The effect of early versus late initiation of renal replacement therapy in patients with acute kidney injury: A meta-analysis with trial sequential analysis of randomized controlled trials

Background The optimal timing for initiating renal replacement therapy (RRT) in patients with acute kidney injury (AKI) remains controversial. Methods We conducted a meta-analysis with trial sequential analysis (TSA) of randomized controlled trials (RCTs) using PUBMED, Cochrane Library databases, and Web of Science (from January 1, 1985, to August 21, 2016). Adult patients with AKI who received RRT with different timing were included. The primary outcome was mortality. The secondary outcomes were intensive care unit (ICU) length of stay (LOS) and hospital LOS. Results We included 9 RCTs with a total of 1636 participants. No differences between the early RRT group and the late RRT group were found with respect to mortality (38% vs 41.4%; relative risk, 0.93; 95% confidence interval [CI], 0.74–1.18). However, TSA showed that the cumulative Z-curve did not cross either the conventional boundary for benefit or the trial sequential monitoring boundary, indicating insufficient evidence. Similarity, there were no findings of benefits in terms of reduction in the ICU LOS (standard difference in the means, −0.32 days; 95% CI, −0.71 to 0.07 days) and hospital LOS (standard difference in the means, −1.11 days; 95% CI, −2.28 to 0.06 days). Meanwhile, the results of TSA did not confirm this conclusion. Conclusions Although conventional meta-analysis showed that early initiation of RRT in patients with AKI was not associated with decreased mortality, ICU LOS and hospital LOS, TSA indicated that the data were far too sparse to make any conclusions. Therefore, well-designed, large RCTs are needed.


Results
We included 9 RCTs with a total of 1636 participants. No differences between the early RRT group and the late RRT group were found with respect to mortality (38% vs 41.4%; relative risk, 0.93; 95% confidence interval [CI], 0.74-1.18). However, TSA showed that the cumulative Z-curve did not cross either the conventional boundary for benefit or the trial sequential monitoring boundary, indicating insufficient evidence. Similarity, there were no findings of benefits in terms of reduction in the ICU LOS (standard difference in the means, −0.32 days; 95% CI, −0.71 to 0.07 days) and hospital LOS (standard difference in the means, −1.11 days; 95% CI, −2.28 to 0.06 days). Meanwhile, the results of TSA did not confirm this conclusion. PLOS  The present meta-analysis was performed according to the recommendations of the Cochrane Handbook for Systematic Reviews of Interventions [21] and was also done in compliance with PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement) guidelines [22] (S2 Table). This systematic review was not registered, and a protocol does not exist. The search was limited to human subjects, and no language restrictions were applied.
The inclusion criteria were as follows: (1) study design: RCTs; (2) comparison: the effect of early versus late initiation of RRT in patients with AKI; and (3) population: critically ill adult patients (>18 years old). Exclusion criteria were as follows: (1) studies that included patients with preexisting chronic kidney disease or previous RRT; (2) data from the published results that could not be extracted and analyzed; and (3) studies that included pregnant patients.
Reference lists from the identified trials and review articles were then manually scanned to identify any additional relevant studies. The updated literature search, data extraction, and quality assessment were done independently by two authors (YMF and RG) using a standardized approach.

Data extraction and quality assessment
Two investigators (YMF and RG) independently performed the study selection. Disagreements between the two investigators were resolved by a third party (YY). A standard form was used to collect data from each study. The form included first author, year of publication, study design, country, patient type, total number of patients, early RRT criteria and primary outcome. The primary outcome was mortality at latest follow-up, including mortality at 28, 60, and 90 days in accordance with the results from the primary authors. Secondary outcomes included intensive care unit (ICU) length of stay (LOS) and hospital LOS. Two investigators (YMF and RG) extracted the data independently.
The definition used for early initiation of RRT was consistent with the criteria used by the original authors in their respective studies. Early RRT was defined based on biochemical markers from the RIFLE classifications (risk, injury, failure, loss of function, and end-stage kidney disease), the Acute Kidney Injury Network (AKIN) stages, or time-based cutoffs (e.g., within a defined time from ICU admission or development of a biochemical "start time").
The quality of the included studies was assessed with standard criteria: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting and other biases. When data were missing or incomplete, the original authors were contacted by written correspondence for clarification, and any relevant information obtained was included in the review.

Grading the quality of the evidence
The quality of evidence for primary and secondary outcomes of this meta-analysis was evaluated independently by two reviewers (FZ and RG) according to the GRADE methodology for risk of bias, inconsistency, indirectness, imprecision, and publication bias; and classified as very low, low, moderate, or high. Summary tables were constructed using the GRADE Profiler (GRADEpro, version 3.6.1) [23].

Statistical analysis
We calculated relative risks (RRs) with 95% confidence intervals (CIs) for dichotomous outcomes and mean differences with 95% CIs for continuous outcomes. Heterogeneity across studies was quantified using the I 2 statistic; 25% I 2 <50% indicated low heterogeneity, 50% I 2 <75% indicated moderate heterogeneity, and I 2 !75% indicated high heterogeneity. A random-effect model was used to analyze the results of trials with significant heterogeneity, and a sensitivity analysis was performed to test the robustness of results. Subgroup analysis was conducted to investigate potential sources of between-study heterogeneity. Publication bias was assessed using the Begg and Egger tests. A P value less than 0.05 was considered to indicate a statistically significant difference. All statistical analyses were performed using commercially available software, including Stata version 13.0 (StataCorp LP), TSA version 0.9 (Copenhagen Trial Unit) and RevMan 5.3 (Cochran IMS).

Trial sequential analysis
Random errors are widely distributed in individual studies, such that repeated tests for significance using traditional meta-analysis may increase the risk of type I errors when sparse and accumulated data are analyzed. Therefore, it is wise to accept a tentative conclusion based on current trials, because it may be reversed by type I errors when subsequent trials are taken into account. TSA is a method that solves this important issue by adapting the monitoring boundaries to evaluate the accumulated evidence [24] and calculating the estimated information size to guide protocols in subsequent trials [25,26]. Basically, TSA recognizes a conclusion from conventional meta-analysis with 1 of the following 4 outcomes: (1) If the cumulative Z-curve exceeds the estimated information size and crosses the trial sequential monitoring boundary, the conclusion is sufficient and no more trials are needed; (2) if the cumulative Z-curve crosses the traditional monitoring boundary but does not cross the trial sequential monitoring boundary or exceed the estimated information size, the current conclusion may be a false-positive result and more trials should be included to clarify this issue; (3) if the cumulative Z-curve does not cross the traditional monitoring boundary and the trial sequential monitoring boundary, the conclusion may be a false-negative finding and more trials should be included to prove this issue; or (4) if the cumulative Z-curve exceeds the estimated information size but does not cross the traditional monitoring boundary, the negative conclusion is sufficient and there are no significant differences between the intervention group and the control group. In this study, we calculated the diversity-adjusted estimated information size using α = 0.05 (twosided) and β = 0.20 (power 80%), with an anticipated 18% reduction in RR of mortality and an anticipated estimated -1.4 days mean difference in ICU LOS and -6.44 days estimated mean difference in hospital LOS.

Study selection
The PRISMA statement flowchart shown in Fig 1 details the process of literature screening, study selection, and reasons for exclusion. The systematic search identified 2347 relevant references. After screening titles and abstracts, we excluded 2251 articles due to irrelevance to the topic or because they did not meet the inclusion criteria. Full texts of the 96 remaining articles were retrieved for formal review. After assessment of the full texts, 88 articles were excluded. After we reviewed the original articles and additional references, 9 RCTs [14][15][16][17][18][27][28][29][30] were included to compare the effect of early versus late RRT in patients with AKI, resulting in a total of 1636 patients in both trial arms (early RRT, n = 827; late RRT, n = 809). The two authors had no disagreements regarding study selection.

Quality of studies
Details of the tool used to assess the risk of bias are shown in Fig 2A. Owing to the nature of the interventions, it was impossible for the medical staff to perform the studies blinded. Therefore, the GRADE Working Group grades of evidence were low for the primary outcome, and very low for the secondary outcomes of ICU LOS and hospital LOS. This was mainly due to the risk of bias and the moderate to high heterogeneity within studies. Full GRADE profiles for the included evidence can be found in S3  Fig 2A). To evaluate the contribution of each individual study to the moderate heterogeneity, sensitivity analysis with consecutive exclusion of one trial each time was performed. However, the meta-analyses performed after the exclusion of individual trials showed no significant effect on the pooled estimate and 95% CI ( Fig 2B). Meanwhile, because of the moderate heterogeneity, the random-effect model of the DerSimonian-Laird (DL) and Sidik-Jonkman (SJ) methods was used for TSA. The results showed that the cumulative Z-curve crossed neither the traditional boundary nor the trial sequential monitoring boundary ( Fig  2C), indicating insufficient power to draw a definitive conclusion.

Secondary outcomes
The secondary outcomes were ICU LOS and hospital LOS. Six of nine RCTs reported ICU LOS and hospital LOS [14,[16][17][18]27,30]. Only one of these studies reported a significant decrease in ICU LOS and hospital LOS following early RRT [27]. Meta-analysis showed that there was no significant decrease in ICU LOS in patients receiving early RRT, with a standard difference in the means of −0.32 (95% CI −0.71 to 0.07, p<0.00001) using a random-effects model and high heterogeneity (tau 2 = 0.21, chi 2 = 51.72, df = 5, p<0.00001, I 2 = 90%; Fig 3A).
To evaluate the contribution of each study to the high heterogeneity, sensitivity analysis with Where is the truth?
consecutive exclusion of single trials was performed. The results of the study conducted by Gaudry et al. [17] were completely out of range of the other results and probably contributed Single trial was excluded each time, however, pooled estimate and 95% CI had no significant changes. C. TSA on mortality in patients with AKI receiving early versus late initiating of RRT, which showed that the cumulative Z-curve did not cross either the conventional boundary for benefit or the trial sequential monitoring boundary for benefit. Therefore, it established insufficient and inconclusive evidence. The estimated required information size of 5185 patients was calculated using α = 0.05 (two-sided) and, β = 0.20 (power 80%), an anticipated relative risk reduction of 18%, and an event proportion of 41.4% in the late RRT group. https://doi.org/10.1371/journal.pone.0174158.g002 Where is the truth?
to the heterogeneity, indicating a lack of reliability in our conclusions (Fig 3B). For TSA using α = 0.05 (two sided) and β = 80% with an estimated mean difference = −1.4, variance = 32.95, and heterogeneity correction = 92%, the Z-curve crossed the conventional boundary but did not reach the trial sequential monitoring boundary or TSA information size (Fig 3C). Forest plots showed that there was no significant decrease in hospital LOS in patients receiving early RRT, with a standard difference in the means of -1.11 (95% CI, −2.28 to 0.06, p<0.00001) using a random-effects model, as well as high heterogeneity (tau 2 = 2.08, chi 2 = 372.25, df = 5, I 2 = 99%; Fig 4A). To evaluate the contribution of individual studies to the high heterogeneity, sensitivity analysis with consecutive exclusion of individual trials was performed. The results of the studies conducted by Combes et al [16] and Zarbock et al [18] were completely out of range of the other results and probably contributed to the heterogeneity, indicating a lack of reliability in our conclusions (Fig 4B). For TSA using α = 0.05 (two-sided) Sensitivity analysis of secondary outcome of ICU LOS. C. TSA on ICU LOS in patients with AKI receiving early versus late initiating of RRT, which showed that the cumulative Z-curve crossed the conventional boundary for benefit but did not cross the trial sequential monitoring boundary for benefit. Therefore, it established insufficient and inconclusive evidence. The estimated required information size of 6596 patients was calculated using α = 0.05 (two-sided) and β = 0.20 (power 80%), an anticipated estimated mean difference reduction of −1.4, and a heterogeneity correction of 92% in the late RRT group. and β = 80% with estimated mean difference = −6.44, variance = 151.87, and heterogeneity correction = 96%, the Z-curve crossed the conventional boundary but did not reach the trial sequential monitoring boundary and TSA information size (Fig 4C).

Publication bias
We assessed potential publication bias for the primary outcome of mortality (P = 0.52 for the Begg test, P = 0.473 for the Egger test) and the secondary outcomes of ICU LOS (P = 0.75 for the Begg test, P = 0.063 for the Egger test) and hospital LOS (P = 0.754 for the Begg test, P = 0.547 for the Egger test). No potential publication bias was observed among the included trials (see S1 Fig). Sensitivity analysis of secondary outcome of hospital LOS. C. TSA on hospital LOS in patients with AKI receiving early versus late initiating of RRT, which showed that the cumulative Z-curve crossed the conventional boundary for benefit but did not cross the trial sequential monitoring boundary for benefit. Therefore, it established insufficient and inconclusive evidence. The estimated required information size of 2874 patients was calculated using α = 0.05 (two-sided) and β = 0.20 (power 80%), an anticipated estimated mean difference reduction of −6.44, and a heterogeneity correction of 96% in the late RRT group. https://doi.org/10.1371/journal.pone.0174158.g004

Discussion
RRT represents a cornerstone in the management of life-threatening AKI. Several aspects of RRT are now well established, but others remain controversial, especially the optimal timing to initiate RRT. To date, there has been clear consensus that timely RRT is required in lifethreatening conditions, such as severe hyperkalemia, marked metabolic acidosis, and/or fluid overload; however, in most critically ill patients with AKI, the decision to initiate RRT is not done mainly on account of overtly life-threatening conditions [31]. Consequently, there are large variations in the timing of RRT initiation in these populations, influencing clinical outcomes. Our meta-analysis, based on the traditional method, provided two conclusions. First, early RRT was not associated with improved clinical outcomes (i.e., reductions in mortality), although this finding should be interpreted with caution, because TSA suggested this evidence may represent a false-negative result. Second, meanwhile, there were no findings to support the idea that early RRT will shorten the ICU LOS and hospital LOS. TSA indicated that the evidence was insufficient and inconclusive.

Relation to other reviews and implication for future research
Because high-quality RCTs are limited, much of the current evidence is derived from observational studies or meta-analyses [12,13,32]. Liu [33] included 9 retrospective cohort studies and 2 RCTs (841 patients). This meta-analysis showed that early initiation of RRT in patients with AKI after cardiac surgery resulted in lower mortality at 28 days (OR = 0.29, 95% CI, 0.16-0.52, P<0.0001, I 2 = 56%) and shorter ICU LOS (3.9 [1.5-6.3] days, P<0.0001, I 2 = 99%). Although the majority of these clinical data appear to support the view that early RRT reduces mortality in patients with AKI, these observational data may suffer from a number of potential methodological limitations, leading to biased results [33].
In addition, our meta-analysis included 9 RCTs with a total of 1636 participants. Hence, we had more sample size as well as updated trials and a higher level of evidence coming from RCTs. Thus, evidence based on our conclusions should have more strength to demonstrate this issue. More importantly, we further performed TSA to assess the robustness of these outcomes which indicated that there was no conclusive evidence to support or oppose the early initiation of RRT in patients with AKI. It is, therefore, encouraging high-quality RCTs to find the truth.
Wierstra et al [34] synthesized the largest pool of clinical data to date (7 RCTs, 10 prospective cohorts, and 19 retrospective cohorts) and concluded that early initiation of RRT in critical illness complicated by AKI did not improve patient survival or reduce ICU or hospital LOS. Although our findings also did not observe significant benefits associated with early initiation of RRT in patients with AKI, our conclusions were based on RCTs only. Because observational studies make no intervention and patients are allocated treatment based on clinical decisions, selection bias may exist, thereby overestimating the benefits of early RRT. Our evidence minimizes this source of bias and may thus be considered more reliable.
In addition, we included 2 more recently published high-quality clinical trials (the ELAIN and AKIKI trials) than Wierstra et al [34]. Although the number of RCTs increased by only 2 clinical trials, the total number of patients involved doubled, from 786 to 1636. Therefore, our analysis may provide stronger conclusions on this issue.
Although a recent meta-analysis of RCTs [35] included the ELAIN and AKIKI trials (6 RCTs with 1257 patients) and made conclusions similar to ours, their search strategy was incomplete. Also, our meta-analysis included 3 additional trials, for a total of 1636 participants (9 RCTs). Furthermore, because the available RCTs had small sample sizes (only 1 trial [17] included more than 300 patients) and the number of RCTs was insufficient (fewer than 10 RCTs), there was no additional ability to determine the risk of random errors due to repeated testing of the accumulated data in the current conclusive meta-analysis. To our knowledge, this is the first study to investigate this effect using TSA method. Although our meta-analysis involved more data and thus should have conclusions with greater power, TSA showed that the optimal timing of initiating RRT remains unclear. More RCTs should be performed to further clarify this issue.
Fortunately, there are two ongoing RCTs that should provide additional information regarding the optimal timing of RRT in critically ill patients with AKI, with special interest in patients with septic shock (IDEAL-ICU) or admitted to general ICU (STARRT-AKI). As well, these clinical trials have enrolled more patients. In addition to short-term follow-up, IDEA-L-ICU further investigates the impact of the two RRT strategies on mortality at 180 and 360 days. Meanwhile, more diverse outcomes are being measured, including cost-effectiveness and new inclusion criteria. Accordingly, these efforts may give physicians better approaches to real-world RRT in patients with AKI.
On the other hand, another implication from our meta-analysis should be noted for planned and ongoing clinical trials. Because there were large variations in the definition of "early" and "late" RRT among physicians and countries (for example, the "late" group in the ELAIN trial was actually earlier than the "early" group in AKIKI), moderate heterogeneity regarding the possible association between timing and mortality was found in our meta analysis. Most trials included in this analysis [14][15][16][17][18]30] used a fixed time point for the timing of RRT initiation with a given anticipated course of AKI. However, the precise time of AKI is elusive and a variety anticipated course of AKI. However, the precise timing of AKI is elusive. A variety of surrogate biomarkers have been used to better define this issue; neutrophil gelatinase-associated lipocalin is the most frequently investigated, and the data are promising [36,37]. However, it is important to remember that a clinical decision to start RRT is inappropriate when it does not take into account a patient's individual condition, specific course of illness, and the different RRT modalities and dosage options that are available.

Strengths and limitations
A major strength of this meta-analysis is the compliance with the PRISMA guidelines and the recommendations of the Cochrane Collaboration, although the protocol of our study was not registered in the international prospective register of systematic reviews (PROSPERO). To increase the robustness of our meta-analysis, we applied TSA to assess the impact of random error and repetitive testing. Finally, we evaluated the quality of evidence for the outcomes using GRADE to help healthcare professionals make better clinical decisions.
Our meta-analysis also has some limitations. The included trials in our meta-analysis were conducted on varying numbers and types of patients, had different designs, and used different criteria to determine early versus latte RRT, the method of RRT treatment, and duration of mortality follow-up. Thus, the risk of introducing potentially significant heterogeneity is imminent. In addition, double-blinding was not performed because of the features of the trials, which may have contributed to performance and detection bias.

Conclusions
Conventional meta-analyses that included recent trial data showed that early initiation of RRT in patients with AKI was not associated with decreased mortality, ICU LOS, or hospital LOS. After TSA adjustment for sparse data and multiple update in the cumulative meta-analysis, we were unable to draw definitive conclusions regarding the ideal timing of RRT in patients with AKI. The results of ongoing and future well-designed, large RCTs are needed to clarify this issue.
Supporting information S1 Fig. A. Publication bias for the primary outcome of mortality using Begg's test. B. Publication bias for the primary outcome of mortality using Egger's test. C. Publication bias for the secondary outcome of ICU LOS using Begg's test. D. Publication bias for the secondary outcome of ICU LOS using Egger's test. E. Publication bias for the secondary outcome of hospital LOS using Begg's test. F. Publication bias for the secondary outcome of hospital LOS using Egger's test. (TIF) S1  Table. The GRADE Evidence Profile for the Primary and Secondary outcomes of this meta-analysis. GRADE Working Group grades of evidence High quality: Further research is very unlikely to change our confidence in the estimate of effect. Moderate quality: Further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate. Low quality: Further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate. Very low quality: We are very uncertain about the estimate. 1 Although most of included RCTs were judged as high risk of performance bias because of without blinding of participants and personnel, the predefined objective outcome was just partly influenced. 2 Heterogeneity (I 2 = 57%) as found. 3 Heterogeneity (I 2 = 90%) was found. 4 Heterogeneity (I 2 = 99%) was found. (TIF)