Ethics Statement

This study received ethics approval from the London School of Hygiene & Tropical Medicine.

Aims

The primary objective of the study was to evaluate the association of RBC transfusion with all-cause mortality at 28 days (or hospital discharge) according to predicted risk of death at hospital admission. The secondary objective was to evaluate the association of RBC transfusion with fatal and non-fatal vascular occlusive events.

Sample

The study cohort included all patients from the CRASH-2 clinical trial. The trial included 20,127 trauma patients with, or at risk of, significant bleeding within 8 h of injury, and evaluated the effect of tranexamic acid on all-cause mortality. The trial was undertaken in 274 hospitals in 40 countries. Detailed information on the methods and results of the CRASH-2 trial have been published previously [9].

Outcomes

The primary outcome of this analysis was death from all causes stratified by baseline risk. We also reported specific causes of death (bleeding, head injury, multi-organ failure, myocardial infarction, stroke, pulmonary embolism, and other causes), and we conducted a secondary analysis exploring the association of RBC transfusion with fatal and non-fatal vascular occlusive events including myocardial infarction, stroke, deep vein thrombosis, and pulmonary embolism. All events were measured at 28 days or hospital discharge. Cause of death was defined by the investigators using their clinical judgment.

Interventions and Comparisons

We compared the association of RBC transfusion with the outcomes versus that of no RBC transfusion. For this analysis we compared two groups: those who received at least one RBC transfusion (transfused) versus those patients who did not receive any RBC transfusion (non-transfused).

Statistical Analysis

The characteristics of patients were tabulated and compared according to whether the patient underwent a transfusion. Univariable comparisons were made using a logistic regression model by treating each variable as a categorical or continuous co-variate as appropriate.

For each patient we estimated the predicted risk of death from all causes using a validated model, and categorised patients into four pre-specified strata (<6%, 6%–20%, 21%–50%, and >50%). The prognostic model we used was developed using 20,127 trauma patients with, or at risk of, significant bleeding within 8 h of injury. The model development was conducted with a backward stepwise approach, and the predictors included in the final model were Glasgow Coma Scale, age, heart rate, systolic blood pressure, time since injury, type of injury, and geographical region. Full details of model development and validation have been published elsewhere [10] (see Text S1). Although risk is a continuous variable, we decided to use risk categories for simplifying its use in clinical practice. The risk categories used were identical to the ones reported in the original prognostic model, and the cutoffs were decided with the feedback from prognostic model users and by looking at previous publications [10].

The number of patients and number of deaths were tabulated by transfusion status. Odds ratios (ORs) and risk differences, together with 95% confidence intervals, comparing RBC transfusion to no RBC transfusion were calculated within each of the pre-specified risk categories as defined previously [10]. Interaction tests were conducted using logistic regression to formally assess whether the impact of RBC transfusion differed according to underlying risk, with risk considered as a continuous variable.

Because RBC transfusion practices vary and could be associated with different risks according to the region of the world, we also examined the association with death from all causes separately for four geographical regions.

To identify a potential non-linear interaction between transfusion and baseline risk, patients were also categorised into ten risk groups containing approximately one-tenth of the primary outcome each, and the association of RBC transfusion with death from all causes was evaluated within each of these categories.

We conducted complete case analyses, as the amount of missing data was very low (1%).

Main Findings

The association of blood transfusion with all-cause mortality appears to vary according to the predicted risk of death. We found that in patients with a predicted risk of ≤20%, transfusion was associated with an increase in all-cause mortality, while in those patients with high predicted risk of death (>50%), transfusion was associated with reduced mortality. This pattern from harmful to beneficial association was also found when the association of transfusion with mortality was analysed in ten risk categories, and when we analysed patients from different geographical regions separately. In spite of these findings, because of potential biases inherent in this observational study, our findings should be considered cautiously.

Because an increase in vascular occlusive events was hypothesized as one of the possible mechanisms by which RBC transfusion might be harmful, we conducted a stratified analysis for these outcomes [11],[12]. Although we found that the association of transfusion with fatal and non-fatal vascular occlusive events varies according to the predicted risk of death, transfusion was positively associated with vascular occlusive events (harmful) across all risk strata regardless of the predicted risk of death category.

Strengths and Limitations

Our study has a number of strengths. The CRASH-2 trial was a prospective cohort of bleeding trauma patients, with standardised collection of data on prognostic factors, a large sample size, few missing data, and low loss to follow-up [9]. It included hospitals from low-, middle-, and high-income countries. The prognostic model used in this analysis has shown good performance when externally validated [10]. The study hypothesis was pre-specified, including the risk strata and the direction of the association of transfusion according to the predicted risk of death. The study protocol was registered.

On the other hand, our study has serious limitations. Although our data were from a randomised clinical trial, blood transfusion was not a randomised intervention, and therefore our inferences are vulnerable to confounding [13]. Potential confounding could be suspected because baseline characteristics for transfused and non-transfused patients were different, and those receiving transfusion were at a higher risk of death due to bleeding. Furthermore, there is the possibility of biases acting in different directions depending on the predicted risk of death. For example, in the high-risk group (>50% risk of death), the negative association (beneficial) of RBC transfusion with all-cause mortality could be due to survival bias, since only those who survive are eligible to receive a transfusion [14]. Unfortunately, it was not possible to conduct a time-updated model whereby the period before transfusion was taken into account, since the time of the transfusion was not recorded. Nonetheless, when we attempted to avoid this bias by limiting the analysis to those patients who survived beyond day 0 and therefore had the same opportunity to be transfused, the interaction remained strong.

Conversely, among the low-risk patients, those receiving RBC transfusion might have been at higher risk of death (“confounding by indication”). Propensity scores are useful in observational studies, as they help the researcher to determine whether groups of users and non-users are comparable, and have the potential to reduce confounding by indication [15]. When we conducted an analysis using propensity scores, the results were similar. One potential limitation of using this analytical approach in our study is that there might be a time gap between the variables used in the propensity score (recorded at hospital admission) and the transfusion indication, and this time gap could result in patients being classified as lower risk than they are at the actual time of transfusion. However, the variables included in the propensity scores have been shown to be good predictors of 28-d mortality (which is the transfusion window included in this analysis), so the potential of “misclassifying” to a lower risk category a large proportion of patients using this approach is low [10].

Another limitation of our study is that we could not consider haemoglobin in our analysis. However, our analysis is still informative for current clinical practice, as the indications for a large proportion of RBC transfusions in trauma patients early after hospital admission are based on clinical signs (such as the ones included in our prognostic model) rather than on haemoglobin levels. Furthermore, as mentioned above, the clinical signs included in our prognostic model, such as heart rate and blood pressure, have been shown to be highly predictive of adverse outcomes in patients with trauma and bleeding, and specifically the prognostic model used in our analysis has shown good predictive performance [10].

Finally, the association of blood transfusion with all-cause mortality could have been influenced by the type of blood product received (i.e., whole blood or RBCs) in different countries. Although we did not have this information available, the same pattern and strong evidence for interaction according to baseline risk was found in all the geographical regions. Furthermore, when we further adjusted by use of platelets, fresh frozen plasma, and cryoprecipitate and by country, results were similar.

Comparison with Previous Studies

Previous studies have shown that RBC transfusions are associated with an increased risk of complications in trauma patients. A systematic review evaluating the association of RBC transfusion with mortality in critically ill patients identified 45 observational studies, and in 42 of them the risks of RBC transfusion outweighed the benefits [6]. The studies included were observational and therefore prone to different types of bias, and, importantly, they did not analyse the association of RBC transfusion with mortality according to baseline risk.

The findings from another systematic review that evaluated the effect of liberal versus restricted transfusion thresholds (haemoglobin or haematocrit triggers) in critically ill patients support the use of restrictive transfusion triggers (haemoglobin levels between 70 and 80 g/l) [7]. Nonetheless, haematocrit level is not a reliable indicator of the extent of bleeding in the early hours after hospital admission, when a substantial proportion of RBC transfusions occur, and clinicians instead use clinical signs and their clinical judgment to decide whether or not to mandate a RBC transfusion. To the best of our knowledge, this is the first study to evaluate the association of RBC transfusion with all-cause mortality stratified by predicted risk of death, using simple clinical variables routinely available at hospital admission.

Implications for Practice and Research

Current recommendations for trauma and critically ill patients state that transfusion is indicated for patients in “haemorrhagic shock” or who are haemodynamically unstable, and that a restrictive strategy (transfusion when haemoglobin <70 g/l) is as effective as a liberal strategy (transfusion when haemoglobin <100 g/l) for haemodynamically stable patients [16],[17]. It is important to highlight that only a small proportion of trauma patients would present with haemorrhagic shock, and the vast majority of trauma patients might be unstable but not at very high risk of death [18],[19]. Although RBC transfusion might be life-saving for patients with haemorrhagic shock, uncertainty remains about the best early transfusion strategy in other patients. Our study suggests that blood transfusion could be harmful for those patients whose predicted risk of death is low. However, as our study was observational, important biases cannot be ruled out, and we cannot claim a causal link. Therefore, this hypothesis should be prospectively evaluated in a randomised controlled trial.