Hemoglobin levels and transfusions in neurocritically ill patients: a systematic review of comparative studies

We designed a search strategy for Ovid MEDLINE (1949 to the present), the Cochrane Central Register of Controlled Trials (1974 to Issue 1, 2011), as well as Embase and Embase Classic (1974 to the present). Abstracts and conference proceedings were searched in BIOSIS previews (1926 to the present) and Web of Science (1898 to the present), whereas the grey literature was searched by using Google Scholar. We sought both randomized controlled trials (RCTs) and comparative nonrandomized studies, both prospective or retrospective. No restriction based on language, year, or type of publication was applied. Keywords and Medical Subject Headings (MeSH) terms (or their EMTREE equivalents) pertaining to the population (neurocritical care) and to the exposure (hemoglobin levels, RBC transfusion, anemia) were combined to form the search strategy (Additional file 1). We used clinicaltrials.gov, controlled-trials.com, and strokecenter.org websites to identify unpublished and ongoing studies. Reference lists from relevant reviews and included articles were manually searched to identify missed studies. The last iteration of the search process was completed on January 31, 2011.

We included comparative studies evaluating the effect of hemoglobin levels on clinical outcomes of neurocritically ill patients admitted to an ICU. Studies were included if at least two different hemoglobin thresholds, levels, targets, or RBC transfusion strategies were compared. Neurocritical conditions encompassed but were not limited to subarachnoid hemorrhage (SAH), stroke, traumatic brain injury (TBI), intracerebral hemorrhage (ICH), and any cerebral neurosurgical conditions. Studies on sickle cell anemia and scoliosis surgery were excluded. We also excluded studies in neonates (< 28 days), but all other age groups were considered.

Two independent reviewers (PD, MHT) screened the studies identified from the systematic search. Non-English language articles were translated as required. A Cohen kappa statistic was calculated to quantify the interrater agreement concerning inclusion of studies. In case of discrepancy, a third reviewer (AFT) was involved to settle the disagreement. Search results from Web of Science, from grey literature sources, and from reference lists of identified studies were reviewed and adjudicated by a single reviewer (PD).

A standardized abstraction form was developed and tested before data collection. Data abstraction was conducted independently, and in duplicate, by two reviewers (PD, MHT). When judged necessary, missing information was requested from corresponding authors.

The primary outcome measure was all-cause mortality at any given time point. Secondary outcomes were neurologic status (irrespective of the scale used), ICU length of stay, hospital length of stay, duration of mechanical ventilation, surrogate measures of brain oxygen delivery, complications (including vasospasm and multiple organ dysfunction score) [22], and serious adverse events (thromboembolic events, myocardial infarction, pulmonary edema or volume overload, transfusion-related acute lung injury (TRALI), and infection). Data pertaining to the study design were also retrieved, as well as characteristics of patients that could act as confounders and affect the outcomes of interest, including age, sex, disease severity, comorbidities, incidence of hypoxemia, incidence of hypotension, and baseline hemoglobin. Information on blood transfusion and the nature, timing, and frequency of co-interventions (hemodilution, blood-conservation strategies, erythropoietin analogues, and use of other blood products) were recorded.

Two reviewers (PD, MHT) independently evaluated the risk of bias in included studies. We used the Cochrane Collaboration tool for assessing risk of bias in RCTs, which was customized for the focus of the review [23]. We judged the overall risk of bias of individual studies as low, moderate, high, or unclear [23]. Additionally, we used the Downs and Black checklist [24] to assess the methodologic quality of both RCTs and nonrandomized studies. This checklist has been validated for reliability and external validity. We put emphasis on how study authors dealt with confounding factors mentioned earlier in nonrandomized studies. The last item of the Downs and Black checklist is an assessment of the adequacy of the sample size of the study, which we performed assuming a two-sided P value of 0.05, 80% power, and a 10% relative difference for the main outcome measure.

A meta-analytic approach was planned by using Mantel-Haenztel random-effect models, if deemed appropriate. We presented outcome data by using odds ratios with 95% confidence intervals. An odds ratio of less than 1 suggests a lower rate of the event among the patients exposed to lower hemoglobin levels. Continuous data, such as length of stay and physiologic parameters, were reported as mean or median. We summarized continuous data as mean difference with 95% confidence intervals. We converted hematocrit to hemoglobin by using a standard published equation [25] (Hb [g/dl] = Hct [%]/3). All data were compiled in Review Manager (version 5.0; The Cochrane Collaboration). A priori sensitivity analyses were planned to explore heterogeneity in study findings, based on age, type of neurocritical condition, risk of bias, and presence of co-interventions.

Our literature search yielded 4,310 studies from major databases after removal of duplicate records (Figure 1). Seven studies were deemed potentially elligible [6, 26‐31], but one was excluded, as it reported only summary data for the overall group [31]. The authors of the latter study were contacted and confirmed that data from each study group were unavailable. Therefore, six studies were included (number of patients = 537) [6, 26‐30]. The overall interrater agreement between reviewers on inclusion was high (Cohen kappa = 0.80). Discrepancies were resolved with the input of a third reviewer on two occasions.

Data for the subgroup of neurocritically ill patients from a previous RCT on transfusion thresholds in pediatric ICUs were obtained from the authors for one study [6]. Attempts to contact the authors of two other studies to obtain data on subgroups of published RCTs were unsuccessful [32, 33]. Two abstracts of potentially relevant studies were retrieved through the review of conference proceedings; however, one reported data on a study already included in our review [29, 34], whereas the other abstract [35] had not yet been published as a full report. Additional data for this study could not be obtained from the corresponding author. Of note, one relevant ongoing study was identified [36].

Among the six studies included, three were from the United States [28‐30], two from Canada [26, 28], and one was completed in Switzerland [27] (Table 1). All studies were published in peer-reviewed English language journals within the last 5 years, and the period over which the studies were conducted spanned from 1994 to 2011. Only one study involved pediatric patients [6], whereas the other five studies took place in adult ICUs. One study was an RCT [30], one was a subgroup analysis of an RCT in the overall critically ill adult population (Transfusion Requirements in Critical Care (TRICC) trial) [26], one was an unpublished subgroup analysis of an RCT in the critically ill pediatric population (Transfusion Requirements in the Pediatric Intensive Care Unit (TRIPICU) trial) [6], and three were retrospective cohort studies [27‐29]. The main diagnosis of patients was traumatic brain injury (n = 4) [26‐29] and subarachnoid hemorrhage (n = 1) [30]. The subset data used from the TRIPICU study [6] included patients representing different neurocritical care conditions: TBI (n = 36), ICH (n = 11), brain tumors (n = 3), neurosurgery (excluding scoliosis surgery) (n = 9), cerebral edema (n = 5) and other space-occupying injuries (n = 2).

All six studies compared two groups of patients whose targets or observed hemoglobin levels differed, having or having not received an RBC transfusion (Table 2). Three studies, all of which were RCTs, described mandatory transfusion protocols [6, 26, 30] based on specified hemoglobin thresholds. Two observational studies [28, 29] compared groups of patients who received or did not receive RBC transfusion for a specified baseline hemoglobin range. One cohort study compared patients who reached and did not reach a target hemoglobin level after the resuscitation phase of TBI [27].

Among studies, lower hemoglobin levels or thresholds ranged from 7 g/dl to 10 g/dl, whereas higher hemoglobin levels or thresholds ranged from 9.3 to 11.5 g/dl. One study divided patients on the basis of achieved hemoglobin levels above or below 9.3 g/dl [27]. The duration of exposure to these hemoglobin levels varied among studies. In three of six studies [6, 26, 30], hemoglobin levels were maintained during the entire ICU stay. In the remaining three studies [27‐29], hemoglobin levels were measured at only one point in time, and thus the exposure was not necessarily sustained. Among the four studies [6, 26, 27, 30] for which the mean hemoglobin levels during the period of exposure could be obtained, all observed a statistically significant difference of at least 1 g/dl (range, 1.2 to 2.78 g/dl) between the two comparison groups. The follow-up period varied between 28 days and 6 months among studies.

One study reported the use of triple-H (hemodilution, hypertension, hypervolemia) therapy in patients with symptomatic vasospasm afer a subarachnoid hemorrhage [30]. The incidence of vasospasm was not significantly different in the higher-hemoglobin and lower-hemoglobin groups (24% and 22%, respectively; P = 1.00). The use of fresh-frozen plasma was reported in two studies [6, 28]. In one study, only two patients (one in each group) received fresh-frozen plasma [6]; whereas in the other, a mean of one more unit was given in the higher-Hb group (P = 0.046) [28]. No other relevant co-interventions susceptible to interfere with hemoglobin levels and outcomes were documented.

All RCTs and subgroups of RCTs presented an overall low risk of bias (Table 3) [6, 26, 30]. Given the nature of the intervention, no study was blinded to allocation during ICU management. All studies reported adequate allocation concealment before and during enrollment. Data analysis of these three studies respected the intention-to-treat principle. Nonrandomized studies generally were of lower methodologic quality than RCTs and subgroup analyses of RCTs. Two of the six studies [27, 30] did not adjust study estimates for important confounding factors, such as severity of the baseline condition.

Given the substantial heterogeneity observed in study designs and participants of included studies, a formal meta-analysis was considered to be inappropriate, and data were not pooled. Study data were therefore descriptively synthesized. For the same reason, we did not quantify statistical heterogeneity, and we did not conduct sensitivity analyses as planned.

Five studies presented data on mortality [6, 26‐29]. None showed a statistically significant effect of lower Hb levels when compared with higher Hb levels (Table 4).

One study reported short and mid-term neurologic outcomes, as defined by the National Institutes of Health (NIH) Stroke Scale (14 days) and the modified Rankin scale (14 and 28 days, and 3 months) [30], and one study evaluated long-term (≥6 months) functional neurologic outcome by using the extended Glasgow outcome scale (GOSe) [29]. When looking at the former, the median scores on the NIH Stroke scale were 1 (Q1 to Q3: 0 to 9.75) in the higher-Hb group versus 2 (0 to 16) in the lower-Hb group (P > 0.10). At 14 days, 13 patients in the higher-Hb group were considered "independent" on the modified Rankin Scale as opposed to 10 in the lower-Hb group, but the difference was not statistically significant (P = 0.25). Findings at 28 days (14 versus 16 patients; P = 0.34) and at three months (18 versus 20 patients; P = 1.00) were similar and did not achieve statistical significance [30]. Compared with a higher-Hb target group (patients transfused when Hb was between 7 and 10 g/dl), investigators of the second study observed a statistically significant increase in the GOSe score (5.7 versus 3.9; P < 0.0005, higher score meaning better outcome) at six months in patients in the lower Hb target group (not transfused when Hb was between 7 and 10 g/dl). This difference remained significant after adjustment was made for the Glasgow Coma Scale score on admission (5.4 ± 0.3 versus 4.1 ± 0.2; P = 0.0005) [29].

Two studies [6, 30] reported the duration of mechanical ventilation and observed no significant difference in the number of days on mechanical ventilation between the higher and lower hemoglobin-level groups (mean difference, 0.57 days (95% CI, -1.78 to 2.92) and -0.63 days (95% CI, -1.85 to 0.60), respectively).

Four of the six studies [6, 26, 28, 29] reported length of stay (Table 5). A significantly different mean ICU length of stay in favor of the lower-Hb group (11.0 days) versus the higher-Hb group (16.7 days; P = 0.02) was observed in one study [28]. In this study, a similar effect was observed for hospital stay, but it did not reach statistical significance. A significant difference in mean hospital stay was, however, observed in one study (mean difference, -11.40 (95% CI, -16.0 to -6.70)) in favor of the lower-Hb group [29]. In contrast, results of two other studies did not show a significant difference in ICU length of stay between the two study groups [6, 26].

Two studies addressed organ failure by using the multiple organ dysfunction score (MODS) [6, 26]. In both studies, neither the progression of organ failure nor the emergence of a new MODS was significantly different between the restrictive and the liberal groups. In the subgroup of neurocritically ill patients from the TRIPICU study, the proportion of patients with new or worsening MODS was 16.6% in the restrictive group versus 8.3% in the liberal group, but this difference did not achieve statistical significance (P = 0.45) [6]. In the subgroup analysis of patients with traumatic brain injury from the TRICC trial, the worsening of MODS was also similar in both intervention arms (3.4 restrictive versus 4.5 liberal; P = 0.49) [26].

We assessed for the reporting of myocardial infarction, pulmonary edema and volume overload, transfusion-related acute lung injury, thromboembolic events, and infections (Table 6). At least one serious adverse event was included as a secondary outcome in five studies [6, 26‐28, 30], but no study reported a systematic method to screen for serious adverse events. The reported incidence of adverse events is shown in Table 6.