Transfusion thresholds for guiding red blood cell transfusion

Jeffrey L Carson<sup>a</sup>; Simon J Stanworth<sup>a</sup>; Jane A Dennis<sup>a</sup>; Marialena Trivella; Nareg Roubinian; Dean A Fergusson; Darrell Triulzi; Carolyn Dorée; Paul C Hébert

doi:10.1002/14651858.CD002042.pub5

Transfusion thresholds for guiding red blood cell transfusion

Authors' declarations of interest

Version published: 21 December 2021 Version history

https://doi.org/10.1002/14651858.CD002042.pub5

Collapse all Expand all

Abstract

available in

Background

The optimal haemoglobin threshold for use of red blood cell (RBC) transfusions in anaemic patients remains an active field of research. Blood is a scarce resource, and in some countries, transfusions are less safe than in others because of inadequate testing for viral pathogens. If a liberal transfusion policy does not improve clinical outcomes, or if it is equivalent, then adopting a more restrictive approach could be recognised as the standard of care.

Objectives

The aim of this review update was to compare 30‐day mortality and other clinical outcomes for participants randomised to restrictive versus liberal red blood cell (RBC) transfusion thresholds (triggers) for all clinical conditions. The restrictive transfusion threshold uses a lower haemoglobin concentration as a threshold for transfusion (most commonly, 7.0 g/dL to 8.0 g/dL), and the liberal transfusion threshold uses a higher haemoglobin concentration as a threshold for transfusion (most commonly, 9.0 g/dL to 10.0 g/dL).

Search methods

We identified trials through updated searches: CENTRAL (2020, Issue 11), MEDLINE (1946 to November 2020), Embase (1974 to November 2020), Transfusion Evidence Library (1950 to November 2020), Web of Science Conference Proceedings Citation Index (1990 to November 2020), and trial registries (November 2020). We checked the reference lists of other published reviews and relevant papers to identify additional trials. We were aware of one trial identified in earlier searching that was in the process of being published (in February 2021), and we were able to include it before this review was finalised.

Selection criteria

We included randomised trials of surgical or medical participants that recruited adults or children, or both. We excluded studies that focused on neonates.

Eligible trials assigned intervention groups on the basis of different transfusion schedules or thresholds or 'triggers'. These thresholds would be defined by a haemoglobin (Hb) or haematocrit (Hct) concentration below which an RBC transfusion would be administered; the haemoglobin concentration remains the most commonly applied marker of the need for RBC transfusion in clinical practice. We included trials in which investigators had allocated participants to higher thresholds or more liberal transfusion strategies compared to more restrictive ones, which might include no transfusion. As in previous versions of this review, we did not exclude unregistered trials published after 2010 (as per the policy of the Cochrane Injuries Group, 2015), however, we did conduct analyses to consider the differential impact of results of trials for which prospective registration could not be confirmed.

Data collection and analysis

We identified trials for inclusion and extracted data using Cochrane methods. We pooled risk ratios of clinical outcomes across trials using a random‐effects model. Two review authors independently extracted data and assessed risk of bias. We conducted predefined analyses by clinical subgroups. We defined participants randomly allocated to the lower transfusion threshold as being in the 'restrictive transfusion' group and those randomly allocated to the higher transfusion threshold as being in the 'liberal transfusion' group.

Main results

A total of 48 trials, involving data from 21,433 participants (at baseline), across a range of clinical contexts (e.g. orthopaedic, cardiac, or vascular surgery; critical care; acute blood loss (including gastrointestinal bleeding); acute coronary syndrome; cancer; leukaemia; haematological malignancies), met the eligibility criteria. The haemoglobin concentration used to define the restrictive transfusion group in most trials (36) was between 7.0 g/dL and 8.0 g/dL. Most trials included only adults; three trials focused on children.

The included studies were generally at low risk of bias for key domains including allocation concealment and incomplete outcome data.

Restrictive transfusion strategies reduced the risk of receiving at least one RBC transfusion by 41% across a broad range of clinical contexts (risk ratio (RR) 0.59, 95% confidence interval (CI) 0.53 to 0.66; 42 studies, 20,057 participants; high‐quality evidence), with a large amount of heterogeneity between trials (I² = 96%).

Overall, restrictive transfusion strategies did not increase or decrease the risk of 30‐day mortality compared with liberal transfusion strategies (RR 0.99, 95% CI 0.86 to 1.15; 31 studies, 16,729 participants; I² = 30%; moderate‐quality evidence) or any of the other outcomes assessed (i.e. cardiac events (low‐quality evidence), myocardial infarction, stroke, thromboembolism (all high‐quality evidence)). High‐quality evidence shows that the liberal transfusion threshold did not affect the risk of infection (pneumonia, wound infection, or bacteraemia). Transfusion‐specific reactions are uncommon and were inconsistently reported within trials.

We noted less certainty in the strength of evidence to support the safety of restrictive transfusion thresholds for the following predefined clinical subgroups: myocardial infarction, vascular surgery, haematological malignancies, and chronic bone‐marrow disorders.

Authors' conclusions

Transfusion at a restrictive haemoglobin concentration decreased the proportion of people exposed to RBC transfusion by 41% across a broad range of clinical contexts. Across all trials, no evidence suggests that a restrictive transfusion strategy impacted 30‐day mortality, mortality at other time points, or morbidity (i.e. cardiac events, myocardial infarction, stroke, pneumonia, thromboembolism, infection) compared with a liberal transfusion strategy.

Despite including 17 more randomised trials (and 8846 participants), data remain insufficient to inform the safety of transfusion policies in important and selected clinical contexts, such as myocardial infarction, chronic cardiovascular disease, neurological injury or traumatic brain injury, stroke, thrombocytopenia, and cancer or haematological malignancies, including chronic bone marrow failure.

Further work is needed to improve our understanding of outcomes other than mortality. Most trials compared only two separate thresholds for haemoglobin concentration, which may not identify the actual optimal threshold for transfusion in a particular patient. Haemoglobin concentration may not be the most informative marker of the need for transfusion in individual patients with different degrees of physiological adaptation to anaemia. Notwithstanding these issues, overall findings provide good evidence that transfusions with allogeneic RBCs can be avoided in most patients with haemoglobin thresholds between the range of 7.0 g/dL and 8.0 g/dL. Some patient subgroups might benefit from RBCs to maintain higher haemoglobin concentrations; research efforts should focus on these clinical contexts.

PICOs

Population

Intervention

Comparison

Outcome

The PICO model is widely used and taught in evidence-based health care as a strategy for formulating questions and search strategies and for characterizing clinical studies or meta-analyses. PICO stands for four different potential components of a clinical question: Patient, Population or Problem; Intervention; Comparison; Outcome.

See more on using PICO in the Cochrane Handbook.

Plain language summary

available in

Is it safe to use lower blood counts (haemoglobin levels) as a trigger for blood transfusion in order to give fewer blood transfusions?

Key messages

•There is no evidence that giving blood transfusions to patients with lower blood counts (haemoglobin levels of 7.0 g/dL to 8.0 g/dL) compared to higher blood counts (9.0 g/dL to 10.0 g/dL) affects risks of death, heart attack, myocardial infarction, stroke, pneumonia, blood clots or infection.

• Giving blood only to patients with lower blood counts (7.0 g/dL to 8.0 g/dL) would reduce the amount of blood transfused substantially. It would also reduce the risk of unnecessary transfusions (transfusions can have harmful effects).

• More research is needed to:

‐ establish the blood count at which a blood transfusion is needed in people who have suffered a heart attack, brain injury, or have cancer; and to

‐ improve our understanding of outcomes other than death, including quality of life.

What happens in people who need blood transfusions?

Doctors and healthcare professionals often give blood transfusions to people who lose blood through surgery, bleeding, or illness. For example, blood transfusions may help patients with anaemia to recover after surgery, but they should only be given when they help people to get better from their medical condition. Blood is a limited resource and transfusion is not risk‐free, especially for people in low‐income countries where the blood used in transfusions may not be tested for harmful viruses such as HIV or hepatitis.

What did we want to find out?

The blood count measures the amount of haemoglobin in the blood. Haemoglobin is a protein that gives blood its red colour and carries oxygen around the body. A normal blood count is around 12 grams a decilitre (12 g/dL). We wanted to find out if it is safe to withhold blood transfusion until the blood count drops to between 7.0 g/dL to 8.0 g/dL, rather than transfusing sooner at higher blood counts of between 9.0 g/dL to 10.0 g/dL.

What did we do?

We examined the results of studies that allocated patients to one of two groups by chance (for example, by flipping a coin). In one group, the patients only received blood transfusions if their blood count fell below a higher threshold (typically, 9.0 g/dL to 10.0 g/dL). In the other group, the patients only received blood transfusions if their blood counts fell below a lower threshold (typically, 7.0 g/dL to 8.0 g/dL).

What did we find?

We found 48 studies that involved 21,433 patients. The patients had been hospitalised for a range of reasons including: bone (orthopaedic), heart (cardiac) or vascular surgery; critical care; acute blood loss (for example, through bleeding in the stomach or intestines); heart diseases; cancer and blood cancers. The studies compared higher or lower blood count thresholds for blood transfusion. (The ‘threshold’ is the blood count level that would need to be met before a transfusion would be given.)

Transfusion

We found that patients who received transfusions only at lower blood count thresholds were 41% less likely to receive a blood transfusion than those who received them only at higher blood count thresholds. If the lower threshold were applied routinely by medical staff, it would lead to a substantial reduction in the quantity of blood needed.

Death and harmful events

There was no clear difference in the risk of dying within 30 days of receiving, or not receiving, a transfusion for patients in the two different threshold groups.

There was also no clear difference between the low and high threshold groups for the number of serious harmful events that occurred after patients received, or did not receive, blood transfusions. The harmful events recorded included infection (pneumonia, wound infection, and blood poisoning), heart attacks, strokes, and problems with blood clots.

What are the limitations of the evidence?

We found that most of the studies provided a high quality of evidence; they were adequately conducted and used methods that minimised biases that could make the validity of the results uncertain.

We are confident in the evidence regarding likelihood of receiving a transfusion, death within 30 days of transfusion, heart attack, stroke and infection. We are moderately confident in the evidence for problems caused by blood clots, but too few occurred in either group for us to be more confident.

Too few studies evaluated quality of life for us to be able to see whether it varied between groups.

How up to date is this evidence?

This Cochrane Review updates our previous work on this subject (last published in 2016). Seventeen new studies are included. The evidence is up to date to November 2020.

Authors' conclusions

Implications for practice

Analysis of published evidence reveals that transfusing at a restrictive strategy of 7.0 g/dL to 8.0 g/dL, compared with a liberal haemoglobin threshold of 9.0 g/dL to 10.0 g/dL, across a broad range of hospitalised patients does not have an adverse effect on clinical outcomes, including 30‐day mortality, myocardial infarction, congestive heart failure, and infection.

Given there is no evidence of additional benefit of red blood cell (RBC) transfusion at higher haemoglobin concentration thresholds (9.0 g/dL to 10.0 g/dL), and that blood for transfusion is a costly and scarce biological resource with finite risks, a restrictive transfusion trigger policy (7.0 g/dL to 8.0 g/dL) could be widely adopted. A restrictive transfusion policy is not associated with increased adverse events and reduces both risk of exposure to RBC transfusion and the total number of units transfused.

Trial interventions varied on the haemoglobin concentration used to define the restrictive transfusion group. About half of the trials used a 7.0‐g/dL threshold, and the other half used a threshold of 8.0 g/dL to 9.0 g/dL. However, within each clinical subgroup, the number of clinical trials (and the total numbers of enrolled patients) testing restrictive thresholds at 7.0 g/dL varied. The exact implications for transfusion practice regarding the nature of restrictive haemoglobin thresholds will, therefore, vary by clinical group.

In critical care trials, a 7.0‐g/dL threshold was used most frequently and shown to have a similar safety profile to higher thresholds for mortality (RR 1.06, 95% CI 0.85 to 1.32). Similarly, a restrictive threshold of 7.0 g/dL was used in trials including patients with acute blood loss from gastrointestinal bleeding; evidence indicates that these patients have lower risk of 30‐day mortality with restrictive transfusion that uses a 7.0‐g/dL threshold.

In patients undergoing cardiac surgery, a restrictive threshold of 7.5 g/dL (rather than 7.0 g/dL) was used in the largest trials and shown to have a risk for mortality which was similar to that of higher thresholds.

In trials of orthopaedic surgery, the restrictive strategy used most frequently was 8.0 g/dL, which had a similar risk profile for mortality as higher transfusion thresholds. In this clinical subgroup, it is not possible to conclude that 7.0 g/dL is as efficacious as 8.0 g/dL, without testing lower thresholds in trials.

In other clinical subgroups, the results do not provide adequate evidence to conclude which specific restrictive transfusion threshold should be applied. These subgroups include vascular surgery and haematological malignancies, where trials are insufficient in number or recruit only small numbers of participants.

The analysis does provide some evidence that a restrictive strategy might be appropriate for patients with underlying cardiovascular disease. The REALITY trial conducted in patients with acute myocardial infarction found fewer deaths and fewer major adverse cardiac events (MACE) with a restrictive threshold of 8.0 g/dL (Ducrocq 2021). However, pooled analysis of all three trials in people with acute myocardial infarction (820 participants) reveals that the risk ratio of 1.61 and very wide 95% confidence intervals (0.38 to 6.88) are also consistent with the possibility of significant benefit for more liberal transfusion policies.

In summary, it is not possible to suggest a single restrictive transfusion threshold across all clinical groups and patients with anaemia. While it is possible that a 7.0‐g/dL threshold could be used in most adult patients, in some settings trial data for thresholds of 7.0 g/dL do not exist. Without these data, it is impossible to be certain of the effects of higher or lower thresholds in these settings. Trials that should clarify the optimal threshold in some of the most important subgroups that currently lack data are now underway (Ongoing studies).

Evidence is insufficient to evaluate the effects of different strategies on functional recovery. Quality of life is an important outcome in many trial settings, for example, people who are transfusion dependent and are managed in outpatient settings. In our review, most included randomised trials were based on patients hospitalised for the management of 'acute' anaemia. In contrast, patients with chronic bone marrow failure, such as myelodysplasia, are transfusion dependent for prolonged periods of time at home, and this may persist for years, yet our understanding of the impact of different transfusion policies on quality of life and functional outcomes for these patients is incomplete.

For countries where there are concerns about microbiological screening and the safety of donated blood, the data in this updated review constitute a strong basis for avoiding liberal RBC transfusion in many clinical settings. The benefits of minimising allogeneic RBC transfusion are likely to be greatest when there is doubt about the safety of the blood supply (WHO 2016). There is a need for practice and research to implement our review findings, with support for education and training and updating of robustly constructed guidelines (Kwan 2020; Pavenski 2018; Smith 2020; Vlaar 2020).

Implications for research

The totality of research evidence now allows us to make firmer recommendations for research priorities. Further randomised trials should be targeted to address specific research questions when the strength of evidence‐based recommendations has significant uncertainty, rather than repeating trials in clinical settings or at haemoglobin thresholds for which the evidence base is better defined. Acute cardiovascular disease is a high‐priority area, for which a restrictive approach may not be as safe, but ongoing larger trials have the potential to provide additional evidence (NCT02981407; Shah 2020). Limited data are available for participants in clinical contexts of acute coronary syndrome, myocardial infarction, neurological injury/traumatic brain injury, acute neurological disorders, and stroke. Areas of uncertainty in cancer and haematological malignancy include chronic bone marrow failure and the role of transfusion at different thresholds for patients receiving radiotherapy (Hoff 2011). Liberal thresholds for red cell transfusion could provide important additive benefits for important outcomes such as fatigue and quality of life for elderly patients with chronic bone marrow failure, who may be transfusion dependent for many years (Stanworth 2020).

Although one large pragmatic trial has been undertaken in critically ill children (Lacroix 2007), many children with anaemia, although eligible, were not recruited into this trial, and further research is warranted to examine the generalisability of these trial findings for all groups of sick infants and children, including those with cardiac disorders. Patients with severe burn injuries who require large volumes of red cells may also present in a clinical context that requires further research.

In summary, we believe that in these selected clinical contexts, clinical goals and pathophysiology preclude generalisation from the completed trials included in this review to date, and there remains uncertainty regarding optimal transfusion practices in these subgroups.

There is a need to continue to update this review, given the large number of ongoing trials, which reflects an active programme of research in the field of red cell transfusion research. All new trials should be adequately powered and apply consistent definitions for clinical outcomes, such as infection, myocardial infarction and ischaemic heart disease (Docherty 2018). Outcomes of importance in trials will continue to include mortality, along with outcomes that are more specific and relevant to the clinical setting, such as function and quality of life measures and bleeding endpoints in transfusion‐dependent patients with cancer and haematological malignancy. All new trials should be prospectively registered, to assist researchers in assessing risk of selective outcome reporting and other matters related to research integrity.

Trials are also needed to evaluate haemoglobin concentrations below 7.0 g/dL, such as 6.0 g/dL (Yao 2020), which may be especially relevant in countries with suboptimal blood safety and inadequate blood supply (Maitland 2019). One randomised feasibility trial identified as ongoing in children undergoing allogeneic haematopoietic stem cell transplantation is comparing restrictive versus liberal red cell transfusion strategies using haemoglobin concentrations of < 6.5 g/dL and < 8.0 g/dL respectively (ISRCTN17438123). Further research should recognise the need for gender‐specific reference ranges for haemoglobin concentration (Butcher 2017).

A limitation of this trial‐level meta‐analysis is the difficulty in analysing subgroups of patients with varying underlying diseases, age, and clinical settings. This is especially important when considering transfusion thresholds because there are pathophysiological and clinical data to suggest that the optimal transfusion threshold could differ according to the patient's underlying co‐morbidity (e.g. cardiovascular disease), gender and age. One approach to address this limitation would be to conduct individual patient data meta‐analysis by obtaining and analysing detailed information on each patient enrolled in the trials.

Research is needed to identify methods used to measure oxygen delivery to vital organs directly, and to define the need for red cell transfusion more precisely. This recognises the challenge of applying haemoglobin concentration as an imperfect surrogate marker of transfusion requirements (Baek 2019; Mueller 2019; Ochocinska 2020). Although it is beyond the scope of this review, further research should explore factors related to the red cell product used for transfusion, including differences in processing and characteristics of the product. As one example of a donor‐specific factor that may be highly relevant for interpretation of trial results, our analysis revealed that only 16 trials clearly specified use of leuco‐depleted red cells in the trial protocol (Turner 2018).

Consideration should be given to aspects of trial design in the future. The most common design identified in eligible studies in our review remains the parallel two‐arm trial, in which two (arguably arbitrarily defined) thresholds for haemoglobin concentration are compared. A two‐arm trial design, while simple and pragmatic, may be an inefficient approach for identifying the exact optimal threshold for transfusion in a setting ‐ it may indeed be a different threshold than those tested in the trial. We and other groups have highlighted the fact that the actual separation of haemoglobin concentrations attained in trials varies considerably (Trentino 2020a; Trentino 2020b), and this separation rarely approaches the hoped for differences defined by protocol trial interventions.

Allied research is needed to address the optimal target haemoglobin concentration post transfusion, which will depend on the dose of transfusion given. Doses of red cells in adults are increasingly recommended as single‐unit transfusions in non‐bleeding patients, but the evidence base is limited (Shih 2018). The question of optimal dose for transfusion is particularly relevant for children and infants, given that a common paediatric dose is 10 mL/kg to 15 mL/kg, which is much greater proportionately by weight than the single unit dose (or bag) commonly used in adults (New 2016).

Ultimately, the optimal threshold for transfusion is likely to vary between patients, and new trial designs are needed that can test and evaluate targeted or personalised approaches to the need for transfusion, possibly incorporating individual physiological parameters. Indeed, it has been argued that current trials do not allow a genuine 'standard of care' arm (Wang 2010), which would mirror current clinical practice by clinicians at the bedside.

Summary of findings

Open in table viewer

Summary of findings 1. Liberal compared with restrictive transfusion protocols for guiding red blood cell transfusion

Outcomes	*Anticipated absolute effects (95% CI)**		Relative effect (95% CI)	№. of participants (studies)	Certainty of the evidence (GRADE)	Comments
Liberal compared with restrictive transfusion protocols for guiding red blood cell transfusion
Patient or population: adults and children (haemodynamically stable) with potential need for RBC transfusion Setting: inpatients Intervention: restrictive transfusion threshold Comparison: liberal transfusion threshold
Outcomes	Risk with liberal transfusion protocol	Risk with restrictive transfusion protocol	Relative effect (95% CI)	№. of participants (studies)	Certainty of the evidence (GRADE)	Comments
Participants exposed to blood transfusion (all studies)	Study population		RR 0.59 (0.53 to 0.66)	20,057 (42)	⊕⊕⊕⊕ High
Participants exposed to blood transfusion (all studies)	815 per 1000	481 per 1000 (432 to 538)	RR 0.59 (0.53 to 0.66)	20,057 (42)	⊕⊕⊕⊕ High
30‐Day mortality	Study population		RR 0.99 (0.86 to 1.15)	16,729 (31)	⊕⊕⊕⊕ High
30‐Day mortality	83 per 1000	83 per 1000 (71 to 96)	RR 0.99 (0.86 to 1.15)	16,729 (31)	⊕⊕⊕⊕ High
Myocardial infarction	Study population		RR 1.04 (0.87 to 1.24)	14,370 (23)	⊕⊕⊕⊕ High
Myocardial infarction	32 per 1000	33 per 1000 (28 to 40)	RR 1.04 (0.87 to 1.24)	14,370 (23)	⊕⊕⊕⊕ High
Congestive heart failure	Study population		RR 0.83 (0.53 to 1.29)	7247 (16)	⊕⊕⊝⊝ Low^a
Congestive heart failure	35 per 1000	29 per 1000 (19 to 45)	RR 0.83 (0.53 to 1.29)	7247 (16)	⊕⊕⊝⊝ Low^a
Cerebrovascular accident ‐ stroke	Study population		RR 0.84 (0.64 to 1.09)	13,985 (19)	⊕⊕⊕⊕ High
Cerebrovascular accident ‐ stroke	17 per 1000	14 per 1000 (11 to 19)	RR 0.84 (0.64 to 1.09)	13,985 (19)	⊕⊕⊕⊕ High
Rebleeding	Study population		RR 0.80 (0.59 to 1.09)	3412 (8)	⊕⊕⊕⊝ Moderate^b
Rebleeding	158 per 1000	126 per 1000 (93 to 172)	RR 0.80 (0.59 to 1.09)	3412 (8)	⊕⊕⊕⊝ Moderate^b
Thromboembolism	Study population		OR 1.11 (0.65 to 1.88)	4201 (13)	⊕⊕⊕⊝ Moderate^c
Thromboembolism	15 per 1000	17 per 1000 (10 to 28)	OR 1.11 (0.65 to 1.88)	4201 (13)	⊕⊕⊕⊝ Moderate^c
Infection	Study population		RR 0.97 (0.88 to 1.07)	17,104 (25)	⊕⊕⊕⊕ High
Infection	143 per 1000	139 per 1000 (126 to 153)	RR 0.97 (0.88 to 1.07)	17,104 (25)	⊕⊕⊕⊕ High
*The risk in the intervention group (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; OR: odds ratio; RBC: red blood cell; RR: risk ratio
GRADE Working Group grades of evidence. High certainty: we are very confident that the true effect lies close to that of the estimate of the effect. Moderate certainty: we are moderately confident in the effect estimate: the true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different. Low certainty: our confidence in the effect estimate is limited: the true effect may be substantially different from the estimate of the effect. Very low certainty: we have very little confidence in the effect estimate: the true effect is likely to be substantially different from the estimate of effect.
^aWe downgraded once for inconsistency, as there was no consistency in the direction of the effect (despite the relatively low statistical heterogeneity), and we downgraded once for imprecision, as there were very low numbers of events. ^bDespite relatively low statistical heterogeneity, there was no consistency in the direction of the effect, hence we downgraded once for inconsistency. ^cDowngraded once for imprecision, as there were few events (and hence a wide CI).

Background

Description of the condition

Patients who are ill in hospital are frequently anaemic, with low haemoglobin concentrations. The causes of anaemia are diverse, including loss of blood from surgery, bleeding, excessive blood sampling for laboratory tests, or as a consequence of illness. Additionally, patients with cancer may develop anaemia because the underlying disease, or chemotherapy, affects production of red cells in their bone marrow. Anaemia both decreases the oxygen content of blood supplied to the tissues, including the myocardial muscles of the heart, and increases myocardial oxygen demand by requiring higher cardiac output to maintain adequate oxygen delivery throughout the body (Sabatine 2005).

It is well known that anaemia is linked with multiple clinical symptoms; it is also associated with worse outcomes among patients who are anaemic before and after surgery or critical illness, or who have cardiovascular disease (Carson 1996; Kunz 2020; Shander 2014). However, it does not necessarily follow that correction of anaemia will improve outcomes, whether by red blood cell (RBC) transfusions (addressed in this review) or by alternative treatments such as intravenous iron (Richards 2020). Anaemia is generally well tolerated by many people, therefore, the benefits of administering potentially corrective treatments such as red cell transfusion need to be weighed against the risks.

Description of the intervention

The main treatment option for raising the haemoglobin concentration rapidly in patients with anaemia is RBC transfusion. RBCs for transfusion are collected from whole blood donations from blood donors. These are centrifuged to concentrate them before they are added to anticoagulant and storage solutions. Autologous transfusions, which are collected from and stored for the same individual, are not indicated for sicker hospitalised patients with anaemia.

Red cell transfusions are life‐saving for patients with major bleeding. Red cell transfusions will treat severe anaemia successfully and may reduce the risks of major complications related to severe anaemia, such as myocardial infarction and heart failure. Uncertainties about the role of red cell transfusions are less clear for patients with less severe degrees of anaemia, and this is the focus of this review.

There are recognised risks of blood transfusion, as with any medical intervention (Delaney 2016). These risks, and the general availability of RBC transfusion vary throughout the world. In countries with well‐regulated blood supplies and effective blood donor screening policies, the safety of allogeneic red cell transfusion has improved significantly over the past 30 years, and overall risks are very low. These risks continue to be well monitored through national haemovigilance systems (e.g. the UK's Serious Hazards of Transfusion; SHOT Annual Report 2019), which document very few cases of transfusion‐transmitted infection; these findings reinforce earlier data from many countries (Zou 2009; Zou 2010). In resource‐limited countries, the supply of blood remains inadequate, with highly variable rates of donation per 1000 individuals. Furthermore, blood may not be as safe in these countries as it is in resource‐rich countries because it is not tested rigorously, and countries may lack quality control for viral pathogens, specifically transfusion‐transmissible infections such as HIV, hepatitis B, hepatitis C, and syphilis (WHO 2016). In some resource‐limited countries, a significant proportion of the blood supply is collected from family or paid blood donors ‐ not from voluntary unpaid donors ‐ and donor screening policies may not be efficiently applied. The prevalence of diseases such as HIV can be higher in low‐income countries than in high‐income countries, which presents a risk for transfusion transmission. All these points are described in the latest report on Global Blood Safety and Availability produced by the Blood Transfusion Safety Programme in the World Health Organization (WHO) Department of Service Delivery and Safety (WHO 2016).

Other risks of transfusion that have been described include acute transfusion reactions, volume overload, and transfusion‐related acute lung injury (Delaney 2016; SHOT Annual Report 2019; Toy 2012). Less well‐defined, but potentially important, adverse effects include loss of red cell nitric oxide production, which is thought to induce local vasodilatation; pro‐thrombotic effects from factors in the supernatant or changes in blood viscosity following transfusions; and immunomodulatory (or pro‐inflammatory) effects of different cellular products in the red cell component (Youssef 2017). Such harmful effects of RBC transfusions may be manifested, for example, as increased risks of infection (Rohde 2014).

Blood transfusion is expensive when one considers that around two million components (of which 1.6 million are units of RBCs) are issued across the UK alone each year (www.shotuk.org). The direct cost of each collected bag of red cells fails to capture the many associated costs related to hospital blood‐banking practice and safe patient administration (Stokes 2018). In 2008, the mean payment for one unit of leuco‐reduced RBCs in the USA was USD 223 (Whitaker 2011). However, when costs of administration as well as acquisition expenses of RBC transfusion were considered, the estimated cost derived from four USA and European hospitals rose to USD 761 per unit (standard deviation ± USD 294) (Shander 2010).

The impact of the storage age of red cells has been addressed in other systematic reviews (Shah 2018; Steiner 2015; Trivella 2019). Treatment options other than red cell transfusions for anaemia include erythropoietin and oral, or intravenous, iron therapy, which have been the topics of other recent trials and reviews (Richards 2020; Roman 2020).

How the intervention might work

The main clinical rationale for transfusing RBCs in anaemic patients is to improve oxygen delivery to tissue beds and vital organs such as the myocardium and brain. Transfusions may reduce compensatory work done by the heart to increase cardiac output in the face of anaemia. These benefits may manifest as better functional activity in patients and, ultimately, improved survival. Red blood cell transfusion is one of the few readily available treatments that consistently raises haemoglobin concentration and may restore tissue oxygenation adequately when oxygen demand exceeds supply (Wang 2010).

There is a long history of randomised controlled trials that have compared outcomes for participants allocated to different policies or schedules of red cell transfusion; these have now been completed and reported (Mueller 2019; NIH 1988). These studies presented results after randomising participants to either 'restrictive' triggers (where, typically, participants are transfused only when their haemoglobin concentration falls below 7.0 g/dL to 8.0 g/dL) or 'liberal' triggers (where participants are transfused at a higher haemoglobin concentration of around 9.0 g/dL to 10.0 g/dL). Historically, the widely accepted clinical standard was to transfuse patients when haemoglobin level dropped below 10.0 g/dL or when haematocrit fell below 30% (Wang 2010). Many guidelines based on the evolving evidence base now recommend that a range of haemoglobin values between 6.0 g/dL to 10.0 g/dL can be safely used for directing transfusions, depending on the presence of serious comorbidity (AAGBI 2008; ASA 2006; Carson 2012a; Carson 2016a; Mueller 2019; Napolitano 2009).

Why it is important to do this review

Much of the earlier evidence comparing restrictive and liberal thresholds for red cell transfusion comes from trials based in critical care. In 1999, the landmark TRICC trial (transfusion requirements in critical care) reported similar mortality in participants transfused at a restrictive trigger less than 7.0 g/dL compared with a liberal trigger less than 10.0 g/dL (Hébert 1999). The number of randomised trials continues to expand, as has been reported in previous iterations of this Cochrane Review (Carless 2010b; Carson 2012b; Hill 2000; Hill 2002; Hill 2005). By 2012, the number of participants enrolled in trials had doubled from 6264 to 12,587 (Carson 2012b); this number rose to 19,049 participants in a targeted update published in 2018, which specifically focused on patients with cardiovascular disease (Carson 2018). As further new trials continue to be published, there remains an ongoing need to update this systematic review, to ensure that new and updated guidelines on the use of red cell transfusions are based on the most recent literature reports of the effectiveness and safety of RBC transfusion (Carson 2016a). In addition, new studies focus on relevant and specific clinical contexts, for which previous levels of evidence for supporting best practice were very limited. This allows this updated review to inform transfusion practice in relevant subpopulations of patients.

The purpose of this updated review was to identify, appraise, and summarise the data from all randomised controlled trials (RCTs) that studied the clinical impact of varying thresholds for transfusion with RBCs. We remain interested in whether results of RCTs support the trend for increasingly restrictive RBC transfusion practices across different trial settings without harm to patients and to what extent RBCs need to be given more liberally in selected patient subgroups.

Objectives

The aim of this review update was to compare 30‐day mortality and other clinical outcomes for participants randomised to restrictive versus liberal red blood cell (RBC) transfusion thresholds (triggers) for all clinical conditions. The restrictive transfusion threshold uses a lower haemoglobin concentration as a threshold for transfusion (most commonly, 7.0 g/dL to 8.0 g/dL), and the liberal transfusion threshold uses a higher haemoglobin concentration to direct transfusion (most commonly, 9.0 g/dL to 10.0 g/dL).

Methods

Criteria for considering studies for this review

Types of studies

To examine evidence for the effects of transfusion thresholds on the use of red blood cell (RBC) transfusions and evidence for any change in clinical outcomes, we included randomised controlled trials (RCTs) in which comparison groups were assigned on the basis of a transfusion 'threshold' (sometimes termed a 'trigger'), defined as haemoglobin concentration or haematocrit level (with or without a specified level of haemodynamic instability) that had to be reached before RBC transfusion was administered. We required trials in which groups of participants were transfused with RBCs at higher haemoglobin or haematocrit levels (transfusion threshold) than those in a lower transfusion group, or were compared to those transfused in accordance with current standard transfusion practices. We excluded trials that were not designed to include any clinical outcomes relevant to this review.

Types of participants

We included trials of surgical or medical participants, involving adults or children, or both. We excluded studies enrolling neonates, given the distinct pathophysiology and clinical features of neonate anaemia, which is the topic of a separate Cochrane Review (Whyte 2011).

Types of interventions

The intervention considered was use of transfusion thresholds ('triggers') as a means of guiding allogeneic or autologous RBC transfusion, or both. A liberal transfusion threshold most often refers to transfusion of blood when the haemoglobin level falls below 9.0 g/dL to 10.0 g/dL. A restrictive transfusion threshold most often refers to transfusion of blood when the haemoglobin level falls below 7.0 g/dL to 8.0 g/dL.

We also included trials that compared transfusion and no transfusion while defining the no transfusion group as the restrictive strategy. Such trials may define a second threshold as a lower limit under which participants' haemoglobin should not fall without initiation of transfusion; this is consistent with all other trials in which clinical discretion is allowed for severe symptomatic anaemia.

Types of outcome measures

We evaluated clinical outcomes for efficacy, and we assessed complications of transfusion for safety.

Primary outcomes

The primary outcome for the analysis was 30‐day mortality. Mortality is a clinically relevant outcome that is widely cited in studies including patients with acute illness, critical illness, and perioperative care.

Secondary outcomes

We examined three categories of secondary outcomes:

mortality at different time intervals;
morbidity outcomes;
subgroups for mortality and morbidity.

We recorded and analysed mortality at different time points, including during hospital admission, at 90 days, and over the long term (median follow‐up, 3.1 years).

We evaluated morbidity that occurred during hospitalisation, including cardiac events (both as a composite outcome that included myocardial infarction, cardiac arrhythmias, cardiac arrest, pulmonary oedema, and angina, and individually when feasible), non‐fatal and fatal myocardial infarction, congestive heart failure, cerebral vascular accident (stroke), rebleeding, infection, thromboembolism, renal failure, mental confusion, function, and fatigue.

Infection was defined in three ways: sepsis or bacteraemia, pneumonia alone, or pneumonia plus wound infection. For the 2021 update, we added a specific outcome of 'transfusion‐specific reactions', as defined and reported in included studies. These events are uncommon, but they are important.

We defined all morbidity outcomes according to the definitions provided in individual trials. We evaluated subgroups based on transfusion thresholds and clinical context.

We recorded information on quality of life and functional outcomes. We also compared use of RBC transfusion as a measure of implementation of the transfusion intervention between groups by proportions of participants exposed to transfusion, units of blood transfused, and mean haemoglobin levels.

As this review is an update, we have continued to include some of these secondary outcomes for historical reasons. As stronger evidence is accrued, we believe that in future updates of this review, reporting of some of these outcomes may need to be modified or omitted.

Search methods for identification of studies

Electronic searches

We searched the following databases and ongoing trial registries:

CENTRAL (Cochrane Central Register of Controlled Trials; 2020, Issue 11), in the Cochrane Library (www.cochranelibrary.com);
MEDLINE via OvidSP (from 1946 to 16 November 2020);
Embase via OvidSP (from 1974 to 16 November 2020);
PubMed (for e‐publications ahead of print only, on 16 November 2020);
Transfusion Evidence Library (www.transfusionevidencelibrary.com; 1950 to 16 November 2020);
Web of Science Conference Proceedings Citations Index (CPCI‐S, 1990 to 16 November 2020);
US National Institutes of Health Ongoing Trials Register ClinicalTrials.gov (www.clinicaltrials.gov; searched to 16 November 2020);
World Health Organization International Clinical Trials Registry Platform (apps.who.int/trialsearch; searched to 16 November 2020).

We combined searches in MEDLINE and Embase with adaptations of the Cochrane RCT search filter as detailed in Chapter 6 of the Cochrane Handbook for Systematic Reviews of Interventions (Lefebvre 2011). We did not restrict our search by date, language, or publication status. We present search strategies for the 2012 update in Appendix 1, for the 2016 update and trial registries in Appendix 2 and Appendix 3, respectively, and for the 2020 update in Appendix 4.

Searching other resources

We checked the references of all identified trials, relevant review articles, and current treatment guidelines for further literature. We limited these searches to 'first‐generation' reference lists (i.e. reference lists of papers retrieved directly by database searches).

We contacted experts in the field to identify information relevant to the review. When possible and when necessary, we contacted authors of published studies for clarification of trial methods and data. We emailed all authors of trials that did not report our primary outcome of 30‐day mortality, but this was not possible for older trials for which contact information was not available. We searched the reference lists of relevant reviews and transfusion trials.

Data collection and analysis

Selection of studies

Two review authors (JLC and SJS) independently screened the titles or abstracts of the search results, or both, and selected trials that met the inclusion criteria. We resolved disagreements by discussion until we reached consensus. We identified trials in which participants were randomised to a restrictive transfusion strategy (transfusion threshold or protocol, or both) or to a control group that was randomised to a liberal transfusion strategy.

Data extraction and management

JLC and Paul Carless (a prior review author) extracted all data for earlier versions of this review. For this 2021 update, JLC and SJS independently extracted study characteristics and outcomes of new trials added since the last review, using a data extraction form. Information recorded on the extraction form included study type, presence of a transfusion threshold, transfusion protocol, type of surgery involved, clinical setting, treatment outcomes, and general comments, as well as details relevant to assessment of risk of bias for key domains described below. JLC entered data into Review Manager 5.4; NR checked data; JD added new items into tables to meet contemporary MECIR (Methodological Expectations for Cochrane Intervention Reviews) standards, which were checked by both JLC and SJS. We contacted authors of trials to request missing data.

We used the data extraction form to record data on the following outcomes:

number of participants exposed to allogeneic blood;
amount of allogeneic blood transfused;
number of participants receiving any transfusion (allogeneic blood, autologous blood, or both).

For trials involving surgical participants, we recorded the following:

postoperative complications (infection, haemorrhage, non‐fatal myocardial infarction, cardiac events, renal failure, stroke, thromboembolism, pulmonary oedema, mental confusion);
mortality, blood loss, haemoglobin and haematocrit levels (on admission, pre‐ and post‐transfusion, and at discharge);
demographics (age, sex);
type of surgery; and
medical condition.

We extracted data for allogeneic blood transfusion if it was expressed as packed RBCs. We documented information regarding the use of fresh frozen plasma or platelets, or both.

Assessment of risk of bias in included studies

We used the Cochrane tool for assessing risk of bias as described in Section 8.5 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011).

JLC, SJS, and JD assessed the following domains for each study:

sequence generation;
allocation concealment;
blinding (assessments were made separately with regard to objective (e.g. mortality) and subjective (e.g. self‐reported quality of life) outcomes);
incomplete outcome data;
selective outcome reporting; and
other potential sources of bias.

We completed a risk of bias table for each trial, incorporating a description of the trial's performance against each of the above domains and our overall judgement of the risk of bias for each entry as follows: 'low', 'unclear' (indicating unclear or unknown risk of bias), or 'high' risk of bias.

Measures of treatment effect

We obtained the risk ratio (RR) for allogeneic blood transfusion in the intervention group compared with the control group and corresponding 95% confidence intervals (CIs) for each trial. We adopted a similar approach for other outcomes of transfusion. When the event rate was low, we considered using the Peto odds ratio when criteria for this method were fulfilled. We also entered the mean number of units of RBCs transfused to each group and the corresponding standard deviations. We used the mean difference (MD) and 95% CI to express average mean reduction in the number of units of RBC administered to the intervention group compared with the control group.

Unit of analysis issues

The unit of analysis was the participant. In all trials except one (Jairath 2015), randomisation was done at the individual participant level. In this trial in people with gastrointestinal bleeding, randomisation was done at the level of the hospital (cluster), but analysis occurred at the level of the individual participant. The intraclass correlation coefficient (ICC) was very low (0.0001) for the outcome of mortality; therefore we included the data and considered the participant as the unit of randomisation and ignored the clustering. We performed a sensitivity analysis from which we excluded this trial, to see what effect, if any, this had on the analysis. We did not evaluate any outcomes with repeated measures.

Dealing with missing data

We performed all analyses on an intention‐to‐treat basis. We undertook no imputations for missing data. We received information on 30‐day mortality from three authors (DeZern 2016; Villanueva 2013; Webert 2008). Levels of missing data were never higher than 10%, which we consider acceptable.

Assessment of heterogeneity

We examined statistical heterogeneity using both the I² statistic and the Chi² test. The I² statistic describes the percentage of total variation across studies due to heterogeneity rather than chance. A value of 0% indicates no observed heterogeneity, and larger values show increasing heterogeneity; moderate or substantial heterogeneity is considered to exist when I² exceeds 50% or 85%, respectively (Higgins 2011). For the Chi² test, we used a P value < 0.10 to indicate the presence of statistically significant heterogeneity. Because of the anticipated significant clinical heterogeneity of trials, we analysed data using a random‐effects model. We also anticipated a high level of heterogeneity related to transfusion rates because practice in the different specialties of the trials would vary considerably according to specialty‐specific protocols. Therefore, as described later, we chose to provide a summary statistic for the outcomes of transfusion even when I² was very high, because of the clinically relevant information this provides.

Assessment of reporting biases

When more than 10 studies were available, we examined funnel plots for the primary outcome of 30‐day mortality and the proportion of participants transfused, to assess the potential for publication bias. We used the proportion of participants transfused because all trials reported this outcome, and this may reflect overall risk of publication bias better than 30‐day mortality, which was not reported in all of the trials. We sought evidence of selective outcome reporting by comparing plans from described registrations/protocols (when available) with final reports.

Data synthesis

We performed all analyses using Review Manager 5.4 software (Review Manager 5a). We entered data for numbers of participants exposed to red cell transfusions, anticipated to be allogeneic blood in most trials and patients. We present the results using haemoglobin concentration in grams per decilitre (g/dL). Based on study reporting, we converted haematocrit to haemoglobin concentration by dividing by three. When studies presented transfusion volume as millilitres (mL), we converted these amounts to units by dividing by 300 (as in most countries, a standard unit of red blood cells is 300 mL). We pooled data for all outcomes and presented data stratified by subgroups for the primary outcome of 30‐day mortality and proportion of participants transfused by using a random‐effects model (Der Simonian 1986), and we presented the pooled result along with its 95% CI. We used Peto odds ratios for outcomes with event rates less than 1%. For continuous variables, we estimated the pooled mean difference and the 95% CI by using the generic inverse variance method.

Subgroup analysis and investigation of heterogeneity

Prespecified subgroups, as established in prior reviews, consisted of the following clinical contexts:

acute blood loss/trauma;
cancer;
cardiac surgery;
critical care;
orthopaedic surgery;
myocardial infarction;
vascular surgery; and
haematological malignancy.

We examined 30‐day mortality and the proportion of participants exposed to transfusion stratified by the transfusion threshold (difference between liberal and restrictive transfusion thresholds: ≥ 2.0 g/dL and < 2.0 g/dL; and restrictive transfusion threshold < 7.0 g/dL versus one of 8.0 g/dL to 9.0 g/dL). We also examined a post hoc subgroup of enrolled participants with myocardial infarction compared with all other clinical specialties, and we combined cardiac surgery with myocardial infarction because of emerging evidence that participants with acute myocardial infarction might differ from other anaemic participants (Carson 2013).

For the primary outcome of 30‐day mortality, we also compared findings between prospectively registered trials and those that were unregistered, or were registered long after recruitment began. Blood components are not subject to the same legal requirements for prospective registration as medical devices or pharmaceutical interventions. As in prior versions of this review, we did not exclude unregistered trials published after 2010 (as per the Cochrane Injuries Group policy), and we did conduct analyses to consider differential impact of the results of all trials for which proof of prospective registration could (or could not) be confirmed.

Sensitivity analysis

We performed a sensitivity analysis to assess effects of studies with high risk of bias for allocation concealment for the primary outcome; however, as in earlier versions of the review, sensitivity analyses for secondary outcomes were not informative. We repeated the analysis while excluding the cluster randomised trial (Jairath 2015).

Summary of findings and assessment of the certainty of the evidence

We have presented judgements about the quality of evidence in a summary of findings table (according to guidelines developed by the GRADE Working Group) (Schünemann 2011). We rated the quality of evidence as 'high', 'moderate', 'low', or 'very low', according to the following five GRADE domain considerations of: risk of bias, inconsistency, indirectness, imprecision, and publication bias.

This table includes the following outcomes:

number of people receiving blood transfusions;
30‐day mortality;
myocardial infarction;
congestive heart failure;
cerebrovascular accident (stroke);
rebleeding; and
thromboembolism.

Results

Description of studies

Details of the selection process for, and characteristics of, the included studies are offered below, along with information about interventions and trial design.

Results of the search

See Figure 1 for the PRISMA flowchart describing trial selection for the present update.

Figure 1

Flow of studies for 2021 update

In the previous review, published in 2016, we included 31 studies.

For this 2021 update, we identified and analysed 17 additional trials (Akyildiz 2018; Bergamin 2017; Ducrocq 2021; Gillies 2020; Gobatto 2019; Hoff 2011; Jansen 2020; Koch 2017; Kola 2020; Laine 2018; Mazer 2017; Møller 2019; Palmieri 2017; Robitaille 2013; Stanworth 2020; Tay 2020; Yakymenko 2018) (Figure 1), leading to a total of 48. We identified one of these trials ‐ Ducrocq 2021 ‐ as an ongoing trial when we searched trials registers in November 2020; it has subsequently been published and is included in the analysis. We reviewed eligibility for one analysis after peer review and included data from it only in a narrative synthesis (Hoff 2011). This analysis treated the outcomes of two linked studies (DAHANCA 5 and DAHANCA 7) as a single trial (Hoff 2011a; Hoff 2011b; Overgaard 1998; Overgaard 2003).

Many of the included trials have been reported within multiple papers, which are included as secondary references. Whilst the focus of this review was the main (first) report of outcome data, reports of secondary or subgroup analyses (e.g. long‐term outcomes) occasionally offered complementary information useful for data extraction and assessment of bias.

Included studies

Participants

This updated systematic review includes a total of 21,433 trial participants (at baseline) across 48 trials described in 70 publications. By comparison, in the 2016 Cochrane review (Carson 2016b), we reported on an analysis of 31 trials that enrolled 12,587 participants.

The clinical context of the 48 trials was varied:

11 studies involved orthopaedic surgery (Carson 1998; Carson 2011; Fan 2014; Foss 2009; Gillies 2020; Gregersen 2015; Grover 2005; Lotke 1999; Nielsen 2014; Parker 2013; So‐Osman 2013);
seven involved critical care (de Almeida 2015; Gobatto 2019; Hébert 1995; Hébert 1999; Holst 2014; Palmieri 2017; Walsh 2013);
six examined acute blood loss (Blair 1986; Jairath 2015; Kola 2020; Prick 2014; Topley 1956; Villanueva 2013), four of which concerned gastrointestinal bleeding, one postpartum haemorrhage (Prick 2014), and one trauma (Topley 1956);
eight included cardiac surgery (Bracey 1999; Hajjar 2010; Johnson 1992; Koch 2017; Laine 2018; Mazer 2017; Murphy 2015; Shehata 2012);
three involved acute coronary syndrome (Carson 2013; Cooper 2011; Ducrocq 2021);
eight involved cancer, leukaemia, and haematological malignancies (Bergamin 2017; DeZern 2016; Hoff 2011; Jansen 2020; Stanworth 2020; Tay 2020; Webert 2008; Yakymenko 2018);
two were in vascular surgery (Bush 1997; Møller 2019);
three enrolled paediatric participants (Akyildiz 2018; Lacroix 2007; Robitaille 2013). Two trials were conducted in paediatric intensive care units (Akyildiz 2018; Lacroix 2007), and one trial involved bone marrow transplant recipients (Robitaille 2013).

Interventions

We noted variation in the definitions of transfusion strategies specified in the protocols, but most commonly, haemoglobin concentrations were used as 'triggers'. Four trials specified haematocrit values for the threshold (Cooper 2011; Hajjar 2010; Koch 2017; Johnson 1992). Four trials incorporated symptoms in addition to haemoglobin threshold in the restrictive transfusion strategy (Carson 2011; Carson 2013; Parker 2013; Prick 2014).

Transfusion thresholds by haemoglobin concentration in restrictive transfusion arms (44 trials) varied from 7.0 g/dL to 9.7 g/dL. The most common restrictive haemoglobin threshold for interventions was between 7.0 g/dL to 8.0 g/dL (35 trials). Two trials recruited patients in the outpatient chronic transfusion‐dependent population setting based on haemoglobin concentrations (Jansen 2020; Stanworth 2020), and thresholds for the intervention arms in these trials were higher, as might be expected for this population. Three trials defined a no‐transfusion strategy for the 'restrictive' arm (Hoff 2011; Parker 2013; Prick 2014), with provisions made for participants with clear signs of anaemia.

Restrictive haematocrit varied between 24% and 25% (equivalent to haemoglobin levels of around 8 g/dL) (Cooper 2011; Hajjar 2010; Johnson 1992; Koch 2017).

The most common transfusion threshold by haemoglobin concentration in the liberal transfusion arm was 9.0 g/dL to 10.0 g/dL. However, the liberal transfusion threshold varied and included:

100% of 'normal red cell volume' (Topley 1956);
two units of blood irrespective of clinical state (immediately in one trial (Blair 1986), postoperatively in another (Lotke 1999)); and
transfusion sufficient to maintain haemoglobin levels:
- above 10 g/dL (Gregersen 2015; Hoff 2011; Jansen 2020; Robitaille 2013; Stanworth 2020; Webert 2008; Yakymenko 2018);
- at 10 g/dL (Bush 1997; Carson 1998; Carson 2011; Carson 2013; Ducrocq 2021; Foss 2009; Grover 2005; Hajjar 2010; Hébert 1995; Hébert 1999; Jairath 2015; Parker 2013);
- at 9.5 g/dL (Lacroix 2007; Shehata 2012);
- at 9 g/dL (Bergamin 2017; Bracey 1999; de Almeida 2015; Gillies 2020; Gobatto 2019; Holst 2014; Murphy 2015; Tay 2020; Villanueva 2013; Walsh 2013);
- at 8.9 g/dL (Prick 2014); and
- at 8 g/dL (DeZern 2016; Kola 2020; Møller 2019).

Four trials used haematocrit levels when determining triggers (Cooper 2011 and Hajjar 2010 specified the liberal triggers as haematocrit levels of 30%; Koch 2017 specified a level of 28%, and Johnson 1992 a level of 32%).

Trial setting and design

See Table 1.

Open in table viewer

Table 1. Trial setting details

Study ID	Number of participants at baseline	Country/Countries	Number of sites	Setting(s)	Year recruitment started
Mazer 2017	5092	19 countries^a	73	73 sites ‐ varied	2014
Carson 2011	2016	USA, Canada	47	47 sites ‐ varied	2004
Murphy 2015	2003	UK	17	17 sites ‐ varied	2009
Holst 2014	1005	Denmark, Sweden, Norway, Finland	32	32 general ICUs	2011
Jairath 2015	936	UK	6	University teaching hospitals	2012
Villanueva 2013	921	Spain	1	General hospital	2003
Hébert 1999	838	Canada	25	Tertiary (22), community ICU (3)	1994
Koch 2017	722	USA (1), India (1)	2	1 academic affiliated hospital in the USA, a private hospital in India	2007
Ducrocq 2021	668	France, Spain	35	35 sites ‐ varied	2016
Lacroix 2007	648	Canada, Belgium, USA, UK	19	Tertiary paediatric ICU	2001
So‐Osman 2013	603	Netherlands	3	Varied ‐ university and general hospitals	2001
Prick 2014	519	Netherlands	37	Varied ‐ university and general hospitals	2004
Hajjar 2010	512	Brazil	1	University teaching hospital	2009
Hoff 2011	466	Denmark	??	Oncology centres	1986
Bracey 1999	428	USA	1	University teaching hospital	1997
Palmieri 2017	345	US (16 sites), Canada (1), New Zealand (1)	18	Specialist burn centres	2010
Tay 2020	300	Canada	4	HCT sites	2011
Bergamin 2017	300	Brazil	1	University teaching hospital	2012
Gregersen 2015	284	Denmark	1	University teaching hospital	2010
Grover 2006	260	UK	3	Acute hospitals	Not stated
Kola 2020	224	India	1	Tertiary hospital	2015
Parker 2013	200	UK	1	General hospital	2002
de Almeida 2015	198	Brazil	1	Tertiary oncology university hospital	2012
Fan 2014	192	China	1	University teaching hospital	2011
Akyildiz 2018	180	Turkey	1	University teaching hospital	2014
Yakymenko 2018	133	Denmark	1	University teaching hospital	2010
Lotke 1999	127	USA	1	University teaching hospital	Not stated
Foss 2009	120	Denmark	1	University teaching hospital	2004
Carson 2013	110	USA	8	8 sites ‐ varied	2010
Walsh 2013	100	UK	6	Varied ‐ university and general hospitals	2009
Bush 1997	99	USA	1	University teaching hospital	1995
DeZern 2016	89	USA	1	Tertiary referral centre for oncology	2014
Carson 1998	84	USA (3), UK (1)	4	University teaching hospitals	1996
Laine 2018	80	Finland	1	University teaching hospital	2014
Hébert 1995	69	Canada	5	Tertiary hospitals	1993
Nielsen 2014	66	Denmark	2	University teaching hospital and general hospital	2009
Gillies 2020	62	UK	1	University teaching hospital	2017
Webert 2008	60	Canada	4	Tertiary oncology centres	2003
Møller 2019	58	Denmark	1	General hospital	2015
Shehata 2012	50	Canada	1	University teaching hospital	2007
Blair 1986	50	UK	1	University teaching hospital	Not stated
Gobatto 2019	47	Brazil	1	University teaching hospital	2014
Cooper 2011	45	USA	2	Veterans' Affairs hospital centres	2003
Johnson 1992	39	USA	1	University teaching hospital	Not stated
Stanworth 2020	38	UK, Australia, New Zealand	12	12 sites ‐ varied	2015
Topley 1956	22	UK	1	'Accident hospital'	Not stated
Januarysen 2020	19	Netherlands	3	1 university hospital, 2 general hospitals	2002
Robitaille 2013	6	Canada	1	Not identified	2009

^aMazer 2017 (TRICS‐III): majority of sites in USA; sites also in Australia, Brazil, Canada, China, Colombia, Denmark, Egypt, Germany, Greece, India, Israel, Malaysia, New Zealand, Romania, Singapore, South Africa, Spain, and Switzerland.

The included studies were conducted at a total of nearly 400 sites within 26 countries. High‐income countries including Canada, the UK, and the USA contributed the bulk of both single‐site and multicentre studies, as well as co‐ordinating international multicentre studies. The next most common countries, in terms of providing settings for eligible trials, were Denmark, the Netherlands, Brazil, and France. Recruitment start dates for studies included within this review ran between 1955 and 2017, with a marked increase in the rate of new studies commencing recruitment from 2009 onwards.

A total of 24 of the 48 included studies were unregistered or were registered by investigators long after recruitment began. Although a majority of unregistered trials were relatively old, lack of prospective registration is a problem that persists to the present day.

In 47 of the 48 trials, the participant was the unit of randomisation and analysis. One trial used cluster randomisation by hospital (Jairath 2015). Sample sizes of included studies varied enormously (from 6 to 5092 participants randomised at baseline). Twenty‐six trials included 100 or more participants, and four trials included over 1000 participants each (Carson 2011; Holst 2014; Mazer 2017; Murphy 2015). Eleven of the included studies were described as pilot or feasibility studies (Carson 2013; DeZern 2016; Gillies 2020; Gobatto 2019; Hébert 1995; Jairath 2015; Møller 2019; Shehata 2012; Stanworth 2020; Webert 2008; Yakymenko 2018). We counted two linked studies in patients with head and neck squamous cell carcinoma before radiotherapy as one trial for the purpose of this review (Hoff 2011); the two component studies, DAHANCA 5 and DAHANCA 7, tested the same main trial intervention (nimorazole) and then applied a similar subrandomisation question to evaluate transfusion versus no transfusion, given concerns about poorer responses to radiation therapy due to a hypothesis of hypoxia‐induced radio‐resistance.

Excluded studies

In 2016, this review contained records of four excluded studies. In this 2021 update, we have excluded a further 17 studies; data for these studies were published in 22 publications (see Characteristics of excluded studies). Of the 21 excluded studies, 17 are ineligible RCTs (excluded largely on grounds of intervention or population); the remaining four are non‐randomised studies of different designs.

Studies awaiting classification

Brief details of five trials that are awaiting assessment are shown in the Studies awaiting classification section. Four have been completed but remain unpublished; we are considering how to handle data reported in the fifth (published) trial, which was of a complex, multifactorial design.

Ongoing studies

Brief details of 14 ongoing studies identified from searches of international trial registers are shown in the Ongoing studies section. When completed, and if eligibility criteria remain stable, results from these studies may add data from approximately 14,880 participants to this review, with five trials aiming to recruit over 1000 participants each. The latter (larger) studies are focusing on populations that are currently under‐represented in the studies included in this review, specifically, those with traumatic brain injury or cardiac/vascular disease.

Risk of bias in included studies

The risk of bias tables detail the assessment of studies for each domain and are summarised in Figure 2 and Figure 3.

Figure 2

'Risk of bias' graph: review authors' judgements about each 'Risk of bias' item presented as percentages across all included trials. Forty‐eight trials are included in this review.

Figure 3

'Risk of bias' summary: review authors' judgements about each methodological quality item for each included trial

Allocation

Sequence generation

We judged 41 trials to be at low risk of bias for this domain. We judged three trials to be at high risk of bias: one for basing the randomisation sequence on hospital record number, one for using coin‐tossing, and one because it mentioned using both a table of random numbers and odd/even (restrictive/liberal) allocation strategy. The remaining four trials presented insufficient information for us to be able to assess the adequacy of sequence generation, so we rated them as being at unclear risk.

Allocation concealment

We judged the risk of bias for this item to be low for 36 trials that used central allocation or sealed envelopes if appropriate safeguards (e.g. sequentially numbered envelopes) were used. We judged four trials to be at high risk of bias: one of these trials used a cluster design, so everyone in all hospitals knew to which group all participants had been assigned (Jairath 2015), one used a coin toss, one used hospital numbers that could be seen, and one used closed envelopes. We rated the risk for eight trials as unclear because the publications did not provide any information about how allocation was concealed.

Blinding

Performance bias

The nature of the intervention meant that blinding of clinicians involved in the care and administration of blood transfusions would not have been possible. Blinding of personnel for this intervention is also not feasible. In our view, for objective outcomes such as mortality (the primary outcome used within this review), it is appropriate to assess risk of bias as low.

Detection bias

Outcomes are assessed optimally when assessors are blinded to assignment. It is possible to blind the assessment of many outcomes by using, for example, an adjudication committee. In contrast, for some outcomes such as death, blinded assessment is less relevant. We classified risk of bias on the basis of the primary outcome of the trial (mortality) and on subjective outcomes, if reported, including functional measures and quality of life. We judged the risk of bias to be high for 11 trials for subjective outcomes.

Incomplete outcome data

We rated seven trials as being at high risk of bias for this domain, as data were missing for a large proportion of participants (20% to 45% of data for an outcome important to this review in six cases) or were missing disproportionately between arms (one trial).

Selective reporting

We rated 20 trials as being at unclear risk of bias for this domain, largely because evidence of prospective registration could not be confirmed. One trial (the oldest in the review, which recruited in the early 1950s) was assessed as being at high risk of bias for not reporting the groups in which deaths occurred. The remaining trials were assessed as having a low risk of bias for this domain.

Other potential sources of bias

We identified few other sources of bias. Small trials, including feasibility or pilot studies (which account for 20% of included trials), often reported small imbalances at baseline, as might be expected. Some trials were obliged to terminate prematurely due to slow recruitment. Only a limited number of trials described protocol violations for transfusions in detail, but these applied to both intervention arms. Overall, we assessed six of the 48 trials as having unclear risk of bias for this domain.

Effects of interventions

See: Summary of findings 1 Liberal compared with restrictive transfusion protocols for guiding red blood cell transfusion

Substantial variation in outcomes was reported in the included trials, which, in part, reflects their clinical settings.

Nearly all trials contributed to the analysis comparing the proportion of participants transfused in liberal and restrictive transfusion groups. Despite the heterogeneous methods and transfusion triggers reported in these RCTs, it was possible to pool data, to varying degrees, for each of the review outcomes. See summary of findings Table 1.

Primary outcome

30‐Day mortality

The primary outcome of 30‐day mortality was reported by 31 trials (including 16,729 participants) in a form suitable for meta‐analysis. There was no difference in 30‐day mortality between restrictive and liberal transfusion strategies (risk ratio (RR) 0.99, 95% confidence interval (CI) 0.86 to 1.15; Analysis 1.1). Heterogeneity between these trials was not important (Chi² = 40.06, degrees of freedom (df) = 28 (P = 0.07); I² = 30%). The funnel plot demonstrates that the RR for 30‐day mortality is symmetrically distributed, which indicates there is not likely to be publication bias for this outcome (Figure 4).

Figure 4

Funnel plot of comparison: 1 Mortality, outcome: 1.1 30‐Day mortality

Subgroup analysis of 30‐day mortality: restrictive threshold of 7.0 g/dL to 7.5 g/dL versus 8.0 g/dL to 9.0 g/dL

We examined 30‐day mortality and stratified it by the restrictive transfusion threshold used in the trials. Fifteen trials with 11,572 participants used a 7.0‐g/dL restrictive threshold. The RR for 30‐day mortality was 1.00 (95% CI 0.83 to 1.19; Analysis 1.2). Sixteen trials with 5157 participants used a restrictive threshold of 8.0 g/dL to 9.0 g/dL. The RR for 30‐day mortality was 0.97 (95% CI 0.75 to 1.24; Analysis 1.2). The test for subgroup differences did not show any differences between subgroups (Chi² = 0.04, df = 1 (P = 0.83), I² = 0%), indicating there was no difference in the mortality risk between the two thresholds.

Subgroup analyses of 30‐day mortality: clinical context

We examined 30‐day mortality and stratified it by the clinical context used in the trials: cardiac surgery, orthopaedic surgery, vascular surgery, acute blood loss or trauma (analyses for this grouping for 30‐day mortality included gastrointestinal (GI) bleeding only), critical care, acute myocardial infarction, and haematological malignancies. The overall RR for 30‐day mortality stratified by clinical specialty was 0.99 (95% CI 0.86 to 1.14; 31 trials, 16,729 participants; Analysis 1.3). There were no differences in 30‐day mortality between subgroups (Chi² = 6.73, df = 6 (P = 0.35); I² = 10.9%).

Cardiac surgery

Four trials conducted in 7411 patients undergoing cardiac surgery reported 30‐day mortality. The RR for 30‐day mortality for a restrictive compared to a liberal transfusion strategy was 0.99 (95% CI 0.74 to 1.33; Analysis 1.3.1).

Orthopaedic surgery

Eight trials of orthopaedic surgery contributed data from 3111 participants for 30‐day mortality. There was no clear effect of a restrictive compared to a liberal transfusion threshold (RR 1.16, 95% CI 0.75 to 1.79; Analysis 1.3.2).

Vascular surgery

Two trials contributed data from 157 participants for 30‐day mortality. The RR for 30‐day mortality was 0.98 (95% CI 0.30 to 3.25; Analysis 1.3.3).

Acute blood loss or trauma

Three trials reported mortality at 30 days among 1522 participants with acute blood loss or trauma (GI bleeding). Mortality was significantly lower when a restrictive strategy rather than a liberal strategy was used (RR 0.65, 95% CI 0.43 to 0.97; Analysis 1.3.4).

Critical care

Nine trials including 3529 participants receiving critical care for heterogeneous reasons contributed data for this outcome. The RR showed no clear effect of a restrictive compared to a liberal transfusion strategy (RR 1.06, 95% CI 0.85 to 1.32; 9 trials, 3529 participants; I² = 55%; Analysis 1.3.5).

Acute myocardial infarction

Three trials provided data from 820 participants with acute myocardial infarction and evaluated mortality; for this subgroup, mortality risk was higher in the restrictive strategy group than in the liberal strategy group (RR 1.61, 95% CI 0.38 to 6.88, Analysis 1.3.6). We carried out a post hoc subgroup analysis that compared 30‐day mortality for acute myocardial infarction participants versus all other participants but found no differences. The P value for subgroup differences was 0.50 (Chi² = 0.45, df = 1; I² = 0%; Analysis 1.4). Although we observed a moderately elevated RR for myocardial infarction participants (RR 1.61), the three included trials were modest in size, and hence, the pooled 95% confidence interval is very wide.

Haematological malignancies

Two small trials provided data on 30‐day mortality among 149 participants. The 95% confidence interval was very wide, and no conclusions can be drawn for this subgroup (RR 0.37, 95% CI 0.07 to 1.95; 2 trials, 149 participants; I² = 0%; Analysis 1.3.7).

Mortality by cardiac surgery, vascular surgery, myocardial infarction, and all others

We examined 30‐day mortality and stratified it by the clinical context used in trials in a grouping comparing cardiac surgery, vascular surgery, myocardial infarction, and a group combining all other included trials. The overall RR for 30‐day mortality stratified by clinical specialty was (to repeat findings above) 0.99 (95% CI 0.86 to 1.14; 31 trials, 16,729 participants; Analysis 1.3). Again there were no differences in 30‐day mortality (test for subgroup differences: Chi² = 0.43, df = 3 (P = 0.93), I² = 0%; Analysis 1.5).

Subgroup analysis of 30‐day mortality: prospectively registered versus unregistered trials or trials for which registration was post hoc

We stratified 30‐day mortality according to whether or not trials were prospectively registered. Of the 31 trials that contributed data to our primary outcome, 18 (with 12,932 participants) were prospectively registered. The RR for 30‐day mortality provided by pooling data from these trials was 1.08 (95% CI 0.89 to 1.31). Pooling of the 13 unregistered trials (3797 participants) led to a RR of 0.81 (95% CI 0.66 to 1.00). The test for subgroup differences indicated a difference between subgroups: Chi² = 4.06, df = 1 (P = 0.04), I² = 75.4%; Analysis 2.1), but in neither group was there a clear effect for either transfusion strategy.

Sensitivity analysis

There were no differences in 30‐day mortality between trials with low versus unclear or high risk of bias in one bias domain (i.e. allocation concealment) (Analysis 3.1). The RR was 1.01 (95% CI 0.87 to 1.18) in trials with low risk of bias for allocation concealment and 0.84 (95% CI 0.51 to 1.39) for trials with unclear or high risk of bias for allocation concealment. Testing for subgroup differences yielded the following: Chi² = 0.47, df = 1 (P = 0.49); I² = 0%.

Secondary outcomes

As detailed below, none of the other analyses on mortality or morbidity showed differences between the groups compared.

Mortality at other time intervals

We analysed mortality at hospital discharge (15 trials; 6597 participants; Analysis 4.1), at 90 days (7 trials, 4143 participants; Analysis 4.2), and at six months or longer (2 trials, 4702 participants; Analysis 4.3).

There were no differences in mortality between transfusion strategies at hospital discharge (RR 0.86, 95% CI 0.72 to 1.03; Chi² = 15.36, df = 13 (P = 0.29); I² = 15%), but the 90‐day mortality was higher for the restrictive strategy (RR 1.13, 95% CI 1.02 to 1.25; Chi² = 5.28, df = 6; P = 0.41; I² = 0%).

The two largest included trials (Carson 2011; Mazer 2017), reported mortality at six months or beyond in separate publications (Carson 2015; Mazer 2018). Results suggest no clear differences (RR 0.98, 95% CI 0.79 to 1.22; P = 0.84; Analysis 4.3). Both trials employed similar transfusion strategies.

The results of mortality analyses at hospital discharge, at 30 days, and at six months are consistent. The results of mortality analyses at 90 days were gathered from a smaller number of participants and are dominated by two particular trials (Bergamin 2017; Holst 2014), limiting interpretation.

Complex analysis in Hoff 2011 had the main purpose of defining a role for the drug nimorazole in patients receiving radiotherapy with head and neck squamous cell carcinoma (HNSCC); additional randomisation steps addressed the value of transfusion in participants who had low haemoglobin levels preradiation. This analysis combined data from a trial comparing the drug with placebo and from another comparing drug delivery at different intervals. We could not incorporate five‐year mortality data within our meta‐analysis, but investigators found that although transfusion improved haemoglobin levels before and during radiation treatment, it did not improve other outcomes for patients and may have had a negative impact on survival.

Clinical outcomes

We noted no differences in any groups compared for any of the clinical outcomes.

Cardiac events

Eleven trials reported data on post‐enrolment cardiac events in 5577 participants. Risks of cardiac events (myocardial infarction, cardiac arrhythmias, cardiac arrest, pulmonary oedema, and angina) were not increased by the use of restrictive transfusion strategies (RR 1.03, 95% CI 0.80 to 1.32; Analysis 5.1). Heterogeneity between these trials was moderate (Chi² = 24.09, df = 10 (P = 0.007); I² = 58%). It is possible that participants were counted in more than one category for this composite outcome because these disorders are clinically inter‐related (e.g. a participant could have angina that might lead to pulmonary oedema).

Myocardial infarction

Twenty‐three trials reported data for myocardial infarction (fatal and non‐fatal) for 14,730 participants after random allocation to liberal or restrictive transfusion arms. There was no difference between restrictive and liberal transfusion strategies (RR 1.04, 95% CI 0.87 to 1.24; Analysis 5.2). We found no evidence of heterogeneity between trials (Chi² = 18.63, df = 21; P = 0.61; I² = 0%).

Congestive heart failure

Sixteen trials reported data for congestive heart failure in 7247 participants. There was no difference between restrictive and liberal transfusion strategies (RR 0.83, 95% CI 0.53 to 1.29; Analysis 5.3). Heterogeneity between trials was moderate (Chi² = 22.06, df = 13; P = 0.05; I² = 41%).

Cerebrovascular accident: stroke

Nineteen trials reported data for stroke in 13,985 participants. There was no difference between restrictive and liberal transfusion strategies (RR 0.84, 95% CI 0.64 to 1.09; Analysis 5.4). Heterogeneity between trials was not important (Chi² = 12.80, df = 18; P = 0.80; I² = 0%).

Rebleeding

Eight trials reported data for rebleeding in 3412 participants. There was no difference between restrictive and liberal transfusion strategies (RR 0.80, 95% CI 0.59 to 1.09; Analysis 5.5). Heterogeneity between trials was not important (Chi² = 12.24, df = 7; P = 0.09; I² = 43%).

Sepsis/bacteraemia

Nine trials reported data for sepsis/bacteraemia in 4352 participants. There was no difference between restrictive and liberal transfusion strategies (RR 1.06, 95% CI 0.86 to 1.30; Analysis 5.6). Heterogeneity between these trials was not important (Chi² = 8.56, df = 7; P = 0.29; I² = 18%).

Pneumonia

Sixteen trials reported data for pneumonia in 6666 participants. There was no difference between restrictive and liberal transfusion strategies (RR 0.97, 95% CI 0.84 to 1.13; Analysis 5.7). Heterogeneity between these trials was not important (Chi² = 11.48, df = 15; P = 0.72; I² = 0.0%).

Infection

Twenty‐five trials including 17,104 participants reported data for all infections defined as sepsis/bacteraemia, pneumonia, and wound infection. There was no difference between restrictive and liberal transfusion strategies (RR 0.97, 95% CI 0.88 to 1.07; Analysis 5.8). Heterogeneity between these trials was not important (Chi² = 21.42, df = 14; P = 0.09; I² = 35%).

Thromboembolism

Thirteen trials reported data for thromboembolism for 4201 participants. We calculated the odds ratio using the Peto method because the risk of thromboembolism was less than 1%. There was no difference between restrictive and liberal transfusion strategies (Peto odds ratio 1.11, 95% CI 0.65 to 1.88; Analysis 5.9). Heterogeneity between these trials was not important (Chi² = 14.48, df = 11; P = 0.21; I² = 24%).

Renal failure

Fifteen trials reported data on renal failure in 12,531 participants. There was no difference between restrictive and liberal transfusion strategies (RR 1.03, 95% CI 0.92 to 1.16; Analysis 5.10). Heterogeneity between these trials was not important (Chi² = 12.77, df = 14; P = 0.55%; I² = 0%).

Mental confusion

Nine trials reported data for mental confusion in 6442 participants. There was no difference between restrictive and liberal transfusion strategies (RR 1.11, 95% CI 0.88 to 1.40; Analysis 5.11). Heterogeneity between these trials was not important (Chi² = 10.29, df = 8; P = 0.24; I² = 22%).

Functional recovery and fatigue

In total, 24 trials reported results showing differing scores for functional and mental outcomes and fatigue (Bracey 1999; Carson 1998; Carson 2011; de Almeida 2015; DeZern 2016; Fan 2014; Foss 2009; Gillies 2020; Gobatto 2019; Gregersen 2015; Hajjar 2010; Jairath 2015; Jansen 2020; Koch 2017; Lotke 1999; Murphy 2015; Nielsen 2014; Parker 2013; Prick 2014; So‐Osman 2013; Stanworth 2020; Tay 2020; Walsh 2013; Yakymenko 2018). However, there was considerable heterogeneity in the methods and questionnaires used, and in the timing of assessments, which precluded meta‐analysis. No larger trials reported significant differences in trial‐specific functional or mental outcomes, or fatigue. Exploratory findings for quality of life measures from two trials in a haematological setting showed possible beneficial effects of liberal transfusion, but these results need to be evaluated in larger trials (Stanworth 2020; Yakymenko 2018). Three trials reported functional outcomes for orthopaedic surgery participants, but assessment of these functional measures in different ways precluded pooling in a meta‐analysis.

Other outcomes ‐ blood transfusions and haemoglobin

Results for transfusion and haemoglobin data were presented across the included trials, and provide key information about the implementation of transfusion protocols in trials. We anticipated high levels of heterogeneity in the analysis of transfusion outcomes, for several reasons. In particular, standard 'control' rates of transfusion practice are highly variable across the clinical specialties in which trials were identified for this update. These differing policies regarding rates of transfusion reflect practice defined in specialty guidelines and recommendations. It is usually recommended that pooled estimates are not presented when heterogeneity is so high. However, we present the pooled results here, as there was consistency regarding the direction of effect; further justification for this is provided in the Discussion.

Proportion of participants transfused

This analysis demonstrates differences in the proportions of participants transfused with RBCs in the liberal and restrictive trial arms. Data on the proportions of transfused participants were available from 42 trials (20,057 participants). The implementation of a restrictive transfusion trigger across all trials reduced the relative risk of receiving at least one RBC transfusion by 41% (RR 0.59, 95% CI 0.53 to 0.66; Analysis 6.1). Heterogeneity between these trials was substantial (Chi² = 1104.24, df = 41 (P < 0.00001); I² = 96%); however, there was consistency in the direction of the effect.

The proportions of participants transfused in liberal and restrictive trial arms were very different across different clinical contexts (Analysis 6.2); differences between subgroups were manifest (Chi² = 25.33, df = 6 (P = 0.0003); I² = 76.3%). There was a tendency for great variation within subgroups also (e.g. in the subgroup of critical care trials, heterogeneity was high (I² = 92%)). The acute blood loss/trauma subgroup included diverse underlying illnesses for haemorrhage, including comorbidities, leading to an I² of 96%. For example, Prick 2014 recruited young (otherwise healthy) women with postpartum haemorrhage, and Jairath 2015 enrolled older participants with gastrointestinal bleeding, characterised by many comorbidities. Prick 2014 contributed to a large extent to the high heterogeneity in this subgroup, and temporarily removing it from the analysis reduced heterogeneity to 77%. By contrast, participants enrolled in the subgroup of cardiac surgery trials demonstrated less variability in risk of transfusion across trials, and in this subgroup, we observed no important heterogeneity. This sensitivity analysis, although post hoc, highlights how transfusion policies in this setting differed from adult protocols in a critical care setting.

When the difference in haemoglobin thresholds between restrictive and liberal arms was 2.0 g/dL or more, the RR of transfusion was 0.57 (95% CI 0.50 to 0.64; Analysis 6.3.1), which means there was a reduction in transfusion of 43% in the restrictive arm compared to the liberal arm. When the difference in haemoglobin transfusion thresholds between restrictive and liberal transfusion arms was less than 2.0 g/dL, the RR was 0.80 (95% CI 0.63 to 1.02; Analysis 6.3.2), which means there was a reduction in transfusion of 20% (test for subgroup differences: Chi² = 6.11, df = 1 (P = 0.01), I² = 83.6%).

There was no clear difference in the proportions of participants transfused between the 17 studies (11,919 partipants) that used a restrictive strategy of 7.5 g/dL or less as a threshold (RR 0.55, 95% CI 0.48 to 0.64) versus the 19 studies (6035 participants) that used a restrictive threshold of 8.0 g/dL to 9.0 g/dL (RR 0.59, 95% CI 0.48 to 0.72; test for subgroup differences: Chi² = 0.18, df = 1 (P = 0.67), I² = 0%; Analysis 6.4).

The forest plot for the proportions of participants transfused displays grouping of trials with a RR around 0.5 for receiving a transfusion in the restrictive transfusion arm (Figure 5), which is consistent with the overall observation that participants in the restrictive arm were transfused approximately half as often as those in the liberal arm. As expected, there were no trials in which participants in the restrictive arm were transfused more often than those in the liberal arm.

Figure 5

Funnel plot of comparison: 2 Blood transfusions, outcome: 2.1 Participants exposed to blood transfusion (all trials)

Quantity of RBCs transfused

Seventeen trials reported the quantities of blood transfused. Most trials (40) provided some information on dose of red cells, or an algorithm or a target haemoglobin for transfusion, although the level of detail varied considerably. Among the three paediatric trials, two trials indicated a range for transfusion dose of 10 mL/kg or 10 mL/kg to 15 mL/kg (Akyildiz 2018; Robitaille 2013), and one trial reported a target haemoglobin of 8.5 g/dL to 9.5 g/dL in the restrictive arm, and 11.0 g/dL to 12.0 g/dL in the liberal arm (Lacroix 2007).

Use of a restrictive transfusion trigger resulted in an average saving of 1.21 units of RBCs per transfused participant (weighted mean difference (MD) ‐1.21, 95% CI ‐1.67 to ‐0.75; Analysis 6.5). Heterogeneity between these trials again was substantial (Chi² = 1173.58, df = 16 (P < 0.00001); I² = 91%); however, there was consistency in the direction of the effect.

Transfusion‐specific reactions

Transfusion reactions appeared to be neither well nor consistently reported, and over half (26/48) of the trials provided no information (a further two trials mentioned collecting data on such reactions but did not report them). Of the remaining 20 trials, eight reported prospectively seeking transfusion‐related reactions, but found that none had occurred in either group. Twelve trials reported transfusion reactions in a heterogeneous manner that was not suited to quantitative pooling, given the variability in methods of reporting and the assigning of severity and causality assessment to transfusion.

Haemoglobin or haematocrit concentration

Nineteen trials reported the difference in haemoglobin or haematocrit levels between liberal and restrictive transfusion arms. Measures included averages (e.g. averages of different data points across a participant's stay in ICU) as well as single data points (e.g. the last measurement before discharge). When we pooled data (without regard to timing), participants assigned to a restrictive strategy had a lower haemoglobin concentration than participants assigned to a liberal transfusion strategy (mean difference ‐1.26, 95% CI ‐1.55 to ‐0.96; analysis not shown). Heterogeneity between these trials was substantial (Chi² = 914.39, df = 18 (P < 0.00001); I² = 98%), however, there was consistency in the direction of the effect.

Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery and in those with myocardial infarction

We analysed morbidity outcomes for a subgroup of participants with underlying cardiovascular disease, defined as those undergoing cardiac surgery or vascular surgery and those with myocardial infarction. There was no difference between restrictive and liberal transfusion strategies for myocardial infarction (RR 1.01, 95% CI 0.81 to 1.26; 8 trials, 8219 participants; Analysis 7.1); renal failure (RR 1.07, 95% CI 0.89 to1.28; 7 trials, 9198 participants; Analysis 7.2); infection (RR 1.00, 0.79 to 1.28; 8 trials, 9219 participants; Analysis 7.3); thromboembolism (RR 1.02, 95% CI 0.11 to 9.55; 3 trials, 239 participants; Analysis 7.5); congestive heart failure (RR 0.77, 95% CI 0.24 to 2.43; 4 trials, 858 participants; Analysis 7.4); or cerebrovascular accident (RR 0.98, 95% CI 0.22 to 4.26; 4 trials, 905 participants; Analysis 7.6).

Economic and costing analyses

The protocol for this review did not include plans for formal economic analysis: findings from included trials are reported as a narrative summary only. Many trials included within this review discussed the potential cost implications of favouring a restrictive strategy or recommended cost‐effectiveness analysis in future research without providing data (e.g. DeZern 2016; Lotke 1999; Stanworth 2020).

Investigators in three trials went further, making estimates based on their own data when different transfusion strategies were compared for patients with severe burns (Palmieri 2017), requiring hematopoietic cell transplantation (HCT) (Tay 2020), or with cardiac issues (Koch 2017). The former two reported that a restrictive strategy would reduce costs considerably; the latter stated only that "health care costs were similar between groups".

Of the six RCTs in which formal economic analysis was specified as an outcome, two were conducted in elective surgery (cardiac or infrainguinal) (Bush 1997; Murphy 2015). The smaller, older trial (n = 99) found a substantial difference favouring the restrictive group; the latter trial included more than 2000 participants and reported, "Total costs did not differ significantly between the groups". One trial on older, critically ill patients requiring mechanical ventilation reported increased costs for this population within the group treated with a restrictive strategy, but this was a small trial, and there was a small difference in survival outcomes (Walsh 2013).

A large trial considering outcomes of women with acute anaemia after postpartum haemorrhage concluded that intervention was more expensive per woman than non‐intervention, with only a small improvement in health‐related quality of life after RBC transfusion (Prick 2014).

A trial on upper GI bleeding was a feasibility trial and did report results, confirming that transfusions were an important driver of costs alongside inpatient stay and endoscopy (Campbell 2015; Jairath 2015). A large trial on myocardial infarction concluded that significant savings were likely with the use of a restrictive strategy, although a formal publication is still pending (Ducrocq 2021). For this trial, the cost‐effectiveness endpoint was the incremental cost‐effectiveness ratio (ICER) at 30 days. It is reported that "the restrictive strategy had an 84% probability of being cost‐saving while improving clinical outcomes, i.e. "dominant" from a medico‐economic standpoint" (conference proceeding; https://www.eurekalert.org/news-releases/775135).

Discussion

Summary of main results

When compared with liberal transfusion strategies, restrictive transfusion strategies did not increase or decrease the risk of 30‐day mortality (risk ratio (RR) 0.99, 95% confidence interval (CI) 0.86 to 1.15; 31 trials, 16,729 participants; I² = 30%; moderate‐quality evidence) or of any of the other outcomes assessed (i.e. myocardial infarction, stroke (high‐quality evidence), thromboembolism (moderate‐quality evidence), and congestive heart failure (low‐quality evidence)). Restrictive transfusion strategies led to a reduction of 41% in the number of participants who received at least one unit of blood; an overall red blood cell (RBC) transfusion requirement that was approximately 1.2 units lower per participant; and a mean haemoglobin concentration that was around 1.26 g/dL lower than in liberal transfusion groups.

These findings are based on an analysis of 48 randomised controlled trials (RCTs) in this updated review that compared outcomes for participants allocated to receive transfusions of RBCs at different haemoglobin concentration thresholds. These trials enrolled 20,967 participants across diverse patient populations; most participants were adults. Since our previous 2016 review, we have included an additional 17 randomised trials.

Meta‐analyses provided no evidence that restrictive transfusion policies harmed participants, or that participants benefited from the use of liberal transfusion policies, within the parameters defined in these trials. Put another way, there was no evidence of an impact on clinically important outcomes when a restrictive RBC transfusion policy rather than a liberal RBC transfusion policy was followed. Results indicate that transfusion strategy did not influence the risk of cardiovascular events, including myocardial infarction, congestive heart failure, or stroke, although statistical heterogeneity was observed in trials that evaluated congestive heart failure (P = 0.01; I² = 57%).

In this updated review, only three trials enrolled children and the results in paediatrics were dominated by findings from a single, large pragmatic trial (Lacroix 2007), which observed no benefit for liberal transfusion in critically ill children. Another very small randomised trial recruited only six children undergoing bone marrow transplantation and was stopped because of concerns about an excess of veno‐occlusive disease in the liberal arm (Robitaille 2013); we await the results of a further ongoing trial in this setting (ISRCTN17438123; see Ongoing studies).

Subgroup analyses

With regard to our predefined clinical subgroups, results indicate that risk of death and other adverse events were not impacted by liberal or restrictive transfusion thresholds for most analyses. This is important because there are pathophysiological reasons to postulate why transfusion might impact clinical outcomes differently in different patient populations as the result of factors such as duration of anaemia (short‐term transfusion dependence in critical illness versus long‐term transfusion dependence in bone marrow failure) or presence or absence of an underlying restriction in cardiac function (Docherty 2016).

However, for patients with acute blood loss and for those with acute myocardial infarction, mortality may be influenced by a liberal or restrictive transfusion strategy, although the test for differences in 30‐day mortality between subgroups showed no differences (P = 0.13; I² = 41.2%). In three trials (1522 participants) including people with gastrointestinal bleeding (included in the acute blood loss or trauma grouping), a restrictive transfusion strategy was associated with a 35% lower risk of 30‐day mortality compared with a liberal transfusion strategy. The mechanism responsible for this significantly reduced risk of death may be lower risk of rebleeding under restrictive transfusion regimens (RR 0.65, 95% CI 0.43 to 0.97; Analysis 1.3.4). The reason for this effect is not known, but it may reflect higher vascular pressures following transfusion in the liberal transfusion group compared with the restrictive transfusion group.

Patients with acute myocardial infarction are another important patient subpopulation. In the 2016 update (Carson 2016b), two small trials included people with myocardial infarction (154 participants) for whom 30‐day mortality was 3.88 times higher in the restrictive transfusion group than in the liberal transfusion group (95% CI 0.83 to 18.13). In one trial, 12.7% of participants (n = 14) undergoing cardiac catheterisation had stable coronary artery disease but did not have acute myocardial infarction (Carson 2013). These results have been extended by inclusion in the meta‐analysis of Ducrocq 2021, which enrolled 666 patients and reported results for a composite major adverse cardiac events (MACE) outcome of all‐cause mortality, stroke, recurrent myocardial infarction, or emergency revascularisation in 11.1% of the restrictive group and in 14.2% of the liberal group. Deaths from any cause occurred in 5.6% of the restrictive group and in 7.7% of the liberal group. With the addition of this trial, the relative risk for 30‐day mortality was closer to 1.0, with wide confidence intervals (RR 1.61, 95% CI 0.38 to 6.88). A 3500‐patient trial in acute myocardial infarction, called 'MINT', is currently under way and will inform this subgroup further (NCT02981407).

The nature of the restrictive transfusion intervention

Around half of the trials identified applied a restrictive threshold of 7.0 g/dL; the other half used 8.0 g/dL. The largest trial including cardiac surgery patients used a 7.5‐g/dL threshold (Mazer 2017). Most participants in the 7.0 g/dL restrictive transfusion threshold trials were based in critical care and acute settings of anaemia. Clinical specialties were more varied in trials that tested an 8.0 g/dL restrictive transfusion threshold and included orthopaedic and cardiac surgery, gastrointestinal bleeding, and acute myocardial infarction. However, there was no apparent difference in risk of death at 30 days between the two strata.

We compared 30‐day mortality in trials where the difference between liberal and restrictive transfusion thresholds in the trial protocol was at least 2.0 g/dL versus trials where the difference was less than 2.0 g/dL. Again, there was no evidence of a dose effect on clinical outcomes of RBC transfusion by different threshold levels of haemoglobin concentration.

Risks of infection and outcomes of recovery

In view of potential immunomodulatory effects of blood transfusion, we compared the risk of infection in three ways. We did not find evidence of an increased risk of infection associated with liberal transfusion. We combined all infections and also examined sepsis or bacteraemia and pneumonia (alone); comparative risks of infection between the two transfusion strategies were nearly identical for all of these analyses. These results varied from prior analyses in another systematic review, which reported an elevated risk of infection in the liberal transfusion group (Rohde 2014). However, they are consistent with later analyses (Carson 2016b; Carson 2018), which were based on a substantially larger number of trials.

Although 23 trials assessed functional recovery or quality of life and fatigue, these trials applied different measures or tools for assessment; therefore quantitative meta‐analysis could not be supported for these outcomes.

Overall completeness and applicability of evidence

As the number of trials expands, the completeness of evidence continues to increase. Clinical trials have now evaluated many of the most common clinical specialties in which RBCs are transfused. Thus, the findings of this review are widely applicable to most clinical contexts. However, we continue to lack knowledge and sufficient precision about the safety of different transfusion thresholds for some groups of patients who frequently receive transfusion, as an insufficient number of trials with adequate power have been published for these groups. These understudied clinical contexts include people with myocardial infarction, neurological injury/traumatic brain injury, acute neurological disorder, stroke, cancer/haematological malignancy, and chronic bone marrow failure. We anticipate some of these gaps may be filled relatively soon as new trials are completed, for which details and recruitment targets are listed in Characteristics of ongoing studies.

A core rationale for RBC transfusion is to improve tissue and cellular oxygenation, but technologies for monitoring this directly, or at a cellular level, are not available routinely. Therefore, haemoglobin concentration continues to be applied as the main surrogate marker of need for transfusion in our included trials, but it may not be a reliable biomarker and it may not support a more precise or personalised approach to transfusion therapy (Baek 2019). Trials that evaluate mechanistic and physiological variables alongside haemoglobin concentration are required (Møller 2019).

Quality of the evidence

Overall, the quality of evidence across trials is good and continues to show improvement over time. The number of trials and enrolled participants has increased substantially, and the precision of the estimate of the effect of transfusion improves with updates of this review. We found relatively little heterogeneity for each clinical outcome across all analyses.

Risk of bias evaluations revealed a variety of methodological issues between trials. For more recent trials, including those reporting on larger sample sizes, evaluations of risk of bias remained generally low risk. We applied Cochrane methods for defining high or low risk of bias to all trials, but we acknowledge a number of challenges, including how to assign a single level of bias for multiple outcomes, for example, incomplete data or blinding (masking). We therefore considered risk of bias for objective (mortality) and subjective (functional and quality of life) outcomes separately. We will explore this further in future updates of this review, possibly by employing the new Cochrane 'Risk of bias 2' tool. We recognise that blinding the use of transfusion at the bedside is difficult to achieve unless trial personnel are assigned to each participant, which would be an expensive procedure. The importance of blinding will differ according to the choice of primary trial outcome; mortality is a hard endpoint (as in this review) that is less open to bias than other functional outcomes.

Outcome assessment by observers who are blind to the treatment group is probably the most rigorous practical approach for transfusion threshold trials, but this is less relevant for outcomes such as mortality. We judged the risk of bias to be high for 11 trials for subjective outcomes, including functional outcomes and quality of life. This issue of detection bias for subjective outcomes will be explored in greater detail in a further update of this review, and informed by additional trial data (see Characteristics of ongoing studies; Characteristics of studies awaiting classification). Maintaining the integrity of the randomisation process becomes important if the trial is not to overestimate the benefit of the intervention (Schulz 1995). We judged the risk of bias for allocation concealment ‐ a key methodological domain ‐ to be low for 36 trials. Only a few trials in this review did not report the methods used to conceal the allocation sequence from treating clinicians.

We recognise a number of further limitations to the quality of trial evidence, beyond those considered by the Cochrane risk of bias assessment. As described by authors of other reviews, these include variable degrees of difference in numbers of transfusions between arms, and variable degrees of actual separation of haemoglobin concentration achieved between trial arms, which often is less than defined thresholds for trial interventions stated in the protocol (Trentino 2020a; Trentino 2020b). Reasons for protocol violations, whether given as extra transfusions in the restrictive arm, or for lack of transfusions in the liberal arm, are often not reported in sufficient detail.

We considered the policy of the Cochrane Injuries Group for research integrity (Roberts 2015). We have explored possible reasons for subgroup differences between trials that were prospectively registered and those not prospectively registered (Chi² = 4.06, df = 1 (P = 0.04), I² = 75.4%). First, we created a funnel plot and found little evidence to support selection bias (Figure 6). Second, it is possible that this statistically significant finding is a chance observation given the large number of statistical tests performed in this review. Third, we note that of the 13 trials that were not prospectively registered, seven were conducted before 2000 or 2001 when legislation regarding prospective registration was introduced in the USA and European Union, respectively, and nine of the trials were conducted before September 2007 when it was mandated by the Food and Drug Administration (FDA).

Figure 6

Funnel plot of comparison: 2 Subgroup analysis by prospective registration, outcome: 2.1 30‐Day mortality

Whilst we cannot exclude the presence of selection bias, we are confident we did not miss any large trials published before 2007 that would have impacted our inferences. Indeed, the overall RR for an analysis including all trials was 0.99 (95% CI 0.86 to 1.15), while an analysis limited to the trials with prospective registration was RR 1.08 (95% CI 0.89 to 1.31).

We observed a substantial amount of statistical heterogeneity in analyses evaluating the proportions of participants transfused, the quantity of RBCs transfused, and differences in haemoglobin/haematocrit concentrations. It is conventional practice not to pool data from studies in which there is a large amount of heterogeneity, however, we chose to present pooled results for these transfusion outcomes for several reasons. First, the impact of restrictive transfusion on the proportion of participants transfused varied only by the magnitude of the reduction in transfusion ‐ not the direction. In all trials, participants in the restrictive transfusion group received fewer transfusions, although the number varied because transfusion protocols were different and clinical contexts required different frequencies of transfusion. Second, we expected this heterogeneity because of the variety of contexts for clinical trials, including participant age, degrees of comorbidity, and policies for standard transfusion practice, which, in turn, reflect specialty‐specific guidelines and recommendations. At one extreme, nearly all participants, if not all, with leukaemia and cancer were transfused (DeZern 2016; Stanworth 2020). Transfusion risk in participants in critical care (Hébert 1995; Hébert 1999; Lacroix 2007), or with acute blood loss (Villanueva 2013), was about 50% at the time of the trials.

In summary, we have chosen to present pooled results for outcomes of transfusion because we are evaluating the effects of restrictive transfusion practice, and because all trial estimates for changes in transfusion are consistently in the same direction. The substantial heterogeneity, therefore, reflects diversity in the strength of estimates, rather than efficacy of the policy. Reasons for diversity in the strength of trial estimates include known and expected clinical contexts and different practice guidelines used by different specialties. Subgroup explorations for transfusion outcomes reported earlier demonstrated these differences.

Potential biases in the review process

We performed extensive searches in an attempt to identify all eligible trials irrespective of publication status or language. Inspection of funnel plots did not reveal a major risk of publication bias (Figure 4; Figure 5).

Other trial limitations apply to the findings of this review. Timing of mortality reporting varied between trials, in part as a consequence of the clinical context. To address this issue, an initiative to undertake an individual patient data analysis has been commenced, including contact with all trial investigators to explore willingness to share trial data.

Randomised trials in this review may not have evaluated important clinical outcomes adequately that are specifically relevant to the use of RBC transfusions, such as quality of life. The identified trials evaluated the effects of transfusion only in hospitalised patients, and only two small trials have tested different thresholds in an outpatient population (Jansen 2020; Stanworth 2020), for whom function and fatigue would be more important endpoints.

Different grades of severity of cardiovascular events, such as myocardial infarction, congestive heart failure, or stroke, and different risks of overall infection will occur in patients; these events may present in ways that are not always clinically overt and so are more subjective in interpretation. This is important because RBC transfusions may have both harmful and beneficial effects on the risk of these outcomes, for example, balancing prothrombotic tendencies against protective mechanisms to limit restrictions in myocardial oxygen delivery. Future trials need to establish robust definitions of all outcomes (Docherty 2016). Despite the large number of participants included in these trials, there remains inadequate power for many outcomes.

Agreements and disagreements with other studies or reviews

The results of this review are consistent with previous published systematic reviews and guidance documents (Carson 2012a; Docherty 2016; Holst 2015; Meybohm 2020). A review of reviews reported no evidence that a difference in mortality exists between patients assigned to a restrictive or a liberal transfusion strategy (Trentino 2020a). These overall findings provide no evidence that restrictive transfusion policies harm patients within the limits defined by the trials.

Multiple reviews have addressed outcomes in selected subgroups or subpopulations of patients. A review by Fominskiy aimed to assess effects of liberal and restrictive RBC transfusion strategies on mortality in perioperative and critically ill adult patients through a meta‐analysis of relevant trials (Fominskiy 2015); a more recent review focused on cardiac surgery (Shehata 2019). A meta‐analysis of trials in gastrointestinal bleeding reported evidence of harm with application of liberal thresholds in this patient group (Odutayo 2017), as found in our review, although our review has identified one further trial in gastrointestinal bleeding (Kola 2020). We suggest some caution in interpretation of systematic reviews that report only separate subpopulations of the wider trial literature. Anaemia is the common presenting clinical problem for all patients when a red cell transfusion is considered, irrespective of clinical setting. There is a risk that the patterns of findings in different clinical contexts are inappropriately selective to a small subpopulation. The clinical decision process for transfusion in one clinical context may need to draw on findings of safety as reported across all randomised trials in different clinical settings, for a common intervention of RBC.

The results of our meta‐analyses need to be viewed against studies or reviews of large observational studies that have reported comparisons of clinical outcomes at varying haemoglobin levels in transfused and non‐transfused patients. Publications of the observational literature have reported findings at variance with the randomised trial literature (Carson 1998; Hébert 1997; Hébert 1999; Patel 2015; Spiess 1998; Wu 2001; Wu 2007; Wu 2010). Reviews of observational data have reported an increase in risk of death associated with transfusion (Chatterjee 2013; Marik 2008). However, a limitation of observational studies is that there may be residual confounding by indication, despite extensive statistical adjustment of the results. It is possible that differences in patient characteristics between those who were transfused and those who were not transfused may not have been identified or adjusted for adequately. In contrast, results of the meta‐analysis of clinical trials performed in this review update show no increase in risk of death for liberal transfusion thresholds compared with restrictive transfusion thresholds. Despite assertions to the contrary (Benson 2000; Concato 2000), we continue to believe there is a need for adequately powered, rigorously performed, randomised trials to provide the highest level of evidence when effects of different transfusion policies are tested, as the way of overcoming these limitations.

The transfusion policies reviewed here represent fairly small but significant modifications to routine clinical practice. They are consistent with the recommendations of published clinical practice guidelines (AAGBI 2008; ASA 2006; BCTMAG 2003; Carson 2012a; Carson 2016b; Mueller 2019; Napolitano 2009; NBUGI 2001; Retter 2013; STSBCGTF 2011). Transfusion triggers (in terms of haemoglobin levels) were most often in the range of 7.0 g/dL to 10.0 g/dL. In fact, the 'restrictive' transfusion triggers in some trials were equivalent to the 'liberal' triggers used in other trials. Nevertheless, trials documented significant reduction in exposure of patients to unnecessary RBC transfusion. Our findings for red cell transfusion strategies should be interpreted alongside findings for the use of 'alternative' agents to red cells or blood‐sparing techniques, such as intravenous iron (Richards 2020), cell salvage/blood conservation (Carless 2010a; STSBCGTF 2011), and antifibrinolytic drugs (Henry 2011). Adoption of a conservative transfusion threshold appears to be as effective, if not more effective, in the context of Patient Blood Managment implementation, and is likely to cost less (Roman 2020).

Some guidelines have recommended RBC transfusion for symptoms or haemodynamic instability, rather than for a specific trigger haemoglobin level (AAGBI 2008; ASA 2006; Napolitano 2009; NBUGI 2001). Three studies tested this approach to transfusion: a pilot study involving 84 participants (Carson 1998), a trial involving 2016 participants (Carson 2011), and a 110‐participant trial for acute myocardial infarction (Carson 2013), in which patients could be transfused if they exhibited symptoms or had a haemoglobin concentration less than 8.0 g/dL. These studies found no differences in functional recovery, mortality, or morbidity among patients in the restrictive (symptomatic) transfusion group in orthopaedic surgery trials (Carson 1998; Carson 2011), although in the trial involving patients with acute myocardial infarction (Carson 2013), there was a tendency towards worse outcomes in the restrictive transfusion group.

Figure 1

Flow of studies for 2021 update

Navigate to figure in ReviewOpen in new tab

Figure 2

'Risk of bias' graph: review authors' judgements about each 'Risk of bias' item presented as percentages across all included trials. Forty‐eight trials are included in this review.

Navigate to figure in ReviewOpen in new tab

Figure 3

'Risk of bias' summary: review authors' judgements about each methodological quality item for each included trial

Navigate to figure in ReviewOpen in new tab

Figure 4

Funnel plot of comparison: 1 Mortality, outcome: 1.1 30‐Day mortality

Navigate to figure in ReviewOpen in new tab

Figure 5

Funnel plot of comparison: 2 Blood transfusions, outcome: 2.1 Participants exposed to blood transfusion (all trials)

Navigate to figure in ReviewOpen in new tab

Figure 6

Funnel plot of comparison: 2 Subgroup analysis by prospective registration, outcome: 2.1 30‐Day mortality

Navigate to figure in ReviewOpen in new tab

Analysis 1.1

Comparison 1: Mortality at 30 days, Outcome 1: 30‐Day mortality

Navigate to figure in ReviewOpen in new tab

Analysis 1.2

Comparison 1: Mortality at 30 days, Outcome 2: 30‐Day mortality subgroup by restrictive haemoglobin level

Navigate to figure in ReviewOpen in new tab

Analysis 1.3

Comparison 1: Mortality at 30 days, Outcome 3: 30‐Day mortality subgroup analysis by clinical specialties

Navigate to figure in ReviewOpen in new tab

Analysis 1.4

Comparison 1: Mortality at 30 days, Outcome 4: 30‐Day mortality by clinical specialties: myocardial infarction vs all others

Navigate to figure in ReviewOpen in new tab

Analysis 1.5

Comparison 1: Mortality at 30 days, Outcome 5: Mortality by cardiac surgery, vascular surgery, myocardial infarction, and all others

Navigate to figure in ReviewOpen in new tab

Analysis 2.1

Comparison 2: Subgroup analysis by prospective registration, Outcome 1: 30‐Day mortality

Navigate to figure in ReviewOpen in new tab

Analysis 3.1

Comparison 3: Sensitivity analysis by allocation concealment, Outcome 1: 30‐Day mortality

Navigate to figure in ReviewOpen in new tab

Analysis 4.1

Comparison 4: Mortality: other time intervals, Outcome 1: Hospital mortality

Navigate to figure in ReviewOpen in new tab

Analysis 4.2

Comparison 4: Mortality: other time intervals, Outcome 2: 90‐Day mortality

Navigate to figure in ReviewOpen in new tab

Analysis 4.3

Comparison 4: Mortality: other time intervals, Outcome 3: 6‐Month mortality

Navigate to figure in ReviewOpen in new tab

Analysis 5.1

Comparison 5: Morbidity: clinical outcomes, Outcome 1: Cardiac events

Navigate to figure in ReviewOpen in new tab

Analysis 5.2

Comparison 5: Morbidity: clinical outcomes, Outcome 2: Myocardial infarction

Navigate to figure in ReviewOpen in new tab

Analysis 5.3

Comparison 5: Morbidity: clinical outcomes, Outcome 3: Congestive heart failure

Navigate to figure in ReviewOpen in new tab

Analysis 5.4

Comparison 5: Morbidity: clinical outcomes, Outcome 4: Cerebrovascular accident (CVA) ‐ stroke

Navigate to figure in ReviewOpen in new tab

Analysis 5.5

Comparison 5: Morbidity: clinical outcomes, Outcome 5: Rebleeding

Navigate to figure in ReviewOpen in new tab

Analysis 5.6

Comparison 5: Morbidity: clinical outcomes, Outcome 6: Sepsis/bacteraemia

Navigate to figure in ReviewOpen in new tab

Analysis 5.7

Comparison 5: Morbidity: clinical outcomes, Outcome 7: Pneumonia

Navigate to figure in ReviewOpen in new tab

Analysis 5.8

Comparison 5: Morbidity: clinical outcomes, Outcome 8: Infection

Navigate to figure in ReviewOpen in new tab

Analysis 5.9

Comparison 5: Morbidity: clinical outcomes, Outcome 9: Thromboembolism

Navigate to figure in ReviewOpen in new tab

Analysis 5.10

Comparison 5: Morbidity: clinical outcomes, Outcome 10: Renal failure

Navigate to figure in ReviewOpen in new tab

Analysis 5.11

Comparison 5: Morbidity: clinical outcomes, Outcome 11: Mental confusion

Navigate to figure in ReviewOpen in new tab

Analysis 6.1

Comparison 6: Blood transfusions, Outcome 1: Participants exposed to blood transfusion (all trials)

Navigate to figure in ReviewOpen in new tab

Analysis 6.2

Comparison 6: Blood transfusions, Outcome 2: Participants exposed to blood transfusion by clinical specialties

Navigate to figure in ReviewOpen in new tab

Analysis 6.3

Comparison 6: Blood transfusions, Outcome 3: Participants exposed to blood transfusion (by transfusion threshold)

Navigate to figure in ReviewOpen in new tab

Analysis 6.4

Comparison 6: Blood transfusions, Outcome 4: Participants exposed to blood transfusion by transfusion threshold

Navigate to figure in ReviewOpen in new tab

Analysis 6.5

Comparison 6: Blood transfusions, Outcome 5: Units of blood transfused

Navigate to figure in ReviewOpen in new tab

Analysis 7.1

Comparison 7: Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI, Outcome 1: Myocardial infarction

Navigate to figure in ReviewOpen in new tab

Analysis 7.2

Comparison 7: Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI, Outcome 2: Renal failure

Navigate to figure in ReviewOpen in new tab

Analysis 7.3

Comparison 7: Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI, Outcome 3: Infection

Navigate to figure in ReviewOpen in new tab

Analysis 7.4

Comparison 7: Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI, Outcome 4: Congestive heart failure

Navigate to figure in ReviewOpen in new tab

Analysis 7.5

Comparison 7: Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI, Outcome 5: Thromboembolism

Navigate to figure in ReviewOpen in new tab

Analysis 7.6

Comparison 7: Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI, Outcome 6: Cerebrovascular accident

Navigate to figure in ReviewOpen in new tab

Summary of findings 1. Liberal compared with restrictive transfusion protocols for guiding red blood cell transfusion

Outcomes	*Anticipated absolute effects (95% CI)**		Relative effect (95% CI)	№. of participants (studies)	Certainty of the evidence (GRADE)	Comments
Liberal compared with restrictive transfusion protocols for guiding red blood cell transfusion
Patient or population: adults and children (haemodynamically stable) with potential need for RBC transfusion Setting: inpatients Intervention: restrictive transfusion threshold Comparison: liberal transfusion threshold
Outcomes	Risk with liberal transfusion protocol	Risk with restrictive transfusion protocol	Relative effect (95% CI)	№. of participants (studies)	Certainty of the evidence (GRADE)	Comments
Participants exposed to blood transfusion (all studies)	Study population		RR 0.59 (0.53 to 0.66)	20,057 (42)	⊕⊕⊕⊕ High
Participants exposed to blood transfusion (all studies)	815 per 1000	481 per 1000 (432 to 538)	RR 0.59 (0.53 to 0.66)	20,057 (42)	⊕⊕⊕⊕ High
30‐Day mortality	Study population		RR 0.99 (0.86 to 1.15)	16,729 (31)	⊕⊕⊕⊕ High
30‐Day mortality	83 per 1000	83 per 1000 (71 to 96)	RR 0.99 (0.86 to 1.15)	16,729 (31)	⊕⊕⊕⊕ High
Myocardial infarction	Study population		RR 1.04 (0.87 to 1.24)	14,370 (23)	⊕⊕⊕⊕ High
Myocardial infarction	32 per 1000	33 per 1000 (28 to 40)	RR 1.04 (0.87 to 1.24)	14,370 (23)	⊕⊕⊕⊕ High
Congestive heart failure	Study population		RR 0.83 (0.53 to 1.29)	7247 (16)	⊕⊕⊝⊝ Low^a
Congestive heart failure	35 per 1000	29 per 1000 (19 to 45)	RR 0.83 (0.53 to 1.29)	7247 (16)	⊕⊕⊝⊝ Low^a
Cerebrovascular accident ‐ stroke	Study population		RR 0.84 (0.64 to 1.09)	13,985 (19)	⊕⊕⊕⊕ High
Cerebrovascular accident ‐ stroke	17 per 1000	14 per 1000 (11 to 19)	RR 0.84 (0.64 to 1.09)	13,985 (19)	⊕⊕⊕⊕ High
Rebleeding	Study population		RR 0.80 (0.59 to 1.09)	3412 (8)	⊕⊕⊕⊝ Moderate^b
Rebleeding	158 per 1000	126 per 1000 (93 to 172)	RR 0.80 (0.59 to 1.09)	3412 (8)	⊕⊕⊕⊝ Moderate^b
Thromboembolism	Study population		OR 1.11 (0.65 to 1.88)	4201 (13)	⊕⊕⊕⊝ Moderate^c
Thromboembolism	15 per 1000	17 per 1000 (10 to 28)	OR 1.11 (0.65 to 1.88)	4201 (13)	⊕⊕⊕⊝ Moderate^c
Infection	Study population		RR 0.97 (0.88 to 1.07)	17,104 (25)	⊕⊕⊕⊕ High
Infection	143 per 1000	139 per 1000 (126 to 153)	RR 0.97 (0.88 to 1.07)	17,104 (25)	⊕⊕⊕⊕ High
*The risk in the intervention group (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; OR: odds ratio; RBC: red blood cell; RR: risk ratio
GRADE Working Group grades of evidence. High certainty: we are very confident that the true effect lies close to that of the estimate of the effect. Moderate certainty: we are moderately confident in the effect estimate: the true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different. Low certainty: our confidence in the effect estimate is limited: the true effect may be substantially different from the estimate of the effect. Very low certainty: we have very little confidence in the effect estimate: the true effect is likely to be substantially different from the estimate of effect.
^aWe downgraded once for inconsistency, as there was no consistency in the direction of the effect (despite the relatively low statistical heterogeneity), and we downgraded once for imprecision, as there were very low numbers of events. ^bDespite relatively low statistical heterogeneity, there was no consistency in the direction of the effect, hence we downgraded once for inconsistency. ^cDowngraded once for imprecision, as there were few events (and hence a wide CI).

Summary of findings 1. Liberal compared with restrictive transfusion protocols for guiding red blood cell transfusion

Navigate to table in Review

Table 1. Trial setting details

Study ID	Number of participants at baseline	Country/Countries	Number of sites	Setting(s)	Year recruitment started
Mazer 2017	5092	19 countries^a	73	73 sites ‐ varied	2014
Carson 2011	2016	USA, Canada	47	47 sites ‐ varied	2004
Murphy 2015	2003	UK	17	17 sites ‐ varied	2009
Holst 2014	1005	Denmark, Sweden, Norway, Finland	32	32 general ICUs	2011
Jairath 2015	936	UK	6	University teaching hospitals	2012
Villanueva 2013	921	Spain	1	General hospital	2003
Hébert 1999	838	Canada	25	Tertiary (22), community ICU (3)	1994
Koch 2017	722	USA (1), India (1)	2	1 academic affiliated hospital in the USA, a private hospital in India	2007
Ducrocq 2021	668	France, Spain	35	35 sites ‐ varied	2016
Lacroix 2007	648	Canada, Belgium, USA, UK	19	Tertiary paediatric ICU	2001
So‐Osman 2013	603	Netherlands	3	Varied ‐ university and general hospitals	2001
Prick 2014	519	Netherlands	37	Varied ‐ university and general hospitals	2004
Hajjar 2010	512	Brazil	1	University teaching hospital	2009
Hoff 2011	466	Denmark	??	Oncology centres	1986
Bracey 1999	428	USA	1	University teaching hospital	1997
Palmieri 2017	345	US (16 sites), Canada (1), New Zealand (1)	18	Specialist burn centres	2010
Tay 2020	300	Canada	4	HCT sites	2011
Bergamin 2017	300	Brazil	1	University teaching hospital	2012
Gregersen 2015	284	Denmark	1	University teaching hospital	2010
Grover 2006	260	UK	3	Acute hospitals	Not stated
Kola 2020	224	India	1	Tertiary hospital	2015
Parker 2013	200	UK	1	General hospital	2002
de Almeida 2015	198	Brazil	1	Tertiary oncology university hospital	2012
Fan 2014	192	China	1	University teaching hospital	2011
Akyildiz 2018	180	Turkey	1	University teaching hospital	2014
Yakymenko 2018	133	Denmark	1	University teaching hospital	2010
Lotke 1999	127	USA	1	University teaching hospital	Not stated
Foss 2009	120	Denmark	1	University teaching hospital	2004
Carson 2013	110	USA	8	8 sites ‐ varied	2010
Walsh 2013	100	UK	6	Varied ‐ university and general hospitals	2009
Bush 1997	99	USA	1	University teaching hospital	1995
DeZern 2016	89	USA	1	Tertiary referral centre for oncology	2014
Carson 1998	84	USA (3), UK (1)	4	University teaching hospitals	1996
Laine 2018	80	Finland	1	University teaching hospital	2014
Hébert 1995	69	Canada	5	Tertiary hospitals	1993
Nielsen 2014	66	Denmark	2	University teaching hospital and general hospital	2009
Gillies 2020	62	UK	1	University teaching hospital	2017
Webert 2008	60	Canada	4	Tertiary oncology centres	2003
Møller 2019	58	Denmark	1	General hospital	2015
Shehata 2012	50	Canada	1	University teaching hospital	2007
Blair 1986	50	UK	1	University teaching hospital	Not stated
Gobatto 2019	47	Brazil	1	University teaching hospital	2014
Cooper 2011	45	USA	2	Veterans' Affairs hospital centres	2003
Johnson 1992	39	USA	1	University teaching hospital	Not stated
Stanworth 2020	38	UK, Australia, New Zealand	12	12 sites ‐ varied	2015
Topley 1956	22	UK	1	'Accident hospital'	Not stated
Januarysen 2020	19	Netherlands	3	1 university hospital, 2 general hospitals	2002
Robitaille 2013	6	Canada	1	Not identified	2009
^aMazer 2017 (TRICS‐III): majority of sites in USA; sites also in Australia, Brazil, Canada, China, Colombia, Denmark, Egypt, Germany, Greece, India, Israel, Malaysia, New Zealand, Romania, Singapore, South Africa, Spain, and Switzerland.

Table 1. Trial setting details

Navigate to table in Review

Comparison 1. Mortality at 30 days

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1.1 30‐Day mortality Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.15]

1.2 30‐Day mortality subgroup by restrictive haemoglobin level Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.14]

1.2.1 Restrictive 7.0 g/dL to 7.5 g/dL	15	11572	Risk Ratio (M‐H, Random, 95% CI)	1.00 [0.83, 1.19]
1.2.2 Restrictive < 8.0 g/dL to 9.0 g/dL	16	5157	Risk Ratio (M‐H, Random, 95% CI)	0.97 [0.75, 1.24]
1.3 30‐Day mortality subgroup analysis by clinical specialties Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.14]

1.3.1 Cardiac surgery	4	7441	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.74, 1.33]
1.3.2 Orthopaedic surgery	8	3111	Risk Ratio (M‐H, Random, 95% CI)	1.16 [0.75, 1.79]
1.3.3 Vascular	2	157	Risk Ratio (M‐H, Random, 95% CI)	0.98 [0.30, 3.25]
1.3.4 Acute blood loss/trauma	3	1522	Risk Ratio (M‐H, Random, 95% CI)	0.65 [0.43, 0.97]
1.3.5 Critical care	9	3529	Risk Ratio (M‐H, Random, 95% CI)	1.06 [0.85, 1.32]
1.3.6 Acute myocardial infarction	3	820	Risk Ratio (M‐H, Random, 95% CI)	1.61 [0.38, 6.88]
1.3.7 Haematological malignancies	2	149	Risk Ratio (M‐H, Random, 95% CI)	0.37 [0.07, 1.95]
1.4 30‐Day mortality by clinical specialties: myocardial infarction vs all others Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.14]

1.4.1 Myocardial infarction	3	820	Risk Ratio (M‐H, Random, 95% CI)	1.61 [0.38, 6.88]
1.4.2 All but myocardial infarction	28	15909	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.15]
1.5 Mortality by cardiac surgery, vascular surgery, myocardial infarction, and all others Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.14]

1.5.1 Cardiac surgery	4	7441	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.74, 1.33]
1.5.2 Myocardial infarction	3	820	Risk Ratio (M‐H, Random, 95% CI)	1.61 [0.38, 6.88]
1.5.3 Vascular surgery	2	157	Risk Ratio (M‐H, Random, 95% CI)	0.98 [0.30, 3.25]
1.5.4 Others	22	8311	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.83, 1.19]

Comparison 1. Mortality at 30 days

Navigate to table in Review

Comparison 2. Subgroup analysis by prospective registration

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
2.1 30‐Day mortality Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.15]

2.1.1 Prospectively registered trials	18	12932	Risk Ratio (M‐H, Random, 95% CI)	1.08 [0.89, 1.31]
2.1.2 Trials without prospective registration	13	3797	Risk Ratio (M‐H, Random, 95% CI)	0.81 [0.66, 1.00]

Comparison 2. Subgroup analysis by prospective registration

Navigate to table in Review

Comparison 3. Sensitivity analysis by allocation concealment

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
3.1 30‐Day mortality Show forest plot	31	16729	Risk Ratio (M‐H, Random, 95% CI)	0.99 [0.86, 1.15]

3.1.1 Low risk of bias	26	15764	Risk Ratio (M‐H, Random, 95% CI)	1.01 [0.87, 1.18]
3.1.2 Unclear or high risk of bias	5	965	Risk Ratio (M‐H, Random, 95% CI)	0.84 [0.51, 1.39]

Comparison 3. Sensitivity analysis by allocation concealment

Navigate to table in Review

Comparison 4. Mortality: other time intervals

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
4.1 Hospital mortality Show forest plot	15	6597	Risk Ratio (M‐H, Random, 95% CI)	0.86 [0.72, 1.03]

4.2 90‐Day mortality Show forest plot	7	4143	Risk Ratio (M‐H, Random, 95% CI)	1.13 [1.02, 1.25]

4.3 6‐Month mortality Show forest plot	2	4702	Risk Ratio (M‐H, Random, 95% CI)	0.98 [0.79, 1.22]

Comparison 4. Mortality: other time intervals

Navigate to table in Review

Comparison 5. Morbidity: clinical outcomes

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
5.1 Cardiac events Show forest plot	11	5577	Risk Ratio (M‐H, Random, 95% CI)	1.03 [0.80, 1.32]

5.2 Myocardial infarction Show forest plot	23	14370	Risk Ratio (M‐H, Random, 95% CI)	1.04 [0.87, 1.24]

5.3 Congestive heart failure Show forest plot	16	7247	Risk Ratio (M‐H, Random, 95% CI)	0.83 [0.53, 1.29]

5.4 Cerebrovascular accident (CVA) ‐ stroke Show forest plot	19	13985	Risk Ratio (M‐H, Random, 95% CI)	0.84 [0.64, 1.09]

5.5 Rebleeding Show forest plot	8	3412	Risk Ratio (M‐H, Random, 95% CI)	0.80 [0.59, 1.09]

5.6 Sepsis/bacteraemia Show forest plot	9	4352	Risk Ratio (M‐H, Random, 95% CI)	1.06 [0.86, 1.30]

5.7 Pneumonia Show forest plot	16	6666	Risk Ratio (M‐H, Random, 95% CI)	0.97 [0.84, 1.13]

5.8 Infection Show forest plot	25	17104	Risk Ratio (M‐H, Random, 95% CI)	0.97 [0.88, 1.07]

5.9 Thromboembolism Show forest plot	13	4201	Peto Odds Ratio (Peto, Fixed, 95% CI)	1.11 [0.65, 1.88]

5.10 Renal failure Show forest plot	15	12531	Risk Ratio (M‐H, Random, 95% CI)	1.03 [0.92, 1.16]

5.11 Mental confusion Show forest plot	9	6442	Risk Ratio (M‐H, Random, 95% CI)	1.11 [0.88, 1.40]

Comparison 5. Morbidity: clinical outcomes

Navigate to table in Review

Comparison 6. Blood transfusions

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
6.1 Participants exposed to blood transfusion (all trials) Show forest plot	42	20057	Risk Ratio (M‐H, Random, 95% CI)	0.59 [0.53, 0.66]

6.2 Participants exposed to blood transfusion by clinical specialties Show forest plot	41	19977	Risk Ratio (M‐H, Random, 95% CI)	0.59 [0.53, 0.66]

6.2.1 Cardiac surgery	7	8598	Risk Ratio (M‐H, Random, 95% CI)	0.69 [0.66, 0.73]
6.2.2 Orthopaedic surgery	11	3969	Risk Ratio (M‐H, Random, 95% CI)	0.49 [0.38, 0.65]
6.2.3 Vascular surgery	2	157	Risk Ratio (M‐H, Random, 95% CI)	0.79 [0.57, 1.08]
6.2.4 Acute blood loss/trauma	5	2416	Risk Ratio (M‐H, Random, 95% CI)	0.39 [0.23, 0.67]
6.2.5 Critical care	9	3529	Risk Ratio (M‐H, Random, 95% CI)	0.66 [0.57, 0.77]
6.2.6 Acute myocardial infarction	3	821	Risk Ratio (M‐H, Random, 95% CI)	0.38 [0.28, 0.53]
6.2.7 Haematological malignancies	4	487	Risk Ratio (M‐H, Random, 95% CI)	0.88 [0.61, 1.26]
6.3 Participants exposed to blood transfusion (by transfusion threshold) Show forest plot	33		Risk Ratio (M‐H, Random, 95% CI)	Subtotals only

6.3.1 Difference between liberal and restrictive haemoglobin thresholds ≥ 2.0 g/dL	27	15072	Risk Ratio (M‐H, Random, 95% CI)	0.57 [0.50, 0.64]
6.3.2 Difference between liberal and restrictive haemoglobin thresholds < 2.0 g/dL	6	2966	Risk Ratio (M‐H, Random, 95% CI)	0.80 [0.63, 1.02]
6.4 Participants exposed to blood transfusion by transfusion threshold Show forest plot	36	17954	Risk Ratio (M‐H, Random, 95% CI)	0.57 [0.51, 0.64]

6.4.1 Restrictive 7.0 g/dL to 7.5 g/dL	17	11919	Risk Ratio (M‐H, Random, 95% CI)	0.55 [0.48, 0.64]
6.4.2 Restrictive < 8.0 g/dL to 9.0 g/dL	19	6035	Risk Ratio (M‐H, Random, 95% CI)	0.59 [0.48, 0.72]
6.5 Units of blood transfused Show forest plot	17	6253	Mean Difference (IV, Random, 95% CI)	‐1.21 [‐1.67, ‐0.75]

Comparison 6. Blood transfusions

Navigate to table in Review

Comparison 7. Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
7.1 Myocardial infarction Show forest plot	8	8219	Odds Ratio (M‐H, Fixed, 95% CI)	1.01 [0.81, 1.26]

7.2 Renal failure Show forest plot	7	9198	Odds Ratio (M‐H, Fixed, 95% CI)	1.07 [0.89, 1.28]

7.3 Infection Show forest plot	8	9219	Risk Ratio (M‐H, Random, 95% CI)	1.00 [0.79, 1.28]

7.4 Congestive heart failure Show forest plot	4	858	Risk Ratio (M‐H, Random, 95% CI)	0.77 [0.24, 2.43]

7.5 Thromboembolism Show forest plot	3	239	Risk Ratio (M‐H, Random, 95% CI)	1.02 [0.11, 9.55]

7.6 Cerebrovascular accident Show forest plot	4	905	Risk Ratio (M‐H, Random, 95% CI)	0.98 [0.22, 4.26]

Comparison 7. Morbidity outcomes in participants undergoing cardiac surgery or vascular surgery, and with acute MI

Navigate to table in Review

Cochrane Review language

Website language

Abstract

Background

Objectives

Search methods

Selection criteria

Data collection and analysis

Main results

Authors' conclusions

PICOs

PICOs

Population

Intervention

Comparison

Outcome

Plain language summary

Is it safe to use lower blood counts (haemoglobin levels) as a trigger for blood transfusion in order to give fewer blood transfusions?

Visual summary

Authors' conclusions

Implications for practice

Implications for research

Summary of findings

Background

Description of the condition

Description of the intervention

How the intervention might work

Why it is important to do this review

Objectives

Methods

Criteria for considering studies for this review

Types of studies

Types of participants

Types of interventions

Types of outcome measures

Primary outcomes

Secondary outcomes

Search methods for identification of studies

Electronic searches

Searching other resources

Data collection and analysis

Selection of studies

Data extraction and management

Assessment of risk of bias in included studies

Measures of treatment effect

Unit of analysis issues

Dealing with missing data

Assessment of heterogeneity

Assessment of reporting biases

Data synthesis

Subgroup analysis and investigation of heterogeneity

Sensitivity analysis

Summary of findings and assessment of the certainty of the evidence

Results

Description of studies

Results of the search

Included studies

Participants

Interventions

Trial setting and design

Excluded studies

Studies awaiting classification

Ongoing studies

Risk of bias in included studies

Allocation

Sequence generation

Allocation concealment

Blinding

Performance bias

Detection bias

Incomplete outcome data

Selective reporting

Other potential sources of bias

Effects of interventions

Primary outcome

30‐Day mortality

Subgroup analysis of 30‐day mortality: restrictive threshold of 7.0 g/dL to 7.5 g/dL versus 8.0 g/dL to 9.0 g/dL

Subgroup analyses of 30‐day mortality: clinical context

Subgroup analysis of 30‐day mortality: prospectively registered versus unregistered trials or trials for which registration was post hoc

Sensitivity analysis