Risk of Bias table was generated using Excel rather than RevMan due to researcher preference. Authors were not contacted for missing data due to time limitations. Data was re-extracted in order to double check the appropriateness of infectious diseases outcome definition given that this was perceived to be the highest risk of bias within the individual studies. We performed an additional sensitivity analysis excluding those with high detection or unclear bias ratings. Due to concern for differing outcomes due to clinical heterogeneity, we also performed a sensitivity analysis excluding urinary tract infections as an outcome. Additionally, to assess for different effects for clinical heterogeneity in more severe infections, we performed subgroup analysis outcomes of pneumonia and bacteremia. Total infections were counted from individual studies reporting infection totals or if that was not available, individual infectious outcomes reported were manually totaled. When extracting data from individual studies, total infectious outcomes was preferred to summing individual outcomes data (e.g. pneumonia and bacteremia) in order to prevent under- or over-counting.
Study Selection
Our PubMed search retrieved 2,679 abstracts. 150 full-text records were screened for eligibility. From abstract screening, 2,529 studies were excluded. 150 studies were included for full-text screening. 19 trials were included for our meta-analysis after 131 studies were excluded (Figure 1). Prior to exclusion, articles duplicate reporting of results underwent full-text review to ensure additional data was not available from the individual manuscripts. Screening of clinicaltrials.gov of the 692 results led to 21 eligible protocols (see appendix A). None of these had results available that were additional to what was available through our PubMed search.
Study Characteristics
Types of Participants 14372 patients were included in total. Enrollment dates ranged from February 1997 to March 2017. Study sites were mostly in the United States there studies also included sites in Australia (1), Brazil (3), Canada (2), China (2), Colombia (1), Denmark (2), Israel (1), Egypt (1) Germany (1), Greece (2), India (2), Malaysia (1), New Zealand (1), Romania (1), Singapore (1), South Africa (1), Spain (2), Switzerland (1), and the United Kingdom (3). Average age of population ranged from 28 to 83.3 in the restrictive group and from 31 to 81.8 in the liberal group(16-18). Four studies included medical patients. These included critical care patients, patients with a gastrointestinal bleed, those that had traumatic brain injury and lastly those undergoing left heart catheterization for evaluation of ischemic injury(19-22). There were 15 surgical studies included in our analysis. Of these, six were cardiovascular surgical patients and six were orthopedic surgical patents.
Types of Interventions and Comparators All trials in the restrictive group had a goal transfusion threshold which ranged narrowly from 7 g/dL (5 trials) to less than 8 g/dL (9 trials). In the liberal group, thresholds ranged from 8.5 to less than 10 g/dL. 10 of 19 trials used a 10g/dL liberal threshold. Six studies reported using leukoreduced packed red blood cell transfusions. The other studies did not report whether or not the transfusions were leukoreduced. No other pre-treatment (e.g. irradiation or CMV-matching) was reported in any of the studies. Among studies that reported amounts transfused in medians, the amount ranged from 0 to 8 units in the restrictive group and from 2 to 16 units in the liberal groups.
Types of Outcomes The duration of follow-up for infectious outcomes included 5 (1 trial), 30 days (5 trials) 45 days (1 trial), 60 days (3 trials), and 90 days (2 trials). Eight studies followed patients for infectious outcomes until discharge and did not follow patients for outcomes after. The length of stay for individuals in these studies would vary but was less than a month in all studies. Eleven studies reported events of pneumonia. Five studies reported cases of sepsis and four studies reported the outcome of bacteremia. Five studies reported urinary tract infections while four studies reported wound infections. Eight studies reported total infections based on varying criteria.
Risk of Bias of Included Studies (Figure 3. Appendix F)
Bias was assessed using an adapted Cochrane Risk of Bias Tool for Randomized Control Trials for the following types of bias: selection bias, performance bias, detection bias, attrition bias, and reporting bias(11). Refer to protocol appendix for specific questions on form used to assess risk of bias.
Selection Bias: Randomization and Allocation Concealment
Two studies were deemed high risks of bias and six studies had unclear risk of bias. 11 studies were deemed low risk of bias due to their randomization methods. Only one study was found to have high risk of concealment bias(23). Five studies had unclear risk of bias while 13 were deemed low risk of allocation bias.
Performance Bias
Six studies were rated to have high risk of performance bias and seven studies had unclear risk of performance bias. Six studies were rated low risk of performance bias.
Detection Bias
Detection bias had the highest number of high bias trials out of the bias categories. Detection bias was rated high in seven trials and unclear in six trials. The remaining six trials were rated low risk of bias.
Attrition Bias
Attrition bias was uniformly low for all trials except one rate high bias which had an 42 people removed from the analysis after randomization due to difficulty reading the electrocardiogram strip(24). In general, trials had very few lost to follow-up.
Reporting Bias
One study was high risk of reporting bias due to primary outcome that was different the pre-specified outcome on clinicaltrials.gov(22). Reporting bias was unclear in seven trials. Eleven trials were rated low risk of bias.
Assessment of Heterogeneity
Clinical Heterogeneity Participants had wide ranging causes of illness or hospitalization. Medical patients had more long term comorbidities and were older (up to a mean of 83.3) compared to one study included trauma patients with an average age of 28. Study sites were primarily tertiary care centers though sites were located in North America, Europe, Asia, Australia and Africa.
There was a large amount of heterogeneity in the infectious outcomes of interest. Composite outcomes including multiple combinations of infectious outcomes. Additionally, for individual outcomes such as pneumonia or wound infection clinical definitions varied. In eight studies the definitions for infection for not reported so therefore the heterogeneity is unclear but with great potential for different methods of diagnosis. There was also variability in which outcomes such as urinary tract infections or bacteremia were recorded or reported.
Methodological Heterogeneity Ten studies scored high for risk of one or more categories of bias (Figure 2.). The risk of bias was not high enough in any of these studies to merit exclusion from our analysis. The study with the highest ratings of risk of bias reported outcomes ambiguously. This study also had the highest attrition bias of the studies. Besides this study, the loss to follow-up was uniformly low due to the design of observing patients until discharge or until a short period after discharge. Another study was rated as high performance bias and detection bias is noted to have infectious outcome results that appear to be outliers in the Forest plot. Studies had varying primary outcomes and also varying international settings which could affect the way infectious outcomes are assessed.
Statistical Heterogeneity We assessed for statistical heterogeneity with the Q statistics, I-squared (I2), and Tau-squared (T2). Heterogeneity measured by I2 was uniformly less than 30% for overall, subgroup analyses, and sensitivity analyses. Measures of statistical heterogeneity are further described in the section of quantitative synthesis.
Qualitative Synthesis Our studies included a wide-ranging age range from 28 to 83.3, but excluded pediatric patients (Table 1). Study sites were mostly in North America but sites included 18 other countries. One study included the majority of study sites that were from upper middle-income countries.(20) Although leukoreduction is known to affect the risk of infection with transfusion only 6 studies reported using leukoreduced packed red blood cell transfusions and the others did not report the frequency of leukoreduction(25). Therefore, it is unclear what how much heterogeneity is related to the practice of leukoreduction. Clinical heterogeneity accounted for the largest source of heterogeneity in these trials. Though our analysis did not include profoundly immunosuppressed patients with hematologic malignancy, there would be expected difference in the immunocompetence between patient populations. For example, middle-aged patients admitted to the intensive care unit may be more profoundly immunosuppressed than a younger population admitted for injuries from trauma(19, 21). These patient populations may additionally have disparate pathophysiology such as hip fractures compared to gastrointestinal bleeds(20). The largest amount of heterogeneity was that due to differential measurement of infectious outcomes. Composite outcomes included different combinations of infectious outcomes. Clinical definitions varied, and different types of infections were reported. While 11 studies reported pneumonia as an outcome, only four studies reported the outcome of bacteremia. Across studies and within each study, mild infectious outcomes such as urinary infections were counted as equal to more severe infections such as bacteremia or pneumonia. Additionally, due to the inherent variability in how infections are diagnosed, and the unmasked nature of the studies, there is certainly heterogeneity in the way infectious outcomes such as wound infections or urinary infections would be diagnosed in clinical practice.
Quantitative Synthesis (Appendix E-I. Figures 2-6)
Primary outcome The overall pooled risk ratio for infectious outcomes between restrictive and liberal transfusion groups was nonsignificant at 1.04 (CI 0.94 to 1.16) (Figure 3). The heterogeneity statistics were low with an overall I2 of 13.3% and T2 of 0.005. The p-value for the Q statistic was 0.29. By sorting by year, there appeared to be a Proteus phenomenon with regression towards the mean with time.
Subgroup Analyses Pre-specified subgroup analyses were performed for medical versus surgical patient populations which also showed no significant difference in the risk of infection in the surgical (RR 1.04, CI 0.73 – 1.02) or medical subgroups (RR 0.86, 0.73 – 1.02) (Figure 3). The heterogeneity statistics were low. The surgical population had an overall I2 of 8.2% and T2 of 0.0035. The p-value for the Q statistic was 0.36. The medical population had very low heterogeneity with an overall I2 of 0%, T2 of <0.0001, and Q statistic p-value of 0.64.
Surgical subgroups
Pre-specified subgroup analyses further divided surgical subgroup populations between cardiovascular and non-cardiovascular surgery (6 of 9 were orthopedic populations) (Figure 4). The non-cardiovascular surgery subgroup also showed no significant difference in the risk of infection (RR 0.93, CI 0.71 – 1.23) with low overall I2 of 24.1 %, T2 of 0.038, and a Q statistic p-value of 0.23. The regression towards the mean over time appears more uniform in this subgroup. The cardiovascular surgery subgroup also showed no significant difference in the risk of infection (RR 1.09, CI 0.92 – 1.28) with low overall I2 of 11.0 %, T2 of 0.005, and a Q statistic p-value of 0.29.
Clinical outcome subgroups
A pooled analysis was performed using the 11 studies which reported pneumonia data with 3082 patients. The risk of pneumonia was non-significant (RR 0.97, CI 0.82 – 1.15) with low statistical heterogeneity (I2 = 31.1%, Q statistic p = 0.21) (Figure 5). An analysis was performed using the 4 studies which reported bacteremia data with 1,690 patients. The risk of bacteremia was nonsignificant (RR 0.89, CI 0.67 – 1.18) with low statistical heterogeneity (I2 = 0.0%, Q statistic p = 0.80) (Figure 6).
Sensitivity Analyses Because only one study with a high proportion of oncologic patients was included in our meta-analysis, we performed a sensitivity analysis excluding that trial. The relative risk remained non-significant at 0.98 (CI 0.89 – 1.08). Excluding studies with high or unclear risk of detection bias, the relative risk of infection remained non-significant at 1.07 (CI 0.91-1.25) (Figure 7). Excluding studies with high or unclear performance bias the relative risk of infection remained non-significant with a relative risk of 1.04 (CI 0.81 – 1.34) (Figure 9). Excluding studies that reported use of leukoreduced blood, relative risk of infection remained non-significant at 1.03 (CI 0.91 – 1.16) (Figure 8). After excluding studies that reported urinary tract infection as infectious outcomes, relative risk of infection also remained non-significant at 0.97 (CI 0.83 – 1.13).
Investigation of Reporting Bias
Of 19 studies were registered on clinicaltrials.gov. One study was noted to have the primary outcome of trial feasibility rather than the reported composite primary outcome and was therefore rated as high risk of bias(22).
Discussion In our systematic review and meta-analysis, using 19 studies that fit our inclusion criteria we found no evidence of decreased risk of infection with restrictive transfusion strategies compared to liberal strategies in adults. Although there were limitations with clinical heterogeneity, our finding of no difference was stable throughout our subgroup and sensitivity analyses used to address this heterogeneity. A prior meta-analysis performed in 2013 found a decreased risk of infection with an RR of 0.72 (95%CI, 0.53-0.97) for severe infections. This meta-analysis pooled studies which included pediatric patients including neonates receiving critical care. This study also included studies with patient populations receiving chronic transfusions, of whom would be more likely to have already developed iron overload which leads to immunosuppression(26). This often cited conclusion was from a subgroup analysis was also changed post-publication after there were multiple methodological concerns(27). Our findings are consistent with their reverse post-publication conclusion that there is no significant decrease in overall health-care associated infections. Unlike this study our patient population was more heterogeneous. Also unlike this study, we performed additional sensitivity analyses removing studies with detection bias and only looking at specific severe infections (i.e. pneumonia and bacteremia), which supported our conclusion. A more recent Cochrane review evaluating the risks and benefits of different transfusion thresholds which included adults and children populations evaluated specific infectious outcomes of severe infections (i.e. pneumonia, “sepsis/bacteremia,” and “pneumonia or wound infection”) and also found no significant difference in the risk of infection with restrictive thresholds(6).
Limitations of Included Studies
Detection bias was a limitation for all studies included in our review. Infection was poorly defined in 11 of our 19 included studies. There was a high risk for low inter-rater agreement or agreement between studies. Differences in country location, patient population, and age also increase the risk for differences in infection diagnoses. None of the studies masked patients or physicians which increases the risk of performance bias. Physicians may have been more or less likely to make infectious outcome diagnoses depending on treatment arm. This was due to the methodological difficulty in blinding patients or physicians to the process of transfusions. Yet, future studies could potentially mask outcome accessors by using physicians other than the treating physician. Only one study used a masked infectious diseases specialist to adjudicate infection outcomes(22).
Limitations of this Review
Despite using more specific inclusion criteria than prior reviews, there remained considerable clinical heterogeneity among our review. As was a limitation of our index review, including both minor infectious outcomes (i.e. urinary infections) with severe infections (e.g. pneumonia or bacteremia) in the primary outcome raises the concern of decreased ability to extrapolate from these findings.(28) However, in our sensitivity analysis our result of non-significance remained unchanged after excluding urinary tract infections. Our conclusion also remained when looking specifically at pooled results for bacteremia or pneumonia. In order to decrease the heterogeneity of our studies we excluded pediatric populations and certain populations with higher degrees of immunosuppression (hematologic disorders in particular). Therefore, our conclusions may not apply to these populations. Additionally, the studies included no sites from low or low-middle income countries, so our conclusions may be limited to patients in high and upper-middle income countries.
Conclusions In conclusion, despite clinical heterogeneity and methodological limitations, our meta-analysis provides the strongest support of no decreased risk of infections with restrictive strategies among adult populations. Yet, this still supports avoiding the resource-intensive use of a liberal transfusion strategy from a public health standpoint. Liberal transfusion strategies also carry other well-defined clinical risks not covered in our review such as transfusion-related acute lung injury(26). In addition to the recent Cochrane study, our study provides additional support that transfusion strategy does not affect the risk of infection among adult hospitalized patients.