Machine learning phenotypes and heterogeneous albumin effects in cirrhotic AKI: a causal inference study

Current guidelines recommend albumin infusion as a first-line treatment for acute kidney injury (AKI) in patients with cirrhosis. However, recent large-scale randomized trials have questioned its universal benefit. We aimed to identify distinct pathophysiological subphenotypes of cirrhotic AKI using machine learning and to evaluate the heterogeneous treatment effects of albumin infusion across these groups. A retrospective cohort study was conducted using the MIMIC-IV (v3.1) database, including 3209 critically ill patients with cirrhosis and AKI. Unsupervised K-means clustering was applied to eight clinical variables (creatinine, bilirubin, INR, sodium, lactate, platelets, mean arterial pressure [MAP], and baseline albumin) to derive phenotypes. The causal effect of albumin infusion within 48 h of ICU admission on 28-day mortality was assessed using inverse probability of treatment weighting (IPTW)-adjusted Cox proportional hazards models. Three distinct phenotypes were identified: Phenotype A (severe hepatic failure type, N = 405), Phenotype B (hemodynamically stable type, N = 1390), and Phenotype C (typical decompensated type, N = 1414). In the overall population, after adjustment for an expanded set of confounders, albumin infusion was associated with a modest increase in mortality (adjusted HR, 1.211; 95%CI 1.046-1.401). However, significant treatment heterogeneity was observed. In Phenotype B, characterized by the highest MAP (75.8 mmHg) and lowest bilirubin (1.15 mg/dL), albumin infusion was associated with an approximately 1.7-fold increase in the risk of 28-day mortality (adjusted HR, 1.744; 95%CI 1.312-2.320; p < 0.001). Conversely, Phenotype A, representing patients with extreme hyperbilirubinemia, exhibited a potential but non-significant protective trend (HR, 0.895; 95%CI 0.643-1.245). Albumin infusion in cirrhotic AKI exhibits a divergent treatment response dictated by baseline pathophysiological phenotypes. For patients with relatively stable hemodynamics and low inflammatory burden (Phenotype B), aggressive albumin therapy may be harmful, potentially due to fluid overload and cardiorenal congestion. These findings advocate for a phenotype-driven, precision medicine approach rather than a "one-size-fits-all" strategy for volume expansion in cirrhosis.