Introduction

Severe acute kidney injury (AKI) occurs in approximately 6% of patients admitted to an intensive care unit (ICU) [1] and in up to 19% of patients with vasopressor-dependent septic shock [2]. For such individuals, mortality is approximately 60% [1], and survivors are at increased risk of requiring permanent renal replacement therapy (RRT) [3]. Two multicenter randomized controlled trials (RCTs) [4, 5] and two meta-analyses [6, 7] have demonstrated that increasing the dose of RRT above 20 to 25 mL/kg/h of effluent flow for continuous renal replacement therapy (CRRT) or increasing intermittent dialysis frequency beyond alternate days does not improve survival.

In addition to dose, the mode of clearance is also a modifiable component of the RRT prescription that may affect patient outcomes. Convective clearance and diffusive clearance, delivered by hemofiltration and hemodialysis, respectively, can be provided by all continuous and some intermittent RRT machines. Despite similar clearance of small molecules, hemofiltration is reported to achieve higher clearance of medium-sized to larger molecules compared to hemodialysis [8]. Consequently, it is postulated that hemofiltration might benefit critically ill patients with AKI by better clearing large toxic inflammatory cytokines [9]. In the absence of a large, suitably-powered, randomized trial demonstrating the superiority of one mode over the other, practice surveys have shown variability in mode selection among countries and regions [10–15]. Therefore, our objective was to conduct a systematic review and meta-analysis of all RCTs comparing the effects of convective clearance (using hemofiltration) to diffusive clearance (using hemodialysis) in patients with AKI on clinically important outcomes.

Materials and methods

Literature search

We searched OVID versions of MEDLINE, EMBASE Classic and EMBASE and the Cochrane Central Register of Controlled Trials (from inception to June 2012) without language restrictions using a previously described search strategy [6]. We also searched abstracts from critical care and nephrology professional society conferences, including: Society of Critical Care Medicine (2004 to 2012), European Society of Intensive Care Medicine (2001 to 2011), International Symposium of Intensive Care and Emergency Medicine (2004 to 2012), American Thoracic Society (2004 to 2012), American College of Chest Physicians (2003 to 2011), American Society of Nephrology (2003 to 2011), and the European Renal Association - European Dialysis and Transplant Association (2002 to 2012). We also searched bibliographies of included studies and personal files. Two reviewers independently reviewed all citations and retrieved the full text of any citation considered potentially relevant by either reviewer. We attempted to contact selected authors of included studies for clarification of methods and to obtain additional data, where required.

Study selection

Two unblinded reviewers assessed full-text reports and included studies meeting the following criteria: (1) design: either parallel-group (patients assigned to only one treatment) or cross-over (each patient received both treatments in random order) randomized or quasi-randomized (for example, assigning patients in alternating fashion or by hospital registry number) controlled trial, (2) population: adult or post-neonatal pediatric patients with AKI requiring RRT, (3) intervention: hemofiltration compared to hemodialysis, with both modes applied using continuous or intermittent RRT; trials comparing continuous to intermittent RRT were excluded, (4) outcomes: all-cause mortality (primary outcome) or other clinically important outcomes (see below) for the parallel-group trials; or group-specific filter duration, or clearance or plasma concentration measurements of cleared substances for both parallel-group and crossover trials. We also included trials of hemodiafiltration (that is, combined dialysis-filtration) compared to either hemodialysis or hemofiltration in sensitivity analyses as explained below.

Data abstraction and validity assessment

Two unblinded reviewers independently abstracted data from included trials, including study population (number of centers, age, gender, illness acuity proportion with oliguria and sepsis, baseline creatinine and urea), RRT methods (modality, dose, duration of and criteria for starting/stopping study RRT), outcomes (mortality, RRT duration, RRT dependence in survivors, hemodynamics (for example, vasopressor doses), evolution of organ dysfunction, filter duration, measured clearances and plasma concentrations of metabolites) and study quality (including method of sequence generation and allocation concealment, intention-to-treat analysis, loss to follow-up for the outcome of mortality, and early trial stopping for efficacy before the planned enrollment was completed). For the crossover trials we also assessed whether carry-over effects and washout periods were addressed and whether paired data analyses were performed [16]. Disagreements between reviewers regarding study selection and data abstraction that remained after author contact were resolved by consensus.

Data analysis

Our primary outcome was all-cause hospital mortality and, if not available, then mortality at 90, 60, 30 or 28 days after randomization, at ICU discharge or after stopping RRT (in descending order of preference). Secondary outcomes included RRT dependence among survivors at the latest time point available (with the same preferred order of time point as for the mortality analysis), RRT duration until renal recovery or death, filter duration, clearance of selected solutes and plasma concentration measurements.

Binary outcomes are reported as risk ratios (RR) and continuous outcomes using weighted mean differences (MD, a measure of absolute change) or ratio of means (RoM, a measure of relative change) [17]. We used Review Manager 5.1 (The Cochrane Collaboration, Oxford, England, UK) to calculate pooled outcome measures. We considered (two-sided) P <0.05 as statistically significant and reported individual trial and summary results with 95% confidence intervals (CIs). We used random-effects models, which incorporate between-trial heterogeneity and give wider confidence intervals when heterogeneity is present, to pool data. We assessed statistical heterogeneity among trials using I2, the percentage of total variability across studies due to heterogeneity rather than chance [18, 19], and used published thresholds to ascribe low (I2 = 25 to 49%), moderate (I2 = 50 to 74%), and high (I2 ≥75%) heterogeneity [19]. Continuous variables are expressed as mean ± standard deviation, unless otherwise indicated. Because the crossover trials generally did not report within-patient differences, we used the unpaired group-specific means to pool data, recognizing that this approach reduces the statistical power to detect differences [20].

For each outcome, the primary analysis included only trials in which one group was treated only with hemofiltration and the other group only with hemodialysis, with both groups receiving RRT providing similar (that is, within 20%) small-molecule clearances. In assessing equivalency of doses, we determined whether trial authors adjusted for the reduced dose of hemofiltration that results from blood dilution by pre-filter replacement fluid, where applicable. In sensitivity analyses for each outcome, we included data from trials in which doses differed by greater than 20% between the convection and diffusion groups, and from trials comparing combined filtration-dialysis modes to either filtration or dialysis. For trials in which one group was treated with combined filtration-dialysis, this group was considered either as filtration (if compared to a dialysis-only group) or dialysis (if compared to a filtration-only group).

To assess for publication bias we planned to visually examine a funnel plot of study precision versus treatment effect on mortality for evidence of asymmetry, assuming ≥5 trials in the analysis.

Because this study reports an analysis of published data, ethical approval was not required.

Results

Study flow

Our search strategy identified 6,657 citations, 6,324 from Medline, EMBASE Classic and EMBASE and 333 from the Cochrane Central Register of Controlled Trials. We retrieved 41 articles for detailed evaluation, of which 19 studies met criteria for inclusion [21–39] and 22 were excluded [40–61] (Figure 1). They included 10 parallel-group trials [21–30] and 9 crossover trials [31–39]. Among the parallel-group trials, authors of 9 included trials [21, 23–30] provided additional methodological or clinical data, and the author of the 10th trial [22] informed us that no additional information was available. In addition, authors of two studies informed us either that the vast majority of enrolled randomized patients did not have renal failure [40], or that it was not a randomized controlled trial [41] (reference [62] is the full paper for the abstract reference [41]), leading to exclusion of these studies.

Figure 1
figure 1

Flow chart of study selection. *In addition, author contact confirmed that mortality and other clinical outcomes not collected, and clearance rates not measured for this trial. Abbreviation: RCT, randomized controlled trial.

Full size image

Description of included studies (Table 1)

Enrolled patients had high illness severity. The mean or median acute physiology and chronic health evaluation (APACHE) II score [63] was ≥20 or simplified acute physiology score (SAPS) II [64] was ≥60, as reported in 8 of 10 parallel-group trials [21–27, 29] and 3 of 9 crossover trials [37–39]. One parallel-group trial reported only the mean sequential organ failure assessment (SOFA) [65] score of 16 [30]. Where reported, the vast majority of patients required mechanical ventilation or vasopressors [21, 23, 25, 29–32, 35, 36]. AKI was defined by abnormal biochemistry (serum creatinine or urea) or a complication of oliguria (such as volume overload, hyperkalemia, acidosis or uremic symptoms) in 7 of 10 parallel-group trials [21, 23–26, 29, 30] and 3 of 9 crossover trials [34, 35, 38]. The remaining nine trials enrolled patients with AKI requiring RRT without reporting specific AKI criteria [22, 27, 28, 31–33, 36, 37, 39]. Five trials explicitly excluded patients with chronic kidney disease, defined by pre-morbid creatinine or creatinine clearance [23, 25, 26, 29] or recent dialysis [30],

Table 1 Patient characteristics in included trials
Full size table

Description of RRT interventions (Table 2)

Seven of the 10 parallel-group trials used continuous RRT [21, 22, 24–27, 30–39], 1 trial used intermittent RRT [23] and 2 trials used sustained low efficiency RRT provided in sessions of eight hours each day [28, 29]. Only three parallel-group trials, all using CRRT, (n = 24 [24], n = 20 [25], n = 78 [30]) compared exclusively hemofiltration to exclusively hemodialysis at similar small-molecule doses. Of these, one trial allocated patients to higher (2.5 L/h) or lower (1 L/h) dose continuous veno-venous hemofiltration (CVVH) or continuous veno-venous hemodialysis (CVVHD) in a 2 × 2 factorial design [24]. In six of the remaining seven parallel-group trials, patients treated with hemofiltration (n = 20 [21], n = 206 [26], n = 96 [27]) or hemodialysis (n = 13 [22], n = 39 [23], n = 60 [29]) were compared to a group treated with hemodiafiltration (hemofiltration and hemodialysis). Hemodiafiltration recipients received higher doses, except in three trials where doses in the two treatment arms were similar, although not corrected for lower clearance due to pre-filter addition of replacement fluid [23, 27, 29]. The seventh trial, which used sustained low efficiency RRT and was reported only in abstract form, randomized patients into three groups: two similar-dose groups of hemodialysis (n = 6) and hemodiafiltration (n = 5) and a lower-dose group of hemofiltration (n = 6) [28]. Only one parallel-group trial was multi-centered [30].

Table 2 Renal replacement intervention in the included trials
Full size table

All crossover trials were single-centered, used CRRT, and included few randomized patients (median 11.5 patients (range, 5 to 31)) [31–39]. Six of nine crossover trials [31, 33, 34, 36–38] compared hemofiltration and hemodialysis at similar doses. Of these, only three trials reported post-filter addition of replacement fluid [33, 37] or appropriately adjusted the rate of pre-filter replacement fluid to compare similar clearances [38]. Patients in one of these three trials [37] received three treatments (post-filter hemofiltration, pre-filter hemofiltration and hemodialysis) in random order. In another trial [31], patients received hemofiltration and three doses of hemodialysis in random order. For this trial, we excluded the two higher dose hemodialysis groups and retained the two matched-dose hemofiltration and hemodialysis groups.

In the remaining three crossover trials, hemofiltration was compared to a mixture of hemofiltration and hemodialysis at either higher [32, 35] or lower [39] doses. In one of these trials [32], patients were crossed over twice and received their initial treatment a second time; only data from the first two treatments were included.

One parallel-group trial used different fixed doses based on weight ranges [26] while two others used weight-based dosing [27, 30]. The parallel-group trials with sustained low efficiency RRT used fixed doses for the hemodialysis and hemodiafiltration groups [29] and weight-based dosing for the hemofiltration group [28]. One crossover trial used weight-based dose prescriptions for both interventions [38], while another used it only for hemofiltration [39]. The remaining trials used either fixed doses [22–25, 31, 33, 34, 36, 37] or dose ranges not directly related to patient weight [21, 32, 35].

Among the parallel-group trials, RRT was discontinued at the clinicians' discretion [21–23, 25, 27, 29], after a fixed duration [24, 28], or when protocol-defined criteria were met [26, 30]. The mean duration of RRT ranged from 2 to 8 days in four parallel-group CRRT trials [24, 25, 27, 30] and just over 20 days in the parallel-group with intermittent therapy [23]. Among the crossover trials, patients were crossed over after a fixed time: 24 h for most trials [32, 34–37] and 0.5 h for one trial [33]. Two trials crossed patients over after filter clotting or failure [38, 39]. One trial did not report when patients were crossed over [31].

Study quality (Table 3)

In the parallel-group trials that provided these data, all patients were analyzed according to the group to which they were initially assigned, and withdrawal of randomized patients from the mortality analysis either did not occur [26, 27] or comprised ≤5% of randomized patients [21, 23, 30]. Caregiver blinding was not practical in any trial, given the nature of the intervention. Four trials reported [21, 23, 26, 30] concealed allocation. In three trials reporting mortality [24, 25, 27] the authors informed us that they allocated patients to interventions in an alternating manner. The author of one trial informed us that the trial stopped early for benefit [26]. This trial was included only in the sensitivity analyses due to the use of differing doses and a mixture of hemofiltration and hemodialysis. Quality measures shown in Table 3 were generally not reported for the crossover trials, although in two crossover trials, authors reported that 3/13 (23%) [34] and 14/45 (31%) [39] patients were not crossed over to the other treatment. Carryover effects were generally ignored, which may have been reasonable given that the impact of clearance mode on solute removal should manifest relatively quickly [8], and subsequent clearance and concentration measurements were collected over a period of 12 to 24 hours after crossover in most trials. The trial with the shortest measurement period (30 minutes) provided a 10-minute equilibration period after crossover prior to data collection [33]. For one other trial [37], we included data collected over 12 hours but not starting until 12 hours after crossover. Most crossover trials reported using paired analyses [33–37, 39], but none provided individual patient data or mean within-patient difference data. Therefore, we were restricted to using group-specific means to perform meta-analyses, as discussed in the Methods.

Table 3 Risk of bias of included trials
Full size table

Clinical outcomes

The three parallel group trials that compared similar doses of hemofiltration to hemodialysis were included in the primary analysis and reported either ICU [24, 25] or hospital [30] mortality. Five additional parallel-group trials comparing similar [23, 27, 29] or different [21, 26] doses of hemodiafiltration to either hemodialysis or hemofiltration reported mortality data and were included in the sensitivity analysis. They reported hospital [21, 23], 28-day [27, 29], or 90-day [26] mortality. We did not identify differences in pooled mortality in the primary (RR 0.96, 95% CI 0.73 to 1.25, P = 0.76; three trials, 121 patients) or sensitivity (RR 1.10, 95% CI 0.88 to 1.38, P = 0.38; eight trials, 540 patients) analyses (Figure 2). Statistical heterogeneity was absent (I2 = 0%) and moderate (I2 = 50%), respectively. Visual inspection of the funnel plot for the sensitivity analysis did not suggest publication bias.

Figure 2
figure 2

Effect of hemofiltration vs. hemodialysis RRT on mortality. The pooled risk ratio was calculated using a random-effects model. Weight refers to the contribution of each study to the overall estimate of treatment effect. Abbreviations: CI, confidence interval; IV, inverse variance.

Full size image

Dialysis dependence in survivors was not different between groups (primary analysis: RR 1.33, 95% CI 0.35 to 5.08, P = 0.67; two trials [25, 30], 37 surviving patients; sensitivity analysis: RR 0.95, 95% CI 0.44 to 2.04, P = 0.89; four trials [25–27, 30], 177 surviving patients), with no statistical heterogeneity (I2 = 0%). Two trials reported no differences in SOFA scores 72 hours after starting therapy (primary analysis: RoM 1.02, 95% CI 0.89 to 1.16, P = 0.79; one trial, 63 patients [30]; sensitivity analysis: RoM 1.00, 95% CI 0.61 to 1.64, P = 0.99; two trials, 124 patients [27, 30]). One trial reported lower SOFA scores in the hemofiltration group primarily between four and seven days, driven primarily by the cardiovascular component (reflecting lower vasopressor requirements); however, this decrease was not statistically significant [30]. Another trial reported no difference in norepinephrine doses or changes in the multi-organ dysfunction score (MODS) [66] or APACHE II score between continuous hemofiltration and hemodialysis groups over the first 72 hours [24]. One parallel-group trial, comparing intermittent hemodiafiltration to hemodialysis, reported similar improvements in MODS between groups over 10 days [23], and another comparing sustained low efficiency dialysis to diafiltration reported similar improvements in blood pressure between groups over three days [29].

Pooled data from two small crossover trials using similar dose CVVH vs. CVVHD [34, 38] suggest that hemofiltration may shorten the time to filter failure, although only the RoM result achieved statistical significance (MD -7.3 hours, 95% CI -19.4 to +4.9, P = 0.24, I2 = 38%; RoM 0.67, 95% CI 0.45 to 0.99, P = 0.04, I2 = 7%; n = 50). Incorporating the results of one parallel-group trial using similar dose CVVH vs. hemodiafiltration (CVVHDF) [27] produced a pooled result that was significant for both effect measures (MD -5.4 hours, 95% CI -9.6 to -1.3 hours, P = 0.01; RoM 0.70, 95% CI 0.56 to 0.88, P = 0.003; I2 = 0% in both analyses; n = 113). Recognizing that higher dose may also affect filter life, two other trials comparing non-equivalent doses in the two treatment arms demonstrated shorter time to filter clotting in the higher dose group. One crossover trial comparing higher dose CVVH to CVVHDF [39] demonstrated an even greater decrease in filter life in the CVVH group, and one parallel group trial comparing lower dose CVVH to CVVHDF [26] demonstrated a non-statistically significant shorter time to filter clotting in the higher-dose group (Figure 3). Including the data from trials with non-equivalent doses in the two treatment arms in the pooled analysis resulted in a similar shortened time to filter failure in the hemofiltration group (MD -5.6 hours, 95% CI -10.4 to -0.9 hours, P = 0.02; RoM 0.69, 95% CI 0.50 to 0.95, P = 0.02; five trials, 383 patients), with higher heterogeneity (I2 = 51 to 66%). All these trials used unfractionated heparin anti-coagulation and pre-filter replacement fluid except for one trial that used a mixture of pre- and post-filter replacement to keep the filtration fraction <20% [38]. This reduction in filter survival time of about one-third is equivalent to a 50% increase in filters required for hemofiltration compared to hemodialysis.

Figure 3
figure 3

Effect of hemofiltration vs. hemodialysis on filter life. The pooled mean difference was calculated using a random-effects model. Weight refers to the contribution of each study to the overall estimate of treatment effect.

Full size image

Clearances of small and large molecules (Table 4)

Few molecules were examined in more than one study, and analyses included few patients. In general, small molecule clearance (for example, urea, phosphate and creatinine) was similar between hemofiltration and hemodialysis, whereas hemofiltration achieved higher clearance of larger molecules (up to around 20 kiloDaltons (kDa)). Pooled data from two studies [24, 34] showed similar clearance of interleukin (IL)-6 between modes, but statistical heterogeneity was high. Single studies also found that hemofiltration delivered significantly higher clearances of protein and albumin (using a high (60 kDa) cut-off filter) [24], and almost all amino acids in one study examining pediatric patients receiving amino acid supplementation [36].

Table 4 Clearance measurements of hemofiltration vs.
Full size table

The impact of clearance mode on serum concentrations of various solutes of interest was reported even less frequently. One crossover trial [34] found the concentration of tumor necrosis factor (TNF) α (but not IL-6, IL-10, SL-selectin, and endotoxin) to be significantly lower in the patients during hemofiltration. A second crossover trial [35] found no differences in concentrations of mediators of endothelial activation. One parallel-group trial [25] measured a larger decrease in IL-6 and smaller increase in TNFα concentration in the hemofiltration group, but these cytokines were only measured in one patient treated with hemofiltration and two patients treated with hemodialysis. Finally, the crossover trial of pediatric patients receiving amino acid supplementation [36] reported lower serum concentrations of amino acids in association with higher clearances in the hemofiltration group.

Discussion

This systematic review and meta-analysis highlights the paucity of data from randomized controlled trials comparing hemofiltration to hemodialysis in the treatment of AKI. Considering clinical outcomes of hemofiltration in parallel-group RCTs, there was no indication of improved mortality or organ dysfunction, although confidence intervals were wide. Our meta-analysis suggests that hemofiltration shortens filter life by about five to six hours (or one-third of total mean filter time). Based primarily on crossover RCTs, we found that hemofiltration increases the clearance of medium to larger molecules compared to hemodialysis. Almost no studies determined whether the enhanced middle-molecule clearance attributed to hemofiltration actually led to lower serum concentrations.

The trials reporting on filter failure rates used primarily pre-filter addition of replacement fluid. Comparing pre-filter to post-filter addition, one trial [37] included in this meta-analysis and other studies [67–69] have demonstrated that replacement fluid requirements are about 15 to 20% higher to achieve similar small molecule clearance rates at doses of around 20 mL/kg/h. Our finding of a 33% shorter time to filter failure with hemofiltration, combined with its higher fluid requirements to achieve similar small molecule clearance, implies that hemofiltration may consume more resources than hemodialysis. This hypothesis merits further evaluation in a formal cost analysis.

Comparing the published practice survey data in different countries and regions [10–15] (summarized in Figure 4), it appears that physicians tend to use a mode with at least some hemofiltration (that is, either CVVH or CVVHDF), perhaps anticipating additional benefit associated with hemofiltration. However, the small number of trials, randomized patients and events does not support this belief. A well-designed and adequately powered trial would be necessary to establish the superiority of hemofiltration.

Figure 4
figure 4

Distribution of mode of RRT used in different countries/regions based on practice surveys. Abbreviations: CRRT, continuous renal replacement therapy; CVVH, continuous venovenous hemofiltration; CVVHD, continuous venovenous hemodialysis; CVVHDF, continuous venovenous hemodiafiltration; ICU, intensive care unit; MD, medical doctor; NZ, New Zealand; RRT, renal replacement therapy; UK, United Kingdom; US, United States.

Full size image

Strengths of our review include methods to minimize bias, such as a comprehensive literature search, duplicate data abstraction, consideration of important clinical outcomes, and inclusion of additional methodological or clinical information from authors. The primary limitation is the small number and size of RCTs comparing pure hemofiltration to pure hemodialysis at similar doses. Sensitivity analyses, including trials whose arms also varied with respect to dose, gave similar results, as expected given recent large trials [4, 5] and meta-analyses [6, 7] that found similar outcomes with different RRT doses. In addition, trials varied in the modality of RRT used, timing of initiation, and types of filters and blood flows, although recent meta-analyses have not found differential outcomes based on these factors [70–73]. We did not consider comparisons of blood clearance modes with peritoneal dialysis [74, 75], which is used in some areas of the world to treat acute kidney failure. Finally, heterogeneity may have been underestimated because these tests are underpowered when there are few trials. Although hemofiltration is of particular interest in patients with sepsis, in whom pro-inflammatory mediators are increased, there was insufficient data to conduct a subgroup analysis in these patients. In addition, we did not examine the role of hemofiltration vs hemodialysis in patients with sepsis who have not yet developed AKI [40] or the role of hemofiltration compared to no RRT [76–79].

Conclusions

Pooled data from a few randomized trials suggest that hemofiltration increases the clearance of medium to larger molecules without improving clinical outcomes, though confidence intervals are wide. Hemofiltration may also reduce filter life. This latter finding, together with the increased replacement fluid requirements to achieve equivalent small-molecule clearance when pre-filter replacement is used, suggests that hemofiltration may be more expensive than hemodialysis. Our findings support the need for additional pilot data [80] to evaluate the impact of hemofiltration vs. hemodialysis on intermediate outcomes, such as vasopressor requirements, that may serve as valid surrogates for important clinical outcomes that could subsequently be evaluated in a large definitive trial.

Key messages

  • Few randomized controlled trials have compared hemofiltration vs hemodialysis for the treatment of acute kidney injury.

  • Pooling the results from these trials does not suggest beneficial clinical outcomes of hemofiltration vs hemodialysis, but confidence intervals are wide.

  • Compared to hemodialysis, hemofiltration may increase clearance of medium to larger molecules, but may also shorten the time to filter failure.

  • Additional pilot trials are needed to evaluate the impact of hemofiltration vs. hemodialysis on intermediate outcomes, such as vasopressor requirements, that may serve as valid surrogates for important clinical outcomes that could subsequently be evaluated in a large definitive trial.