin values throughout their hospital course, a greater burden of in-hospital complications, and a significantly longer length of hospitalization compared with nontransfused patients. Transfused patients were much less likely to undergo angiography or percutaneous coronary intervention and to receive medical therapies such as aspirin, thienopyridines, angiotensin-converting enzyme inhibitors, or angiotensin receptor blockers.
Patients who received a blood transfusion had significantly higher mortality than those who did not (11.0% vs. 5.7%; unadjusted OR: 2.05 [95% confidence interval (CI): 1.76 to 2.40]; p < 0.001). However, there was substantial baseline imbalance in key clinical characteristics, as evidenced by a plot of propensity scores (Central Illustration). In fact, 31,829 (91.1%) patients had nonoverlapping propensity scores for blood transfusion and could not be matched based on propensity for transfusion.
After propensity matching the remaining population (n = 3,108 [n = 1,121 transfused and n = 1,987 not transfused]), there were no longer significant imbalances in clinical characteristics, with all post-matching standardized differences being <10% (Fig. 2). In contrast to the unadjusted analyses, after adequate propensity matching, blood transfusion was associated with a lower in-hospital mortality (OR: 0.73 [95% CI: 0.58 to 0.92]).
In additional analyses, a marked variation was noted in the association between transfusion and mortality stratified according to nadir hemoglobin (Fig. 3⇓). In unadjusted analyses, transfusion was associated with lower risk of mortality among those with nadir hemoglobin values <7 g/dl (OR: 0.52 [95% CI: 0.32 to 0.84]) and between 7 and 8.99 g/dl (OR: 0.73 [95% CI: 0.6 to 0.9]). There was no significant relationship of transfusion with mortality among those with nadir hemoglobin values between 9 and 10.99 g/dl, whereas transfusion was associated with higher mortality among those with nadir hemoglobin values ≥11 g/dl (OR: 6.28 [95% CI: 2.12 to 18.6]). After adjustment for site and patient characteristics, the trend of lower mortality among transfused patients below a nadir hemoglobin of 9 g/dl persisted, but the apparent hazard of transfusion for patients with nadir hemoglobin ≥11 g/dl was attenuated (OR: 1.88 [95% CI: 0.40 to 8.78]).
There was substantial variability in the rate of transfusion across hospitals (Fig. 4). After applying additional exclusions as described in the Methods, 1,175 (7.4%) of the remaining 24,083 patients received at least 1 red blood cell transfusion. Shrinkage-adjusted transfusion rates ranged from 3.1% to 14.5%. The median OR for transfusion was 2.0 (95% CI: 1.5 to 2.5), indicating a 2-fold variability in blood transfusion rates across hospitals for 2 randomly selected patients with identical clinical characteristics.
In this large, multicenter study of AMI patients reflecting real-world clinical practice, we found that patients who received blood transfusions had higher in-hospital mortality than those who were not transfused but that this finding largely reflects their higher risk clinical characteristics. After propensity matching, the relationship between transfusion and higher mortality was eliminated (Central Illustration). We also identified marked variability in case-mix–adjusted transfusion rates across hospitals, likely indicating substantial variation in hospital- and provider-specific blood transfusion practices.
The only randomized trial to assess the impact of transfusion on outcomes at the time of AMI was CRIT (Conservative versus Liberal Red Cell Transfusion in Acute Myocardial Infarction), a small pilot study that found an excess in the composite outcome of death, recurrent myocardial infarction (MI), or heart failure when patients were transfused below a hematocrit level of 30% compared with a more conservative threshold of 24% (17). This finding was driven entirely by the development of heart failure, and the authors observed no difference in the combined endpoint of mortality or recurrent AMI. A host of observational studies have been conducted and have reached varying conclusions. Wu et al. (7) found that transfusion was associated with lower 30-day mortality among elderly patients with MI below a hematocrit threshold of 33%. In contrast, several studies reported increased mortality among patients treated with blood transfusion (9–11). A recent meta-analysis by Chatterjee et al. (10) reported that blood transfusion was associated with a nearly 3-fold increase in mortality. However, many studies included in this meta-analysis had limitations related to the assessment of patient-level characteristics (presence and severity of comorbidities, inclusion of multiple hemoglobin assessments), as well as in the robustness of statistical methods used.
Several previous studies have used propensity score adjustment, which is a less robust tool to minimize confounding compared with propensity matching (12). Our finding that the vast majority of patients in Health Facts had nonoverlapping propensity scores strongly suggests that the use of multivariable regression with or without propensity score adjustment may not be adequate to address the considerable selection bias inherent in comparing the outcomes of transfused and nontransfused patients. Both the large size and detailed patient-level data available in the present study allowed exclusion of patients with nonoverlapping propensity scores, while still preserving the sample size necessary to find a clinically important difference in mortality; this was likely not possible in smaller previous studies. One previous study did match patients on propensity to be transfused; in contrast to our study, the authors found that transfusion was associated with increased mortality (11). However, they did not include nadir hemoglobin in propensity matching, choosing instead to adjust for nadir hemoglobin after the match was completed. In our study, there was a >200% standardized difference in nadir hemoglobin between those who were and were not transfused; given the strong prognostic implications of severe in-hospital anemia, not including nadir hemoglobin in propensity matching may erroneously attribute the adverse prognostic impact of anemia to blood transfusion. Moreover, the propensity model in this study did not include key in-hospital complications, such as acute renal failure, mechanical ventilation, shock, and requirement for hemodialysis. These variables differed greatly between transfused and nontransfused patients in our study (each with a standardized difference >10% before matching), and given the strong prognostic value of these factors, their exclusion could produce a strong confounding effect.
Previous randomized studies examining other patient populations found similar outcomes between conservative and liberal transfusion strategies. Examining critically ill patients randomized to a transfusion threshold of 7 g/dl versus 10 g/dl, Hebert et al. (18) found no significant difference in 30-day mortality. Similarly, there was no difference in the composite of 30-day mortality or morbidity comparing a liberal versus a conservative transfusion strategy among cardiac surgery patients in the TRACS (Transfusion Requirement After Cardiac Surgery) trial (19). Similar results were observed in a trial of patients with coronary disease or coronary disease risk factors undergoing hip surgery (20). A recent meta-analysis summarizing trials of transfusion to date across a variety of populations found a modestly lower in-hospital mortality among patients managed with a restrictive versus more liberal transfusion goal (relative risk: 0.77 [95% CI: 0.62 to 0.95]) but no difference in 30-day mortality (relative risk: 0.85 [95% CI: 0.70 to 1.03]) (21,22). In contrast, a recent trial including patients with upper gastrointestinal bleeding found lower mortality rates among patients assigned to a restrictive transfusion strategy (23). Whether these findings generalize to patients presenting with MI is unclear, and practice guidelines acknowledge this lack of certainty with regard to transfusion practices in AMI (24). Regardless, these data, coupled with our findings, argue against a 2-fold increase in mortality attributable to blood transfusion in AMI as suggested by previous observational literature.
The optimal threshold for blood transfusion during AMI remains a subject of debate. We observed a trend toward improved outcomes associated with transfusion below a hemoglobin threshold of 9 g/dl, which was not seen at higher nadir hemoglobin values. This finding is consistent with previous analyses that have suggested a possible benefit from transfusion for patients with lower nadir hemoglobin (1,7,8). Our results are also consistent with data from a small pilot study of liberal (transfusion threshold of 10 mg/dl) versus conservative transfusion in patients with symptomatic coronary artery disease (25). The authors found a lower rate of death and a trend toward a reduced composite of death, MI, or unscheduled revascularization among patients randomized to the liberal transfusion strategy.
Until additional data from randomized trials are available to guide clinical practice, it seems reasonable to consider transfusion during AMI below “conservative thresholds.” However, these data underscore significant uncertainty in the benefits and risks of transfusion, necessitating clinicians’ careful consideration of individual patient factors, which may influence the decision to provide a blood transfusion. We also observed dramatic variation in the frequency of blood transfusion across hospitals. Importantly, a 2-fold variability remained even after extensive adjustment for patient characteristics, indicating that hospital and provider factors likely drive substantial variability in provision of blood transfusion. This variability likely reflects clinical uncertainty regarding benefits and risks of transfusion during AMI and represents an important target for future research. Given the absence of adequately powered trials to describe the relationship between transfusion and outcomes in patients with AMI, our findings also suggest that it is premature to consider using transfusion rates as a quality metric.
Our results should be interpreted in the context of the following limitations. First, the number of units of blood administered with each transfusion event was not recorded. The number of units transfused has been shown to correlate with outcomes (19), likely reflecting that lower nadir hemoglobin portends a poorer prognosis. Our definition of AMI was based on the combination ICD-9 codes and elevated cardiac biomarkers, and it could have included some patients with a type II non–ST-segment elevation MI. However, type II MI has been shown to be prognostically important (26), and the potential benefit of improved oxygen delivery may be even more important in patients whose AMI is driven primarily by demand ischemia. Finally, although propensity matching is one of the most robust methods to address observed confounding in nonrandomized comparisons, the possibility of residual and unmeasured confounding cannot be excluded. Selection bias cannot be completely eliminated from observational studies examining interventions, such as blood transfusion, for which physicians have strong treatment preferences. Moreover, use of this method, which required exclusion of the majority of patients in Health Facts from the primary analysis, could limit generalizability of our findings.
We found that the majority of patients who received a blood transfusion during AMI hospitalization were not comparable to those who were not transfused, due to marked differences in multiple prognostically important characteristics. Although blood transfusion was initially associated with increased mortality, this association was eliminated by propensity matching, suggesting that this apparent relationship between transfusion and mortality is largely driven by selection bias. Randomized trials are needed to determine appropriate blood transfusion thresholds in patients with AMI.
COMPETENCY IN MEDICAL KNOWLEDGE: Severe anemia is common in patients hospitalized with AMI, but blood transfusion is controversial, given the lack of evidence of a favorable impact on patient outcomes.
COMPETENCY IN PATIENT CARE 1: In a large cohort of consecutive propensity-matched patients with AMI, blood transfusions were not associated with a greater risk of adverse outcomes.
COMPETENCY IN PATIENT CARE 2: In the absence of definitive data from randomized trials, clinicians should determine whether to transfuse blood to individual patients with AMI based on careful consideration of benefits and risks.
TRANSLATIONAL OUTLOOK: Randomized trials are necessary to overcome selection bias and to further investigate the relationship between blood transfusion and clinical outcomes in patients with AMI.
This project was supported by an Outcomes Research Center grant from the American Heart Associationhttp://dx.doi.org/10.13039/100000968. Cerner Corporation provided the data but had no role in funding, design, analyses, drafting, or review of the manuscript. Drs. Salisbury, Spertus, and Kosiborod are funded, in part, by an award from the American Heart Associationhttp://dx.doi.org/10.13039/100000968 Pharmaceutical Round Table and David and Stevie Spina. Dr. Marso has received research grants from The Medicines Company, Volcano Corporation, Amylin Pharmaceuticalshttp://dx.doi.org/10.13039/100004323, Novo Nordisk, Terumo Medical Corporation, and St. Jude Medicalhttp://dx.doi.org/10.13039/100006279. Dr. Amin has served as a consultant to Terumo Medical Corporation and The Medicines Company. Dr. Wang has received research grants from Bristol-Myers Squibbhttp://dx.doi.org/10.13039/100002491/Sanofi Pharmaceuticals, Daiichi Sankyo, Canyon Pharmaceuticals, Eli Lilly, Sanofi-Aventis, Schering Plough, Merck, and The Medicines Company; and has served as a consultant for Medco and AstraZeneca, The Medicines Company, Novo Nordisk, and Terumo Medical Corporation. Dr. Spertus has received research grants from the National Heart, Lung, and Blood Institutehttp://dx.doi.org/10.13039/100000050, American Heart Associationhttp://dx.doi.org/10.13039/100000968/Pharmaceutical Round Table Outcomes Centers, American College of Cardiology Foundation, Johnson & Johnson, Amgenhttp://dx.doi.org/10.13039/100002429, Eli Lilly, Evaheart, and Sanofi-Aventis; has received other research support from Rochehttp://dx.doi.org/10.13039/100004337 and Atherotech; and has served as a consultant on the advisory board for St. Jude Medical, United Healthcare, and Novartis. Dr. Kosiborod has received research grants from the American Heart Association (11GRNT7330005)http://dx.doi.org/10.13039/100000968, Gilead Sciences (IN-US-259-0159)http://dx.doi.org/10.13039/100005564 (with Dr. Spertus), Genentechhttp://dx.doi.org/10.13039/100004328, Sanofi-Aventis, Medtronic Diabetes, Glumetrics, Maquet, and Eisaihttp://dx.doi.org/10.13039/501100003769; and has served as a consultant or on the advisory board of Gilead Sciences, Genentech, Hoffmann-La Roche, Medtronic Diabetes, AstraZeneca, AbbVie, Regeneron, Edwards Lifesciences, ZS Pharma, and Eli Lilly. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
- Abbreviations and Acronyms
- acute myocardial infarction
- confidence interval
- International Classification of Diseases-Ninth Revision-Clinical Modification
- interquartile range
- myocardial infarction
- odds ratio
- Received January 6, 2014.
- Revision received April 10, 2014.
- Accepted May 1, 2014.
- American College of Cardiology Foundation
- Sabatine M.S.,
- Morrow D.A.,
- Giugliano R.P.,
- et al.
- Anker S.D.,
- Voors A.,
- Okonko D.,
- et al.
- Salisbury A.C.,
- Amin A.P.,
- Reid K.J.,
- et al.
- Salisbury A.C.,
- Alexander K.P.,
- Reid K.J.,
- et al.
- Shander A.,
- Javidroozi M.,
- Ozawa S.,
- Hare G.M.
- Shishehbor M.H.,
- Madhwal S.,
- Rajagopal V.,
- et al.
- Austin P.C.
- Bonaca M.P.,
- Wiviott S.D.,
- Braunwald E.,
- et al.