Restrictive versus Liberal Transfusion Strategy in the Perioperative and Acute Care Settings: A Context-specific Systematic Review and Meta-analysis of Randomized Controlled Trials

DESPITE studies suggesting unfavorable outcomes after the administration of erythrocytes,^1–5  the optimal transfusion strategy in surgical and critically ill patients remains unclear. Concerns have been raised about harmful effects of low hemoglobin transfusion thresholds in individuals less able to tolerate anemia, such as the elderly and patients with cardiovascular disease or cancer.^6–9  Previously published meta-analyses were inconclusive: minimizing exposure to allogeneic blood reduced the risk of infection, but patients assigned to these restrictive transfusion strategies seemed also at a higher risk of myocardial infarction (MI).^10,11 

Since variability among studies is inevitable, undertaking meta-analyses generally entails some degree of heterogeneity, of which three different subtypes have been described¹²: (1) clinical heterogeneity, which results from variability in participants, interventions, or outcomes; (2) methodologic heterogeneity, a consequence of variability in study design and risk of bias; (3) statistical heterogeneity, which results from clinical or methodologic heterogeneity, or both. Identifying and addressing each type of heterogeneity remains a key step in undertaking meta-analyses. To date, however, most systematic reviews on transfusion strategies failed to address clinical heterogeneity, thereby limiting their interpretation.^13–16 

Therefore, we conducted a context-specific systematic review and meta-analysis of randomized controlled trials (RCTs) investigating the effects of restrictive transfusion strategies in the perioperative and acute care settings. The rationale for a contextual approach (i.e., stratification of the analysis according to patient characteristics and clinical settings) was based on the prespecified assumption that a high degree of clinical heterogeneity may hinder the identification of group-specific effects: pooling data from various patient populations (young and elderly patients, for instance) or from various settings (such as cardiac surgery and postpartum setting) may result in a dilution of the intervention effects. Since clinical diversity may also result from variability in study interventions, we were also interested in the effects of different transfusion protocols and in the contributing role of cointerventions (i.e., administration of non-erythrocyte blood products, hemostatic agents, or intravenous fluids) in complication rates.

We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.¹⁷  Eligibility criteria, outcomes, and methods of analysis were prespecified (study protocol available at: http://alfredanaesthesia.org.au/research).

Only fully published reports of RCTs were included. For duplicates or follow-up or ancillary studies, the first published article was considered the main study.¹⁸  Crossover designs or studies not adequately controlled were excluded. Trials evaluating a multiinterventional protocol were excluded if the effect of transfusion strategies could not be distinguished from the effect of other interventions. Cluster randomized trials were included only if methods of analysis allowed for clustering.¹⁹ 

Only trials conducted among adult patients (more than 18 yr old) in the perioperative, emergency, or intensive care settings were considered. We excluded studies conducted in patients with sickle cell disease.

We searched for studies comparing two different laboratory values (or using symptoms of anemia) to guide erythrocyte administration. We excluded studies applying transfusion strategies 24 h or less, trials using a hemodilution protocol, and interventions relying on preoperative autologous blood donation, since some effects (such as immunomodulation) were unlikely to develop,²⁰  and this approach is no longer recommended.²¹ 

We were interested in studies reporting events associated with, or worsened by, anemia.^5,22–28  To fully capture all the effects of transfusion strategies,^29,30  individual outcome events were combined into the following categories: “inadequate oxygen supply” (myocardial, cerebral, renal, mesenteric, and peripheral ischemic injury; arrhythmia; unstable angina), “mortality,” and a composite category “inadequate oxygen supply + mortality” (see description provided in Supplemental Digital Content 1, http://links.lww.com/ALN/B275, which describes outcome categories). Only events occurring within 30 days were retrieved because substantial hemoglobin recovery seems to occur within 2 months after surgery/intensive care unit stay.^31,32  Our aim was also to explore the immunomodulatory effects of allogeneic blood: since transfusions have been associated with impaired host defenses,^33,34  we searched for studies reporting new infections occurring within 30 days.

We performed a systematic electronic search in the MEDLINE (Ovid), EMBASE (Ovid), and Cochrane CENTRAL databases. Both MeSH terms and keywords combined with Boolean operators were used (see strategy provided in Supplemental Digital Content 2, http://links.lww.com/ALN/B276, which provides the full search strategy used in this systematic review). The following sources of grey literature were screened: OpenGrey, International Clinical Trials Registry Platform, ClinicalTrials.gov. Additional reports were identified by hand-searching bibliographies. No language or date restrictions were applied. The last electronic search was done on November 17, 2015.

Titles and abstracts were assessed for eligibility by two independent reviewers (Drs. Hovaguimian and Myles). Duplicate publications were identified through comparison of reports for author names, enrolment date, setting, intervention, participant number, or baseline data. Disagreements were resolved through discussion.

Data were extracted from original reports by one reviewer (Dr. Hovaguimian) and entered in a form specifically designed for this review (see description provided in Supplemental Digital Content 3, http://links.lww.com/ALN/B277, which details which information was extracted). The second reviewer (Dr. Myles) verified these data, and queries were resolved through discussion. Missing, unclear, or incomplete data in the original report were clarified by contacting authors (two provided additional data).^35,36  Outcome data were not considered for analysis if no clarification could be obtained. Data from duplicates were extracted and merged under a unique study identification name. Data were subsequently entered into the Cochrane Review Manager software (RevMan, version 5.3.3—The Cochrane Collaboration, The Nordic Cochrane Centre, Denmark, 2014) by one reviewer (Dr. Hovaguimian) and checked by the second reviewer (Dr. Myles).

The risk of bias was assessed using the Cochrane “Risk of bias” tool, which evaluates randomization method, concealment of treatment allocation, blinding of participants and personnel, blinding of outcome assessor, risk of incomplete outcome data, risk of selective reporting, and other sources of bias (ethics approval, informed consent, funding, and conflict of interest).³⁷  Each item was rated at “low,” “unclear,” or “high” risk of bias. The effects of detection and attrition bias were specifically explored, since this may affect studies evaluating adverse events (AEs).³⁸  For cluster randomized trials, we used specific items as recommended elsewhere.¹⁹  Disagreements were resolved through discussion.

Dichotomous outcomes were reported as risk ratios with 95% CIs, while continuous data were expressed as weighted mean differences with 95% CI. All statistical analyses were performed with the Cochrane Review Manager software. Data handling and transformation were necessary for some endpoints. Composite outcome categories were obtained by combining individual outcome data, as performed in previous reports (see description provided in Supplemental Digital Content 4, http://links.lww.com/ALN/B278, which outlines data handling, transformation, and combination).^10,13 

Differences between studies in terms of patient characteristics and/or clinical settings (i.e., clinical diversity) may affect effect estimates.¹²  To control for these sources of heterogeneity, we used a prespecified context-specific approach and stratified the analysis by (1) patient-specific risk of developing complications (according to age, comorbidities, and concomitant medication); (2) setting-related risk of complications (type of surgery). Studies conducted in similar populations and settings were regrouped into risk strata (see description provided in Supplemental Digital Content 5, http://links.lww.com/ALN/B279, which explains the rationale and methods used for strata generation). Meta-analyses were performed only if data were obtained from at least two studies.

To assess if our context-specific approach was sensible, we conducted for each outcome category a nonstratified analysis (i.e., data pooling without controlling for clinical diversity) and performed a test of interaction using the Cochran Q and Higgins I². We considered that different population parameters were represented within each risk stratum when the Cochran Q P-value was less than 0.05 or when I² was greater than 50%.^12,39  When data combination was deemed inappropriate, a qualitative assessment was performed.

Statistical heterogeneity was assessed by visual inspection of forest plots and by using the chi-square test and the I² statistic. When data were heterogeneous (P < 0.1, I² greater than 50%), we searched for methodologic sources of heterogeneity.¹²  We used a fixed-effect model, unless overt clinical or residual statistical heterogeneity was present (see fig., Supplemental Digital Content 6, http://links.lww.com/ALN/B280, which outlines how heterogeneity assessment was performed).^39,40 

We assumed that only studies demonstrating successful transfusion protocol implementation would reflect true intervention effects. Success was arbitrarily defined as a statistically significant difference between transfusion groups in two performance indicators: (1) hemoglobin levels over time and (2) mean erythrocyte units/group. P values were assessed from original reports. We also explored other possible determinants of success, such as hemoglobin thresholds, transfusion-sparing effect, and adherence rates (see description provided in Supplemental Digital Content 7, http://links.lww.com/ALN/B281, which outlines which indicators were used to assess successful protocol implementation).

Several routinely administered drugs or other management measures may participate in the occurrence of AE. Thus, we assessed the effects of nonerythrocyte blood products (cryoprecipitate, fresh frozen plasma, and platelets), antifibrinolytics, clotting factor concentrates, and fluids administration.

We identified 4,684 records from MEDLINE, 858 records from EMBASE, and 1741 records from CENTRAL (fig. 1). Other sources retrieved no additional records. Of these 7,283 records, 7,193 were excluded after preliminary screening. Of the 90 remaining reports, 53 were discarded because the study population was ineligible, the intervention was inappropriate, the design was problematic, or the study presented other issues (see description provided in Supplemental Digital Content 8, http://links.lww.com/ALN/B282, which lists excluded studies). Of 37 relevant reports, 6 were excluded after more thorough examination: 2 were duplicate publications, 1 was a preliminary analysis, and 3 were ancillary or follow-up studies of included studies.^41–46  Thus, we included 31 RCTs comparing restrictive with liberal transfusion strategies in the perioperative or acute care setting.^{6–9,35,36,47–71} 

Included studies were published between 1956 and 2015, and all were reported in the English language. A two-arm parallel design was used in all trials except in a factorial 2 × 2 design investigating transfusion strategies and erythropoietin administration.⁶⁵  One study was a cluster randomized trial.³⁵ 

Trials with similar patient populations and clinical settings were regrouped into five risk strata: (1) patients with cardiovascular disease undergoing cardiac or vascular procedures (surgery or catheterization—8 trials, 3,323 patients)^{6,47,48,51,55,60,66,67} ; (2) elderly population with varying cardiovascular disease undergoing orthopedic surgery (9 trials, 3,777 patients)^{8,9,49,50,52,54,62,63,68} ; (3) mixed surgical and medical patients with varying comorbidities admitted to an acute care facility (emergency or intensive care unit—10 trials, 4,129 patients)^{7,35,36,53,56–59,69,70} ; (4) younger, less comorbid population admitted for acute traumatic brain injury or subarachnoid hemorrhage (2 trials, 244 patients)^61,65 ; (5) other patients or settings: one conducted among anemic women in the postpartum phase and one including thrombocytopenic middle-aged patients with hematologic cancer (2 trials, 579 patients).^64,71 

In the first group, four studies failed in implementing their protocol: in two, the absolute difference in hemoglobin thresholds was only 10 g/l, the transfusion-sparing effect was less than 20%, and nonadherence rates in the restrictive group were more than or equal to 15%;^47,48  in one, transfusion thresholds were higher in some patients of the restrictive group compared to liberal⁶⁷ ; in one, nonadherence rates, transfusion-sparing effects, and transfusion thresholds seemed adequate, but the sample size was small.⁵¹ 

In group 2, data to assess success were lacking in one trial,⁶³  and three studies showed unsuccessful implementation: in one, the transfusion-sparing effect was less than 20%, and nonadherence rates in the restrictive group were more than or equal to 15%⁶² ; in one, the lack of standardized protocol in the liberal group resulted in a negative transfusion-sparing effect⁶⁸ ; and in one, surgical procedures were not adequately balanced between groups (liberal patients were more likely to bleed).⁸ 

In group 3, success was unclear in two studies (“mean erythrocyte units” were reported per transfused patients instead of per strategy group).^7,53  In one unsuccessful study, the transfusion-sparing effect was less than 20%, and nonadherence rates in the liberal group were more than or equal to 15%.³⁵ 

Group 4 included two studies that demonstrated successful implementation but the protocols were heterogeneous: one study used particularly high thresholds (restrictive: hemoglobin 100 g/l; liberal: hemoglobin 115 g/l), while the other compared 70 with 100 g/l.^61,65 

In the last group, success was unclear in one study reporting negligible transfusion-sparing effects and high nonadherence rates in the restrictive group.⁷¹ 

The risk of selection bias was deemed low in only 7 of 31 included studies (see fig., Supplemental Digital Content 9, http://links.lww.com/ALN/B283, which summarizes the risk of bias in included studies).^{7,56,58,66,68,70,71}  All studies were at high or unclear risk of performance bias. Outcome assessors were blinded in 13 studies.^{6–8,36,49,50,54,55,58,60,61,65,71}  The risk of attrition bias was low in all studies except three: in two,^54,67  it was not possible to assess if attrition was balanced between groups, and in one,⁶⁴  attrition rates exceeded 20%. The risk of selective reporting was low in all studies, but in one trial, study findings were reported in three different publications without mention of other existing reports.^9,41,42 

(See fig., Supplemental Digital Content 10, http://links.lww.com/ALN/B284,which illustrates the risk of events reflecting inadequate oxygen supply.) In group 1, early MI was reported in seven studies, arrhythmia in five, angina in two, stroke or transient ischemic attack (TIA) in five, acute kidney injury in five, and mesenteric ischemia in one (see table, Supplemental Digital Content 11, http://links.lww.com/ALN/B285, which outlines outcome reporting across studies). In one study, stroke/TIA was reported in combination with delirium and could not be extracted.⁵⁵  Thus, data from 8 studies (3,322 patients) were combined^{6,47,48,51,55,60,66,67} : in patients with cardiovascular disease assigned to a restrictive strategy and undergoing high-risk surgery, there was a possible increase in events reflecting inadequate oxygen supply (risk ratio [RR], 1.09; 95% CI, 0.97 to 1.22).

In group 2, early MI was reported in five studies, arrhythmia in three, stroke/TIA in five, and AKI in two (see table, Supplemental Digital Content 11, http://links.lww.com/ALN/B285, which outlines outcome reporting across studies). Outcomes could not be extracted in one study reporting “neuropsychiatric complications.”⁶⁸  Thus, data from 7 studies (3,465 patients) were combined^{8,49,50,52,54,63,68} : in an elderly population undergoing orthopedic surgery, events reflecting inadequate oxygen supply were significantly increased in the restrictive group (RR, 1.41; 95% CI, 1.03 to 1.92).

In group 3, six studies reported “inadequate oxygen” events (see table, Supplemental Digital Content 11, http://links.lww.com/ALN/B285, which outlines outcome reporting across studies).^7,56,58,70  Data from one study could not be extracted, since events were reported as composite outcomes.⁵⁶  When data were combined, there was no difference between groups (3,590 patients; RR, 0.89; 95% CI, 0.72 to 1.09).

Among group 4 and 5 studies, only one study reported the incidence of stroke.⁶¹  In this trial, patients with subarachnoid hemorrhage at risk of cerebral vasospasm were allocated to high transfusion thresholds. There was no difference between groups (RR, 1.36; 95% CI, 0.59 to 3.15).

(See fig., Supplemental Digital Content 12, http://links.lww.com/ALN/B286, which illustrates the risk of early mortality.) Most group 1 to 3 studies reported early mortality (see table, Supplemental Digital Content 11, http://links.lww.com/ALN/B285, which outlines outcome reporting across studies). When data were combined, there was a possible increase in events when a restrictive strategy was applied in group 1 (RR, 1.39; 95% CI, 0.95 to 2.04—7 studies, 3,245 patients)^{6,47,48,51,55,60,66}  but not in group 2 (RR, 1.09; 95% CI, 0.80 to 1.49—7 studies, 3,546 patients)^{8,9,49,50,54,63,68}  or group 3 (RR, 0.94; 95% CI, 0.73 to 1.20—7 studies, 2,894 patients).^{7,35,36,56–59}  Mortality was not reported in groups 4 and 5.

The risk of events was significantly increased when patients were assigned to a restrictive transfusion strategy in group 1 (RR, 1.12; 95% CI, 1.01 to 1.24—8 studies, 3,322 patients)^{6,47,48,51,55,60,66,67}  and group 2 (RR, 1.24; 95% CI, 1.00 to 1.54—8 studies, 3,749 patients)^{8,9,49,50,52,54,63,68}  but not in group 3 (RR, 0.90; 95% CI, 0.74 to 1.10—8 studies, 3,762 patients).^{7,35,36,56–59,70}  Data in groups 4 and 5 were too scarce to allow statistical combination.

The incidence of infections was reported in the majority of included studies (see table, Supplemental Digital Content 11, http://links.lww.com/ALN/B285, which outlines outcome reporting across studies). In groups 1 and 3, no difference was found (group 1: RR, 1.11; 95% CI, 0.94 to 1.31—6 studies, 3,141 patients^{6,47,55,60,66,67} ; group 3: RR, 0.99; 95% CI, 0.85 to 1.17—5 studies, 2,616 patients^{7,35,56,59,70} ). In group 2, patients assigned to a restrictive policy seemed to have less septic events (RR, 0.75; 95% CI, 0.53 to 1.04—9 studies, 3,815 patients).^{8,9,49,50,52,54,62,63,68}  In group 4, both studies reported the incidence of infection, but transfusion protocols were deemed too heterogeneous for statistical combination. When individually assessed, no difference was found (RR, 0.77; 95% CI, 0.51 to 1.16⁶⁵ ; RR, 0.91; 95% CI, 0.14 to 5.92⁶¹ ). Finally, in a study conducted in the postpartum setting, there was no difference between groups (RR, 1.08; 95% CI, 0.63 to 1.87).⁶⁴ 

We found significant interaction between risk strata, thereby indicating that our context-specific approach was appropriate (inadequate oxygen supply: Cochran Q P = 0.003, I² = 82.7%; early mortality: P = 0.11 but I² = 54.2%; composite outcome: P = 0.0007, I² = 86.1%; infections: P = 0.04, I² = 69.1%). Data pooling without controlling for clinical heterogeneity (i.e., no context-specific approach) resulted in a dilution of the intervention effect (inadequate oxygen supply: RR, 1.02; 95% CI, 0.94 to 1.11; early mortality: RR, 1.00; 95% CI, 0.89 to 1.12; composite outcome: RR, 1.01; 95% CI, 0.95 to 1.08; infections: RR, 1.97; 95% CI, 0.89 to 1.07).

The effect of successful protocol implementation on the risk of AE was explored by excluding unsuccessful studies from each analysis (see table, Supplemental Digital Content 13, http://links.lww.com/ALN/B287, which outlines how risk estimates varied according to successful protocol implementation). In group 1, risk estimates increased further away from the null: patients in the restrictive group seemed to have more events reflecting inadequate oxygen supply (RR, 1.12; 95% CI, 0.99 to 1.27), a 59% increase in mortality (RR, 1.59; 95% CI, 1.04 to 2.44) and a significant increase in the composite outcome (RR, 1.16; 95% CI, 1.03 to 1.31). However, these findings were not reproducible for groups 2 and 3: risk estimates decreased toward the null or further in favor of a restrictive strategy. The effect of successful implementation on infections was inconsistent.

We also explored the effect of cointerventions on the risk of AE, but data were scarce and comparison across studies was difficult (see table, Supplemental Digital Content 14, http://links.lww.com/ALN/B288, which illustrates cointerventions across studies). The use of clotting factor concentrates or antifibrinolytics was reported in three cardiac and one orthopedic surgery studies.^55,60,62,66  Their administration was well balanced between transfusion groups. The use of cryoprecipitate, fresh frozen plasma, and platelets was reported in four,^35,55,60,66  eight,^{35,52,55,58,60,62,66,70}  and seven studies,^{35,55,58,60,66,70,71}  respectively. Overall, the administration of blood products was similar between transfusion groups (cryoprecipitate: RR, 0.99; 95% CI, 0.77 to 1.27; fresh frozen plasma: RR, 0.89; 95% CI, 0.75 to 1.04); platelet therapy: RR, 0.95; 95% CI, 0.81 to 1.12), but compared to others, patients undergoing cardiac surgery were more systematically exposed to nonerythrocyte blood products. Finally, 10 studies reported intravenous fluid usage: in only three, unbalanced administration was found.^52,57,62 

Exploring the effect of studies at high or unclear risk of detection and attrition bias did not yield clinically meaningful results: because of the small remaining number of studies, 95% CI was large and risk estimates varied only mildly (data not shown).

In this context-specific systematic review, we found that restrictive transfusion strategies were associated with an increased risk of complications in situations combining high-risk patients with major surgery. Those with cardiovascular disease undergoing cardiac or vascular procedures seemed to have more events reflecting inadequate oxygen supply, higher mortality rates, or both. In the elderly orthopedic population, a restrictive policy led to a 40% increase in ischemic events or AKI.

These findings were consistent with physiologic studies suggesting that multiple perioperative factors may undermine normal compensatory responses to anemia^72,73 : cardiac dysfunction (induced by anesthetic drugs or surgical trauma), vasoconstriction (due to endogenous or exogenous catecholamines), or postoperative hypoventilation (due to pain or residual effects of anesthetics) may compromise adequate oxygen delivery to vital organs. In normal conditions, systemic oxygen delivery largely exceeds oxygen consumption, resulting in a positive oxygen reserve.⁷³  In patients having a preexisting low reserve, however, the combination of acute anemia with impaired compensatory responses may induce a state of oxygen supply dependency, resulting in acidosis and organ failure. In this particular situation, administrating erythrocyte could restore the oxygen reserve by increasing blood oxygen content and tissue oxygenation.⁷³ 

Surprisingly, no evidence of harm was found when restrictive strategies were applied in critically ill patients, although similar impairment of compensatory responses was expected. One explanation might be the heterogeneity in oxygen reserve among this mixed population: medical patients might be at lower risk of oxygen supply dependency than their surgical counterparts, who have the additional burden of surgery, pain, and recovery from anesthesia.

When we excluded studies showing unsuccessful protocol implementation, effect estimates differed across risk strata. While harm was more pronounced in patients undergoing cardiac or vascular procedures, it decreased toward the null in those undergoing orthopedic surgery. Two reasons may account for this: first, cardiac patients were probably more likely to enter a state of oxygen supply dependency (due to the combination of advanced cardiovascular disease and high-risk surgery) than the orthopedic population, where ischemic heart disease ranged from 10 to 45%. Second, cardiac patients might have been exposed to context-specific factors increasing their risk of AE: antifibrinolytics and clotting factor concentrates were more systematically administered in this population, while this was rarely reported in orthopedic patients. Although tranexamic acid seems safe in trauma patients,⁷⁴  the thrombogenic potential of other hemostatic agents in nontrauma settings remains unclear.^75–77  Thus, the combination of a restrictive transfusion strategy and highly thrombogenic drugs in patients with advanced cardiovascular disease might be particularly unfavorable.

Our analysis of the risk of infection remained inconclusive. Although there was a possible reduction in septic events in orthopedic patients assigned to a restrictive strategy, effect estimates differed widely across risk strata: the benefit of reduced exposure to allogeneic blood was less clear in cardiac patients, which is consistent with a previously published meta-analysis.¹⁰  One reason might be that immune response impairment was more pronounced in these patients: in our analysis, the cardiac surgery population was more likely to receive nonerythrocyte allogeneic blood products (such as platelets or fresh frozen plasma), which are also known to have immunomodulatory effects.^33,78–80  Alternatively, cardiac surgery itself might induce particularly high levels of perioperative stress, which has also been found to interfere with immune responses.^22,81–83 

Our systematic review differs from others in many aspects. First, in most previous meta-analyses, data were combined despite high clinical heterogeneity,^10,13,14,16  which may hinder the identification of group-specific effects. We addressed this methodologic limitation by performing a context-specific analysis, using strict criteria for risk-strata generation and subsequent data pooling. We were eventually able to show that indiscriminate data combination (i.e., performing analyses without controlling for clinical diversity) resulted in a dilution of the intervention effects. Furthermore, to fully explore the impact of clinical diversity, the issue of transfusion protocol variability was examined using two performance indicators and exploring three different determinants of success. One other meta-analysis investigated protocol diversity, but only hemoglobin thresholds were explored.¹⁰  Additionally, our review was the first to assess the role of cointerventions: we were able to show that in some patients, the combination of restrictive transfusion policies and thrombogenic drugs could be particularly detrimental; we also identified a possible effect of nonerythrocyte blood products and of perioperative stressors on the risk of infection.

This review has some limitations. First, although strict criteria were used to handle unclear/missing outcome data, outcome reporting and definitions varied across studies, and so this may result in residual clinical diversity. However, a certain degree of heterogeneity is desirable to ensure wide applicability of the findings. Second, in order to capture the full spectrum of effects related to transfusion strategies and to improve statistical precision, we used arbitrarily defined outcome categories. However, although endpoints combination might be biologically well founded, individual components may differ in clinical importance, and our categories may have failed to reflect endpoints truly relevant for patients.³⁰  Third, in some risk strata, large studies having high event rates appeared to dominate the analysis, but the risk of a small-study effect was deemed low: we used the Mantel-Haenszel method to account for smaller studies and addressed thoroughly all sources of clinical and methodologic diversity. It seems therefore unlikely that our findings derive solely from the effect of larger studies. Fourth, our assessment of methodologic heterogeneity (detection and attrition bias) was hampered by the scarcity of data. The same problem was encountered with data on cointerventions, which reduced our ability to fully explore their role in the occurrence of complications. Finally, indicators of successful protocol implementation were arbitrarily defined; using a different model might have yielded other results.

This analysis provided clear evidence that the decision to transfuse (or not transfuse) requires more than a “one-size-fits-all” approach. As highlighted recently,^84–86  the identification of populations at higher risk of oxygen supply dependency who might particularly benefit from erythrocyte administration remains a real challenge. New transfusion algorithms should aim to integrate additional clinical parameters, such as patient comorbidities, particular settings, or oxygen reserve estimates.⁷³  We recommend that future trials systematically collect and report data regarding the use of nonerythrocyte blood products, antifibrinolytics, and clotting factor concentrates, since their role in the risk of AE remains unclear.

This meta-analysis suggests that the use of restrictive transfusion strategies might be detrimental in high-risk patients undergoing major surgery. Further research is needed to evaluate the contributing role of cointerventions in the occurrence of complications.

Supported by the Professor Dr. Max Cloëtta Foundation, Zurich, Switzerland (to Dr. Hovaguimian), the Uniscientia Foundation of Vaduz, Zurich, Switzerland (to Dr. Hovaguimian), institutional funds (University Hospital of Zurich, Zurich, Switzerland; to Dr. Hovaguimian), an Australian National Health and Medical Research Council Practitioner Fellowship (APP1042462; to Dr. Myles), Canberra, Australia.

The authors declare no competing interests. The funding organizations had no role in the design and conduct of the study; in the collection, management, analysis, or interpretation of the data; or in the preparation, review, or approval of the manuscript.

Full protocol available from Dr. Hovaguimian: frederique.hovaguimian@usz.ch. Raw data available from Dr. Hovaguimian: frederique.hovaguimian@usz.ch.