Personalized Fluid Management in Patients With Sepsis and Acute Kidney Injury: A Casual Machine Learning Approach

Abstract

Importance: IV fluids are the cornerstone for management of acute kidney injury (AKI) after sepsis but can cause fluid overload. A restrictive fluid strategy may benefit some patients; however, identifying them is challenging. Novel causal machine learning (ML) techniques can estimate heterogenous treatment effects (HTEs) of IV fluids among these patients.

Objectives: To develop and validate a causal-ML framework to identify patients who benefit from restrictive fluids (< 500 mL fluids within 24 hr after AKI).

Design setting and participants: We conducted a retrospective study among patients with sepsis who developed acute kidney injury (AKI) within 48 hours of ICU admission. We developed a causal-ML approach to estimate individualized treatment effects and guide fluid therapy. We developed the model in Medical Information Mart for Intensive Care IV and externally validated it in Salzburg Intensive Care database.

Main outcomes and measures: Our primary outcome was early AKI reversal at 24 hours. Secondary outcomes included sustained AKI reversal and major adverse kidney events by 30 days (MAKE30). Model performance to identify HTE of restrictive IV fluids was assessed using the area under the targeting operator characteristic curve (AUTOC), which quantifies how well a model captures HTE, and compared with a random forest model.

Results: Causal forest model outperformed random forest in identifying HTE of restrictive IV fluids with AUTOC 0.15 vs. -0.02 in external validation cohort. Among 1931 patients in external validation cohort, the model recommended restrictive fluids for 68.9%. Among these, patients who received restrictive fluids demonstrated significantly higher rates of early AKI reversal (53.9% vs. 33.2%, p < 0.001), sustained AKI reversal (34.2% vs. 18.0%, p < 0.001), and lower rates of MAKE30 (17.1% vs. 34.6%, p = 0.003). Results were consistent in the adjusted analysis.

Conclusions and relevance: Causal-ML framework outperformed random forest model in identifying patients with AKI and sepsis who benefit from restrictive fluid therapy. This provides a data-driven approach for personalized fluid management and merits prospective evaluation in clinical trials.

Keywords: Policy Tree; acute kidney injury; causal machine learning; individual treatment effect; restrictive fluids.

Figure 1.

Comparison of causal and random forest models: A, area under the targeting operator characteristic curve scores for individual treatment effects in development and validation cohorts (p values from two-sample t test); B, quantile-based distributions of estimated treatment effects by causal forest and random forest models. MIMIC-IV = Medical Information Mart for Intensive Care IV, SICdb = Salzburg Intensive Care database.

Figure 2.

Policy tree for restrictive fluid strategy in septic patients with acute kidney injury (AKI). Features: age, blood urea nitrogen (BUN), diastolic blood pressure (DBP), heart rate, hematocrit, mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), partial pressure of carbon dioxide (Pco₂), partial pressure of oxygen (Po₂), RBC count, platelet, respiratory rate, SBP, weight, temperature, urine output.

Figure 3.

Impact of restrictive fluid strategy among patients stratified by predicted benefit from receipt of restrictive fluid strategy: A, Proportion (p values from χ² tests), B, odds ratio (OR). AKI = acute kidney injury, MAKE30 = major adverse kidney events by 30 days, MIMIC-IV = Medical Information Mart for Intensive Care IV, SICdb = Salzburg Intensive Care database.