(1328) Leveraging AKI Heterogeneity Characteristics in CRRT Timing: A Decision Analysis

Introduction: In-hospital mortality for Acute Kidney injury (AKI) patients receiving Continuous Renal Replacement Therapy (CRRT) is 40-55%. CRRT timing has been implicated as a cause, but CRRT initiation trials have been inconclusive. Patient and disease heterogeneity characteristics may be the culprit for those results: They preclude from identifying the optimal CRRT initiation time by diluting trial effect sizes. Accordingly, there is an unmet need to reconcile the underlying role of AKI heterogeneity with CRRT timing decisions as this may impact mortality rates.

Objective: We aimed to optimize CRRT initiation using AKI heterogeneity characteristics in patients with AKI.

Methods: The Medical Information Mart for Intensive Care (MIMIC) IV dataset is utilized to augment a decision tree model with 1200 adult ICU patients with a primary or secondary diagnosis of AKI (at or up to 1 week after ICU admission) but who have not yet started CRRT. The decision tree is constructed as a tree root with two main 'trunks', each terminating in a decision node. Time from ICU admission until CRRT initiation is used to define early ( < 12 hours) vs late (12-72 hours) CRRT at those nodes. Each decision node then divides into smaller branches with intermediate nodes containing over 60 AKI characteristics, including patient demographics, ICU admission vitals and procedures, and laboratory values. Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression is used to filter which nodes to keep and their order (lower p-value nodes placed nearer the root). Lastly, our primary outcome (dead or alive up to 90 days post ICU discharge) is placed at terminating nodes or ‘leaves’ to calculate tree branch pay-offs.

Results: Patients from MIMIC IV are identified using ICD-9 codes and the KDIGO criteria for AKI. Every patient is matched to 2 identical tree branches, 1 from each trunk. The tree uses each branch and leaves to calculate the expected value (EV) of the corresponding trunk (decision), highlighting the node with the higher EV.

Conclusions: Our decision tree highlights the priority AKI characteristics to consider when initiating CRRT, with 90-day mortality reduction as the primary target. This tree will be used in future studies to prospectively examine its impact on 90-day mortality in AKI patients undergoing CRRT.