MO363: Target Trial Emulation on Timing of Start of Renal Replacement Therapy in Acute Kidney Injury: Lessons Learned

Abstract

Abstract BACKGROUND AND AIMS Routinely collected data (RCD) are readily available at most ICUs. Their secondary analysis is expected to improve medical decision-making by providing expedited answers to clinical questions that would otherwise require conducting costly and less generalizable RCTs. However, doubts remain whether RCD can produce sound answers to comparative effectiveness questions. Trial emulation studies based on RCD closely mimic trial protocols, limit the risk of avoidable biases and could therefore be a powerful tool to expand our evidence base on patient management. We assess whether such an emulation of the ELAIN trial[1] can replicate results of this landmark study comparing timing strategies for initiating renal replacement therapy (RRT) in ICU patients with acute kidney injury (AKI). METHOD Using serum creatinine and urine output measures, we identified patients admitted to the Ghent University Hospital ICUs (2013–2017) who developed KDIGO stage 2 AKI. Other eligibility criteria from the ELAIN trial available from our linked databases were applied to obtain a final sample that matches the trial population. Based on time from study inclusion to RRT and/or progression to stage 3 AKI or reaching an absolute indication for RRT, we identified patients with a treatment history compatible with assignment either to the early RRT arm, the delayed RRT arm, or both. To avoid immortal time bias, patients who died or were discharged without RRT within the grace period for initiating RRT were also considered compatible with the respective RRT strategies. To avoid selection bias, more weight was given to compatible patients less likely to remain compatible through follow-up (inverse probability weighting). The number of patients receiving RRT and ICU mortality were compared between RRT groups (ad hoc approach) and emulated trial arms (emulation approach). RESULTS Finally, 2756 out of 16 434 patients (16.8%) would have been eligible for the ELAIN trial (Fig. 1). Of these, 2565 (93.1%) did not receive RRT at any time (Fig. 1: bottom, white boxes); 9.6% and 1.7% of patients compatible with assignment to early and delayed RRT, respectively, actually received RRT (Fig. 1: bottom, grey boxes). The majority of patients were discharged without RRT, did not progress to stage 3 AKI or did not reach an absolute indication for RRT. In contrast to the ELAIN trial, there was no evidence that early (versus delayed) RRT reduces mortality, neither using the ad hoc approach (59 versus 51%), or the emulation approach (6.4 versus 8%) (Fig. 2). Furthermore, mortality was comparable to that observed in the ELAIN trial using the ad hoc approach in the delayed arm but was >5 times lower using the emulation approach. CONCLUSION Our observational emulation of the ELAIN trial did not reproduce certain key aspects of the benchmark trial. In each of the emulated trial arms <10% of patients received RRT. Along with the much lower mortality (compared with the ELAIN trial), this may in part be due to too broad patient inclusion into the emulation, e.g. because of failure to emulate eligibility criteria unavailable in RCD or an implicit application of undocumented selection criteria in the ELAIN trial. Considering ICU mortality only in patients who did receive RRT (ad hoc approach) may conceal these discrepancies and lead to a false sense of comparability in terms of baseline risk. Although our bias corrections may better signal this issue, final mortality estimates cannot directly be compared head-to-head due to unadjusted confounding. In sum, emulation of even basic protocol components such as eligibility criteria and allocation to treatment strategies can be cumbersome if the available RCD does not contain sufficient information. Future emulations may therefore need to consider careful selection of available proxies of unavailable eligibility criteria and need to rely on adequate adjustment for confounding.