Summary of main results

Our systematic review and subsequent meta‐analysis examined the effect of different timing of initiation of KRT on death, kidney recovery function, length of stay, and adverse events among 4880 randomised critically ill patients with AKI. Most of the included studies were assessed as having a low or unclear risk of bias for all domains. Two studies were assessed as having a high risk bias, one for incomplete outcome data (attrition bias) and the other for selective reporting (reporting bias).

Within the time of KRT initiation assessed, earlier start may have no beneficial effect on death or recovery of kidney function (in all patients) compared to standard strategy.

Within the time of KRT initiation assessed, earlier start may have little to no difference on death at day 30. The overall estimated effects on risk of death showed clinically small benefits (decreased death by 3%), but the CIs were sufficiently wide to include benefits and harm (imprecision), with a low level of heterogeneity (I² = 29%; inconsistency). The 3% relative risk reduction (RRR) in death at day 30 in the early KRT group is related to a reduction in absolute risk observed in only 12 of 1000 patients (50 less or 35 more than those treated with late KRT), thus assuming little to no effect on death.

Early strategy probably makes little to no difference on death after day 30, with imprecision and without inconsistency (I² = 6%).

Early strategy may make little or no difference to the risk of death or non‐recovery of kidney function at day 90. The overall estimated effects on risk of death showed clinical benefits (decreased death by 9%), but the CIs were sufficiently wide to include benefits and harm (imprecision), with a moderate level of heterogeneity (I² = 66%; inconsistency). However, when we removed Sugahara 2004, the I² is reduced to 46%, and the imprecision was also reduced. There are no significant differences between the groups (subgroup test P = 0.12, I² = 43%). The RR went from 0.91 (95% CI 0.74 to 1.11) to 0.96 (95% CI 0.83 to 1.10), which is compatible with little to no difference in death or non‐recovery kidney function at day 90. This study was assessed as having a high risk of bias by incomplete outcome data (attrition bias) (See Sensitivity analysis and Overall completeness and applicability of evidence).

Early start may make little or no difference to the number of patients who recovered kidney function. CIs included damage (imprecision), with a moderate level of heterogeneity (I² = 55%; inconsistency). There was little to no difference in kidney recovery among survivors between interventions. However, reporting kidney recovery among survivors alone does not preserve the previously achieved randomisation. Therefore, the interpretation of this result may be misleading, given death is a competing endpoint for recovery of kidney function in patients with a high short‐term risk of death (indirectness). However, when we removed three studies (EARLYRRT 2018; Sugahara 2004; Xia 2019), the I² was reduced to 25%. The RR went from 1.07 (95% CI 0.94 to 1.22) to RR 1.00 (95% CI 0.92 to 1.09), which is compatible with little to no difference in the recovery of kidney function. These studies were sources of heterogeneity probably due to selection bias (Sugahara 2004; Xia 2019), attrition bias (Sugahara 2004) and no blinding (EARLYRRT 2018; Sugahara 2004; Xia 2019), thus limiting the internal validity. Xia 2019 and EARLYRRT 2018 used AKI‐biomarker (high level of urinary or serum NGAL) as criteria for early KRT initiation.

It is uncertain whether early KRT initiation increases or reduces the number of patients who experienced any adverse events compared to standard, with a substantial level of heterogeneity (I² = 91%; inconsistency). Nevertheless, the early strategy did increase the risk of hypophosphataemia, hypotension, cardiac rhythm disorder and infections, although it had uncertain effects on thrombocytopenia and the risk of bleeding when compared to standard initiation.

Early start probably reduces the length of ICU and hospital stay (number of days). The magnitude of the possible benefit was clinically relevant (‐1.01 days to ‐2.45 days, respectively). These results should be interpreted with caution owing to the indirectness observed (in this population, death is a competing endpoint for the length of stay).

With a focus on the effect size of the central estimation (magnitude or importance), we observed that early initiation may make little to no difference to death, may improve the recovery of kidney function, probably reduces the length of ICU and hospital stay, while it increased the risk of adverse events. However, all results (except any adverse events and length of stay) were imprecise because the CIs crossed both the important effect threshold and the no difference threshold.

An important limitation of this systematic review was the low to moderate heterogeneity found in the main results, as death at day 30 (I² = 29%), death or non‐recovery of kidney function at 90 days (I² = 66%), and on recovery of kidney function in all patients (I² = 55%). There was no heterogeneity identified for the length of stay, and adverse events (hypophosphataemia, hypotension, cardiac rhythm disorder and infections), except for the number of patients with any adverse event (I² = 91%).

We explored this heterogeneity by prespecified subgroup analyses: aetiology of AKI, according to criteria used to define the timing of KRT initiation, modalities of KRT, and the severity of illness at baseline. The subgroup modality of KRT initiation was identified as a source of heterogeneity in the size of the effect observed in the recovery of kidney function (test for subgroup differences: Chi² = 4.27; P = 0.04; I² = 76.6%). These results should be interpreted with caution as only five small studies contributed to these data. Notably, several studies reported that there were more hypotension events with intermittent haemodialysis, which was more likely to result in haemodynamic instability than CKRT, with a lower likelihood of kidney recovery after AKI.

In the subgroup of aetiology of AKI, we observed a reduction in death (35%) in patients with surgery‐acquired compared to those patients with non‐surgery‐acquired AKI (increased risk 1%). Despite some heterogeneity (I² = 28.3%) between groups, the test for subgroup difference was not statistically significant. This could be explained by the studies being underpowered to detect differences due to the small sample size of the studies with the surgical‐AKI group. However, if we remove Sugahara 2004, the I² is reduced to 13%, and the imprecision is also reduced. The RR goes from 0.65 (95% CI 0.31 to 1.36) to 0.84 (95% CI 0.59 to 1.20). The effect size is lower but still clinically relevant. This study was assessed as having a high risk of bias by incomplete outcome data (attrition bias) (See Overall completeness and applicability of evidence).

In the subgroup of KRT modalities reduction in death (14%) was observed in patients with CKRT compared to those patients with mixed KRT modality (increased risk 2%). Without heterogeneity (I² = 0%) between groups, the test for subgroup difference was not statistically significant. However, when we removed Sugahara 2004, the I² is reduced from 48% to 7%, and the imprecision was also reduced. The RR goes from 0.86 (95% CI 0.65 to 1.14) to 0.93 (95% CI 0.77 to 1.13) (See Overall completeness and applicability of evidence).

In the subgroup of aetiology of AKI, we observed an increase in kidney recovery rate (36%) in patients with surgery‐acquired compared to those patients with non‐surgery‐acquired AKI. Without (I² = 9.4%) between groups, the test for subgroup difference was not statistically significant. This could be explained by the underpowered to detect differences due to the small sample size of the studies with the surgical‐AKI group. However, if we remove Sugahara 2004, the I² is reduced to 0%, and the imprecision was also reduced. The RR goes from 1.36 (95% CI 0.78 to 2.38) to 1.12 (95% CI 0.72 to 1.74). The effect size is lower but still clinically relevant. This study was assessed as having a high risk of bias by incomplete outcome data (attrition bias) (See Overall completeness and applicability of evidence).

For the death or non‐recovery of kidney function at 90 days, the subgroups aetiology (surgical and non‐surgical), initiation criteria KDIGO 2, KDIGO 3, AKI RIFLE‐F stage, and AKIN stage 3, or other criteria) and modality (CKRT or mixed KRT) made little or no difference to this outcome. Early initiation may reduce the risk of death or non‐recovery of kidney function at 90 days in patients with a SOFA score > 12 but not in those with a SOFA score ≤ 12.

RCTs focusing on the timing of KRT initiation for paediatric AKI patients were not available.

Overall completeness and applicability of evidence

Although the analyses included data obtained from a comprehensive and rigorous search, we identified gaps in several areas. The majority of participants in the included studies were adults, limiting the applicability of our finding to children. In general, the incidence of AKI secondary to sepsis in ICU is high; however, in three studies, it was observed that the majority of patients had post‐surgical AKI, and relatively few had sepsis or pre‐existing chronic kidney disease (CKD), limiting the applicability of our results to general ICU population.

Six studies were single‐centre, and all were unblinded, limiting the external and internal validity of the results, respectively.

Data on the number of patients with any adverse events were limited and only provided by five of the 12 studies in our review.

Few studies reported data for KRT dosage and volume overload; we are aware that it is an important issue to consider in relation to death in critically ill patients with AKI.

Most of the studies did not report data on death in patients with pre‐existing CKD.

There were large variations in the definition of the timing of KRT initiation among included studies. Heterogeneous indicators such as different serum urea or SCr levels, urine output, time from randomisation and time to fulfil KDIGO AKI stage, biochemical markers and furosemide test are widely used to measure the timing of KRT; however, this approach provides an incomplete assessment of optimal timing of KRT initiation in this population and limits the applicability of our results.

It is important to highlight the absence of data related to the characteristics and evolution of patients randomised to the standard or late arm who did not receive dialysis treatment. These data would allow us to develop a propensity‐based analysis of patients in the accelerated group and among those who did not receive KRT in the standard/delayed strategy in order to define where these patients could have had a better outcome.

We were unable to address all of the objectives of this review due to the lack of data in the included studies. Also, we did not have individual patient data for the different subgroups of the modality of KRT and aetiology of AKI, being a limitation of our review

The RCTs included as well as recent research by Gaudry 2020, provided new knowledge and tools, such as the use of furosemide stress test or emerging biomarkers of persistent severe AKI and clinical judgment, that will help us define the optimal KRT initiation time in order to recognize when early KRT initiation may be essential for better outcomes or unnecessary due to potential harms for AKI‐patients in ICU.

We included only RCTs with the purpose of reducing bias.

Quality of the evidence

We conducted this review according to the process described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2021). Our review was based on evidence from 12 RCTs (4880 randomised participants) that compared different timing of KRT initiation in critically ill patients with AKI. The certainty evidence for our main outcomes was drawn from studies assessed at low risk of bias for random sequence generation and allocation concealment processes, incomplete outcomes data, intention to treat analysis, selective outcomes reporting, performance and detection bias and other sources of bias; as well as unclear risk for detection bias. Two studies were at high risk of bias by incomplete outcome data (attrition bias) and selective reporting (reporting bias). Three small studies had an unclear risk of selection bias.

Data comparing the effect of early KRT initiation against standard initiation on death at day 30 or after were obtained from 12 and seven well‐conducted RCTs, respectively, but we downgraded the certainty of evidence to low, mainly due to inconsistency (I² = 29%) and imprecision (CIs included a range of plausible value with clinically important benefits, but also harm), and rated it as moderate by imprecision for death after 30 days. Similarly, we downgraded the certainty of evidence to low for recovery of kidney function in all patients due to imprecision and inconsistency (I² = 55%) and rated as low data obtained for recovery of kidney function among survivors by inconsistency (I² = 69%) and indirectness (the recovery of kidney function in this high‐risk group is affected when the risk of death is taken into account).

Data used to assess the impact of early versus standard initiation of KRT on adverse events were obtained from eight well‐conducted RCTs, providing treatment effects with clinically important harms for hypophosphataemia, hypotension, cardiac‐rhythm disorder and infections. We rated this as high‐certainty evidence. Six studies reported the number of patients with bleeding; and were rated as moderate by imprecision. One study provided data on the number of patients with thrombocytopenia; we downgraded the certainty of evidence to very low due to serious imprecision and study limitation (one study with a small sample size). In the same way, we downgraded the certainty of evidence as very low owing to imprecision and substantial inconsistency (I ²= 91%) observed in the number of patients with some kind of adverse event (data provided by five RCTs).

Length of ICU and hospital stay was reported by five and seven RCTs, respectively; we downgraded the certainty of evidence to moderate due to indirectness, as death is a competing endpoint for the length of stay in this population.

Potential biases in the review process

While this review was conducted according to rigorous methods developed by the Cochrane Collaboration, some bias may be present in the review process. We searched for all relevant studies using sensitive and validated strategies in major medical databases and grey literature sources. However, it is possible that some studies (such as unpublished data and studies with negative or no effects) were not identified. An analysis of evidence to assess the risk of publication bias was not possible for all outcomes due to the number of studies available in each meta‐analysis (Figure 4).

Several subgroup analyses were planned to explore potential sources of heterogeneity in our review; however, a lack of data prevented us from performing these analyses.

Agreements and disagreements with other studies or reviews

Our systematic review, in keeping with previous meta‐analyses on timing in KRT (Gaudry 2020; Li 2021; Naorungroj 2021; Pan 2021), found that earlier KRT initiation may have no beneficial effect on death in critically ill patients with AKI compared with later strategy. These results were not consistent with two other systematic reviews that included randomised and observational studies (Seabra 2008; Wierstra 2016) and other meta‐analyses based only on RCTs (Mavrakanas 2017; Wang 2017; Xu 2017)

The hypothesis that critical AKI patients, especially those with acidaemia, fluid overload, or systemic inflammation, could benefit from early KRT was proposed by several researchers. Our review has found that early strategy may have little to no difference on death at day 30. This result is consistent with five multicentre RCTs (AKIKI 2015; Bouman 2002; IDEAL‐ICU 2014; STARRT‐AKI Pilot 2013; STARRT‐AKI 2019) but does not agree with those reported in three individual RCTs (ELAIN 2016; Tang 2016; Sugahara 2004)

It is important to note that differences in death between AKIKI 2015 and ELAIN 2016 were observed (41.6% versus 30.4% at day 30, respectively). These differences may be due to several factors, which include: different severity levels and aetiology of AKI, e.g. prevalence of patients with AKI related to surgical cause in the ELAIN 2016 or septic AKI‐patients was more frequent in AKIKI 2015; both aetiologies have different pathophysiology and prognosis), and variable criteria for defining early KRT initiation (KDIGO AKI stage 3 for AKIKI 2015 and KDIGO AKI stage 2 for ELAIN 2016).

Other timing criteria were observed: serum and urinary biomarkers (EARLYRRT 2018; Xia 2019), or furosemide test (FST 2018), and the equipoise judgment of clinicians for inclusion in the standard arm (STARRT‐AKI Pilot 2013; STARRT‐AKI 2019) (See Overall completeness and applicability of evidence).

There has been increased interest in the recovery of kidney function. Indeed, lack of recovery of kidney function implies the need for long‐term dialysis associated with low quality of life and high health costs. Our review has found that early strategy may have a slightly beneficial effect on the recovery of kidney function in all patients. This finding is consistent with two individual RCTs (ELAIN 2016; Sugahara 2004) (with high kidney recovery rate), and does not agree with the other three multicentre RCTs (AKIKI 2015; Bouman 2002; STARRT‐AKI Pilot 2013). Differences in the recovery of kidney function between studies may be due to the same factors mentioned above. However, in patients with a high short‐term death risk, the interpretation of this result may be misleading, given that death is a competing endpoint for recovery of kidney function (Palevsky 2005).

Patients with AKI experience longer ICU and hospital stays. In our review, the earlier strategy probably reduce ICU and hospital length of stay; this result is consistent with individual RCTs and meta‐analyses (ELAIN 2016; STARRT‐AKI Pilot 2013; Naorungroj 2021) and does not agree with other RCT reports (AKIKI 2015; IDEAL‐ICU 2014) and meta‐analyses (Gaudry 2020; Li 2021). However, the length of stay in this high‐risk population may be affected when death is taken into account.

There was an increased risk in the number of patients who had specific adverse events with early initiation of KRT compared with standard. Our results were consistent with other RCTs (Bouman 2002; STARRT‐AKI 2019) and meta‐analyses (Li 2021; Naorungroj 2021).

Our review has an important limitation due to the heterogeneity observed in the main outcomes. Only in kidney recovery did we find an association between the estimated effect and KRT modality in agreement with a recent meta‐analysis (Pan 2021). We were unable to address all of the pre‐specified subgroup analyses of this review due to the lack of data in the included studies.

Our review includes studies of different countries (Europe, North America and Asia) which increase the applicability of these results.

Previous reviews explored the effect of time to KRT initiation in patients with AKI; however, these reviews included studies that we excluded from our review due to the following factors: different inclusion criteria applied, e.g. hospitalised patients were not in an ICU setting (Pursnani 1997) or did not require AKI for enrolment in the early arm (Durmaz 2003; HEROICS 2015; Jamale 2013; Koo 2006) and differences in the methodological studies design (cohort studies). Although the abundance of cohort studies provided more power (increases the sample size) to find significant clinical differences between both treatments, these studies have important limitations: patients between intervention groups were different (e.g. patients assigned to late arm treatment might have died before initiating the therapy, while others who lived enough to be assigned to the late group might have been less sick or with a high likelihood of recovering kidney function without KRT). A relevant point worth considering is that patients do not have the same opportunity to receive early or standard treatment (allocation or selection bias). Consequently, to minimise the risk of bias in our review, we included only RCTs for our main outcomes.