Author + information
- Received May 4, 2015
- Revision received September 10, 2015
- Accepted September 15, 2015
- Published online December 8, 2015.
- Yew Y. Ding, MBBS, MPH∗,†∗∗ (, )
- Boris Kader, PhD‡,§,
- Cindy L. Christiansen, PhD‡,§ and
- Dan R. Berlowitz, MD, MPH‡,§∗ ()
- ∗Department of Geriatric Medicine and Institute of Geriatrics and Active Ageing, Tan Tock Seng Hospital, Singapore
- †Health Services and Outcomes Research, National Healthcare Group, Singapore
- ‡Center for Healthcare Organization and Implementation Research (CHOIR), Bedford, Massachusetts
- §Boston University School of Public Health, Boston, Massachusetts
- ↵∗Reprint requests and correspondence:
Dr. Dan R. Berlowitz, VA Bedford Hospital, Center for Healthcare Organization and Implementation Research (CHOIR), Edith Nourse Rogers Memorial Hospital, (152), Building 70, 200 Springs Road, Bedford, Massachusetts 01730.
- ↵∗∗Dr. Yew Y. Ding, Department of Geriatric Medicine & Institute of Geriatrics and Active Ageing, Tan Tock Seng Hospital, 11 Jalan Tan Tock Seng, Singapore 308433.
Background There is a paucity of randomized clinical trial data on the use of red blood cell (RBC) transfusion in critically ill patients, specifically in the setting of cardiac disease.
Objectives This study examined how hemoglobin (Hgb) level and cardiac disease modify the relationship of RBC transfusion with hospital mortality. The aim was to estimate the Hgb level threshold below which transfusion would be associated with reduced hospital mortality.
Methods We performed secondary data analyses of Veterans Affairs intensive care unit (ICU) episodes across 5 years. Logistic regression quantified the effect of transfusion on hospital mortality while adjusting for nadir Hgb level, demographic characteristics, admission information, comorbid conditions, and ICU admission diagnoses.
Results Among 258,826 ICU episodes, 12.4% involved transfusions. Hospital death occurred in 11.6%. Without comorbid heart disease, transfusion was associated with decreased adjusted hospital mortality when Hgb was approximately <7.7 g/dl, but transfusion increased mortality above this Hgb level. Corresponding Hgb level thresholds were approximately 8.7 g/dl when comorbid heart disease was present and approximately 10 g/dl when the ICU admission diagnosis was acute myocardial infarction (AMI). Sensitivity analysis using additional adjustment for selected blood tests in a subgroup of 182,792 ICU episodes lowered these thresholds by approximately 1 g/dl.
Conclusions Transfusion of critically ill patients was associated with reduced hospital mortality when Hgb level was <8 to 9 g/dl in the presence of comorbid heart disease. This Hgb level threshold for transfusion was 9 to 10 g/dl when AMI was the ICU admission diagnosis.
Hemoglobin (Hgb) level thresholds drive red blood cell (RBC) transfusion practices in critically ill patients, with additional influence exerted by other patient factors, such as increased age, severity of illness, gastrointestinal hemorrhage, comorbid heart disease, and acute myocardial infarction (AMI) (1–7). Over the past 2 decades, research has led to better understanding of the extent to which these transfusion thresholds influence mortality.
In the landmark TRICC (Transfusion Requirements in Critical Care) trial, patients randomly assigned to a restrictive strategy of receiving transfusions only when Hgb levels fell to <7 g/dl had significantly lower hospital mortality compared with a liberal strategy of transfusing when Hgb was <10 g/dl (2). Subsequent observational studies of intensive care unit (ICU) patients yielded mixed results, with some demonstrating increased mortality with transfusion, whereas others showed the opposite (4,5,8–12). Subgroups that derived greater benefit from transfusions included those with cardiac disease, such as those with severe ischemic heart disease (13) and AMI (14). These findings were by no means consistent across studies, with some showing benefit only when Hgb was <8 g/dl (12,15–18). In contrast, a recent, pilot randomized controlled trial (RCT) documented reduced mortality with a liberal transfusion strategy in the setting of acute coronary syndrome (ACS) or stable angina (19). For congestive heart failure (CHF), the impact of transfusion is even more uncertain (20). Notably, the American Association of Blood Banks clinical practice guidelines recommended a restrictive transfusion strategy for hospitalized patients with pre-existing cardiovascular disease for an Hgb level ≤8 g/dl, but these guidelines did not recommend a restrictive or liberal transfusion strategy for hemodynamically stable patients with ACS. No specific mention of a suggested strategy for patients with CHF was made (21).
More recently, an elegant observational study of approximately 35,000 AMI hospitalizations from 57 centers used propensity score analysis to identify a subset of patients who were well-matched on 45 variables. In this subset, transfusion was associated with a 25% reduction in the odds of hospital mortality (22). However, the accompanying editorial cautioned, that despite several observational studies on transfusion in patients with significant cardiac disease, there is still uncertainty on when to transfuse these patients (23). Nonetheless, there is arguably accumulating evidence that suggests that a restrictive transfusion strategy may not be optimal in this setting (13,14,19,22). Although another recent editorial echoed the sentiment that evidence on appropriate transfusion thresholds for patients with ACS is weak, the possibility was entertained that such patients may yet benefit from transfusion at higher Hgb levels (e.g., 9 to 10 g/dl) (24).
Due to this ongoing debate, we conducted additional analyses of a previously developed dataset to determine how Hgb level modifies the treatment effect of RBC transfusion during ICU admissions with respect to hospital mortality in the context of cardiac disease. Specifically, we sought to estimate the Hgb level below which transfusion is associated with reduced hospital mortality (or transfusion threshold) across key cardiac conditions, such as comorbid heart disease, AMI, unstable angina, and CHF. We hypothesized that transfusion thresholds with these conditions are higher than those advocated in the restrictive transfusion strategy. Ultimately, our goal was to assemble additional evidence to assist physicians in deciding when to transfuse critically ill patients with cardiac disease while we await the results of definitive clinical trials.
We performed secondary analyses of Veterans Affairs (VA) electronic databases at the Center for Healthcare Organization and Implementation Research. Ethical approval was obtained from the VA Bedford Institutional Review Board. Individual-level data on VA ICU admissions during fiscal years 2001 through 2005 were extracted from national-level VA databases. For hospitals with separate surgical and neurological ICUs, we only selected medical ICU episodes. We excluded operative cases with surgical admitting diagnoses. Only the first ICU episode of each year for unique patients was included.
Hospital death was the outcome of interest, whereas RBC transfusion during the first 30 days of ICU admission was the treatment of interest. Transfusion receipt was defined as having an International Classification of Diseases-9th Revision-Clinical Modification (ICD-9-CM) procedure code of 99.04 (transfusion of packed cells) or 99.03 (transfusion of whole blood) in the ICU file. ICU admissions with at least 1 transfusion documented during the first 30 days were collectively assigned as the “transfusion” group; all others were included in the “no transfusion” group. To exclude hospitals that were possible outliers in terms of transfusion practices, we arbitrarily removed data from 8 of 120 hospitals where transfusion rates were >2 SD from the mean.
To specify the nadir Hgb level, we selected the lowest value during ICU admission before blood transfusion. If not transfused, we selected the lowest value during the first 30 days of ICU admission. If none was available, we used the lowest value during hospital admission before ICU admission.
Other explanatory variables included demographic characteristics, admission-related information, comorbid conditions (e.g., Elixhauser comorbidity measures , Acute Physiology and Chronic Health Evaluation III chronic health parameters , and others [27,28]), ICU admission diagnoses categories (adapted from the Healthcare Cost and Utilization Project) (29), and selected blood test values. Comorbid conditions were assigned if their codes were stated at least once in administrative records during the previous 2 years, but not where there was an overlapping admission diagnosis (27). ICU admission diagnoses were defined by the first ICD-9-CM code from ICU files. Chronic kidney disease was defined as a glomerular filtration rate <60 ml/min/1.73 m2 for at least 3 months (30).
We used principal component analyses to create a smaller number of comorbidity groupings. This is a data reduction method that addresses correlation between a set of observed variables, and it develops a shorter list of artificial variables or “principal components” that can be used in further analyses. Details of the principal components and their corresponding comorbid conditions are provided in Online Table 1. Logistic regression was then performed to build explanatory models for hospital mortality. Blood transfusion during the first 30 days of ICU stay was the treatment variable. Other aforementioned explanatory variables, including Hgb, were included. To examine the effect modification of Hgb, comorbid heart disease, and ICU admission for AMI, unstable angina, and CHF on the relationship between transfusion and hospital mortality, we included corresponding interaction variables. To address clustering of admissions within fiscal years, we employed fixed effects for year as dummy variables. To understand how absolute hospital mortality rates with and without transfusion varied with Hgb, we computed model-predicted mortality rates at different Hgb values from 6 to 11 g/dl. Finally, we conducted sensitivity analyses by including selected blood tests, namely blood urea nitrogen, blood total white blood cell, serum albumin, and serum creatinine as explanatory variables for a subset of ICU episodes that had complete data on these investigations. SAS version 9.3 (SAS Institute, Cary, North Carolina) was used for all data analyses.
Among 258,826 ICU episodes, hospital death occurred in 30,086 patients (11.6%). Transfusion was documented in 32,097 (12.4%) episodes during the first 30 days of ICU admission. Patient characteristics for these episodes are shown in Table 1. For 182,792 ICU episodes with complete information on selected blood tests, hospital death occurred in 13%, with transfusion received by 14% during the first 30 days of ICU stay. There were no major differences from the study population as a whole (Online Table 2). As expected, patients who died in hospital were older, had lower Hgb, were more likely to have heart disease and most other comorbidities, and stayed longer in the ICU. Notably, those who died were almost twice as likely to have received transfusion during the first 30 days of ICU admission.
Details of logistic regression analyses for hospital mortality are provided in Online Table 3. The most important results are those that show the interactions of blood transfusion with Hgb and cardiac disease. For the whole group, the odds ratio (OR) for the interaction between blood transfusion and Hgb was 1.22 (95% confidence interval [CI]: 1.20 to 1.25). This indicates that for each 1 g/dl increase in Hgb, transfusion was associated with a 22% increase in odds of hospital mortality. Lower hospital mortality was associated with transfusion at lower Hgb levels. Corresponding ORs for interactions with comorbid heart disease and ICU admission diagnosis of AMI were 0.82 (95% CI: 0.76 to 0.88) and 0.78 (95% CI: 0.69 to 0.88), respectively. The interpretation was that transfusion was associated with lower hospital mortality when these 2 conditions were present. In other words, for any given Hgb, the benefit associated with transfusion was greater when patients had comorbid heart disease or were admitted to the ICU for AMI. The interactions for transfusion with ICU admission for unstable angina and CHF were not significant. This suggests no differences in benefit with transfusion whether or not these 2 conditions were present. The model c-statistic of 0.81 indicates good model discrimination.
For model-predicted hospital mortality rates with and without transfusion at different Hgb levels from 6 to 11 g/dl for 3 key permutations of comorbidity and ICU admitting diagnosis, Figures 1A to 1C (unbroken blue and orange lines) represent point estimates of predicted probability of hospital mortality with and without transfusion, respectively. The corresponding broken lines represent the 95% CIs for those point estimates. Confidence bands around these point estimates for mortality range from approximately 0.02 to 0.06. As expected, mortality decreased with higher Hgb for both transfused and nontransfused patients. However, this decrease was predicted to be steeper for those who were not transfused. In patients without comorbid heart disease and who were admitted to the ICU for a set of reference noncardiac diagnoses, mortality was lower with transfusion compared with no transfusion below an Hgb of approximately 7.7 g/dl, but mortality was higher above that cutpoint. We can infer that the critical point at which the risk–benefit balance of transfusion reversed was at an Hgb in the region of 7.7 g/dl. However, for those with comorbid heart disease admitted to the ICU for these noncardiac diagnoses, mortality was lower with transfusions below an Hgb of approximately 8.7 g/dl, but mortality was higher above this level. This indicates that the corresponding critical point was in this region of Hgb level. Finally, for those with comorbid heart disease who were admitted to the ICU for AMI, mortality was lower with transfusion below an Hgb of approximately 10 g/dl, but was higher above this level, which was the critical point for reversal of the risk–benefit balance.
Beyond obtaining point estimates of the transfusion threshold, having a sense of the uncertainty of these estimates would be helpful. The broken line representing the lower limit of the 95% CI for the transfusion group (blue line) intersects with that of the corresponding upper limit for the no transfusion group (orange lines) at Hgb levels of approximately 0.7 to 0.9 g/dl above the transition point across all graphs. Correspondingly, the broken line representing the upper limit of the 95% CI for the transfusion group (blue lines) and that of the corresponding lower limit for the no transfusion group (orange lines) intersect at Hgb levels approximately 0.7 to 0.9 g/dl below the transition point. These values represent worst-case scenarios with respect to uncertainty around our estimates of the transition point. In other words, the true transition point is very likely within the range of 0.7 to 0.9 g/dl below and above our point estimates.
Sensitivity analyses using additional blood test information on a smaller subset of patients yielded interaction effects in the same direction. However, the interaction of transfusion with ICU admission for AMI was not significant (third column of Online Table 3). Most importantly, the corresponding transfusion thresholds were approximately 1 g/dl lower than those for the whole group, as seen in Online Figure 1. We also conducted sensitivity analyses with robust standard error estimation to account for clustering by facility. However, this made no difference in the results of estimated effects or statistical significance. Finally, we repeated the regression models, but with Hgb specified as 1 g/dl categories rather than as a continuous variable for both the main effect and its interaction with transfusion receipt. The parameter estimates on the logit of hospital mortality varied in an approximately linear fashion across Hgb categories, thereby providing justification for specifying Hgb as a continuous variable. Detailed results are provided in Online Tables 4 and 5, as well as Online Figure 2.
This is the largest observational study of critically ill patients to examine the association of blood transfusions with mortality and effect modification by Hgb level and cardiac disease. Our premise is that Hgb levels exist below which the transfusion risk–benefit balance reverses (31). Combining our primary results and those of sensitivity analyses, we infer that the Hgb level below which transfusion was associated with reduced hospital mortality or transfusion threshold is approximately 7 to 8 g/dl across a heterogeneous population of VA ICU patients. Interestingly, these results approximate those of TRICC and are consistent with a Cochrane Review based on a meta-analysis of 19 RCTs that confirmed the association of restrictive transfusion strategies with reduced hospital mortality (32). When we consider the results of TRICC as offering true estimates of the treatment effect of transfusions, then the convergence of findings confers added confidence with respect to our extended analyses of subgroups with cardiac disease.
It was recognized quite early that critically ill patients with cardiovascular disease may require a different transfusion strategy (33–35). Subsequent analyses of these patients in TRICC revealed that among those with severe ischemic heart disease, the restrictive transfusion group had nonsignificantly higher mortality than the liberal transfusion group (13). In a retrospective study of elderly Medicare patients with AMI, transfusion was associated with significantly lower 30-day mortality if the hematocrit at admission was ≤33% (Hgb 11 g/dl) (14). In a recent pilot RCT of patients with ACS or stable angina who underwent cardiac catheterization, 30-day mortality was lower for the liberal transfusion group compared with the restrictive transfusion groups (19,36). Together, these findings suggest that patients with cardiac disease could benefit from a more liberal transfusion strategy.
In contrast, other studies suggest that a restrictive transfusion strategy might be beneficial. These include a higher predicted 30-day mortality with transfusion when the nadir Hgb level was >8 g/dl or there was an equivalent hematocrit value in the setting of ACS (15,16,37), and higher hospital mortality with transfusion when Hgb was >7 g/dl in the presence of cardiovascular comorbidities (12). The reasons for contrasting effects on mortality in different studies are probably methodological, including different trial designs and selection issues with observational studies and RCTs where participants may not be truly representative of the population of interest. Inadequate statistical handling may occur, too, where transfusion operates as a determinant of predictors of outcomes when not randomized (38).
In our study, having comorbid heart disease raised the transfusion threshold by approximately 1 g/dl to approximately 8 to 9 g/dl (Central Illustration). In other words, the mortality benefit associated with transfusion was accrued at slightly higher Hgb levels in the presence of heart disease. Here, our findings are consistent with the recommendations of the AABB (American Association of Blood Banks). Beyond that, ICU admission for AMI raises this threshold by a further 1 g/dl to approximately 9 to 10 g/dl (Central Illustration). Together, our results provide additional evidence, albeit observational in nature, to support transfusing critically ill patients with these cardiac conditions at higher Hgb levels. It is plausible that increased sensitivity to the adverse consequences of reduced oxygen delivery with anemia exists in the presence of cardiac disease (23). Anemia may exacerbate myocardial ischemia and activate the sympathetic nervous system (39,40). Furthermore, the ability to increase cardiac output to compensate for reduced oxygen delivery with anemia may be reduced with cardiac disease (41). The impact of these effects may vary across the spectrum of cardiac disease, with AMI being at its most severe end. This may explain why transfusions are associated with reduced mortality at higher Hgb levels as we traverse across the continuum from no cardiac disease to comorbid heart disease, and finally to AMI. Then again, our findings do not indicate that these levels should be modified for unstable angina or CHF. However, a national heart failure survey found a nonsignificant trend toward lower 30-day mortality with transfusion among propensity score-matched patients (20).
Bedside decisions on transfusion attempt to balance achieving greater oxygen carriage for myocardial support and minimizing risk of transfusion-related adverse outcomes, including acute lung injury (42–47) and secondary bacterial infection (48). It is plausible that specific comorbid conditions and acute illnesses could shift this balance toward transfusing more liberally. In this respect, we have demonstrated the varying effect of transfusion on hospital mortality across a range of Hgb levels and the effect modification by cardiac conditions. The situation becomes more complicated with multiple concurrent illnesses as commonly encountered in critically ill patients. Suggested transfusion thresholds, at best, serve as a general guide, and physicians should look beyond them for more complex patients with advanced age, multimorbidities, or frailty in addition to their cardiac illnesses. For them, transfusion decisions are best guided by considered clinical judgment at the bedside.
First, as emphasized, this was an observational study. Selection bias from unaccounted factors influencing transfusion receipt may exist despite our best efforts to include a wide range of covariates in our risk-adjustment model. Particularly, we did not have data on some aspects of illness severity, functional status, and do-not-resuscitate orders. As shown, sensitivity analyses using additional test information suggest that the Hgb level at which the risk–benefit associated with transfusion reverses could be approximately 1 g/dl lower. With this information, it is quite likely that the true transfusion threshold lies between these 2 sets of estimates. In line with this possibility, we adjusted our study conclusions accordingly. Second, we did not have VA data on the accuracy of ICD-9-CM codes for transfusion. External evidence on validity of the billing for blood transfusion in a non-VA tertiary care setting hospital indicated favorable sensitivity and specificity for transfusion identification from blood bank records using relevant ICD-9-CM procedure codes (49). Third, only approximately 3% of our patients were women. Although there were >7,000 female ICU admissions, and sex adjustment was made in regression analyses, we remained cautious on extrapolating our results to female patients. As such, our findings relate largely to critically ill male patients who have cardiac disease. Finally, data currency is a potential concern because we used data from 2001 to 2005. However, it is debatable as to whether transfusion and ICU practice have changed enough over the past decade to expect important differences in our results if we had used more recent VA data. Although possible, we are of the opinion that such differences are less likely to be operant. Despite these limitations, we believe that this study contributes to the ongoing debate on when critically ill patients with anemia and cardiac disease should be transfused. Building on previous work such as the recent study on transfusion in the setting of AMI by Salisbury et al. (22), our findings suggest that the restrictive transfusion strategy may not be optimal for patients with selected cardiac conditions. Rather, they support setting transfusion thresholds at Hgb levels 1 to 2 g/dl higher when comorbid heart disease or AMI occur, but not when unstable angina or CHF do.
RBC transfusion during ICU admission is associated with reduced hospital mortality when pre-transfusion Hgb levels were <8 to 9 g/dl in the setting of comorbid heart disease. The transfusion threshold is raised to 9 to 10 g/dl if the ICU admission is for AMI, but not for patients with unstable angina or CHF. Although we are waiting for more definitive evidence from randomized clinical trials, these findings offer possible interim guidance for physicians on transfusing ICU patients with cardiac disease.
COMPETENCY IN PATIENT CARE AND PROCEDURAL SKILLS: Although transfusion decisions should be individualized based on clinical assessment, critically ill patients with comorbid heart disease may benefit from Hgb levels >8 to 9 g/dl, and patients with ACS may need levels >9 to 10 g/dl.
TRANSLATIONAL OUTLOOK: Randomized trials are necessary to determine the Hgb thresholds for transfusion associated with survival benefit in critically ill patients with heart disease, including those with acute coronary syndromes or decompensated heart failure.
The Department of Veterans Affairs provided space and computer support.
For supplemental tables and figures, please see the online version of this article.
Dr. Ding was funded by the National Medical Research Council of Singapore for his research fellowship in Boston.
The authors have reported that they have no relationships relevant to the contents of this paper to disclose.
- Abbreviations and Acronyms
- acute coronary syndrome
- acute myocardial infarction
- congestive heart failure
- intensive care unit
- red blood cell
- randomized controlled trial
- Veterans Affairs
- Received May 4, 2015.
- Revision received September 10, 2015.
- Accepted September 15, 2015.
- American College of Cardiology Foundation
- Engoren M.,
- Arslanian-Engoren C.
- Alexander K.P.,
- Chen A.Y.,
- Wang T.Y.,
- et al.
- Shishehbor M.H.,
- Madhwal S.,
- Rajagopal V.,
- et al.
- Garty M.,
- Cohen E.,
- Zuchenko A.,
- et al.
- Salisbury A.C.,
- Reid K.J.,
- Marso S.P.,
- et al.
- Yeh R.W.,
- Wimmer N.J.
- ↵Healthcare Cost and Utilization Project (HCUP) Clinical Classifications Software (CCS) for ICD-9-CM. Rockville, MD: Agency for Healthcare Quality and Research. Available at: https://www.hcup-us.ahrq.gov/toolssoftware/ccs/ccs.jsp. Accessed May 2, 2015.
- Hill S.R.,
- Carless P.A.,
- Henry D.A.,
- et al.
- Alfonso F.
- Nikolsky E.,
- Aymong E.D.,
- Halkin A.,
- et al.
- van Stein D.,
- Beckers E.A.,
- Sintnicolaas K.,
- et al.
- Segal J.B.,
- Ness P.M.,
- Powe N.R.