Machine learning-derived clinical phenotypes of ascites in cirrhosis: A multi-center latent class analysis with external validation

Background: Ascites development in cirrhosis reduces five-year survival from 80% to 30%, yet substantial heterogeneity exists among patients with comparable ascites severity. Methods: This multicenter retrospective cohort study included adult liver transplant candidates with ascites at four United States centers (2015-2024). Latent class analysis was performed using seven clinical variables: ascites grade, bilirubin, albumin, platelet count, estimated glomerular filtration rate, systolic blood pressure, and portal vein thrombosis. The derivation cohort (n=625) from University of California San Francisco and Columbia University was randomly split 80/20 for model development and internal validation. External validation was performed at University of Southern California (n=59) and Mayo Clinic (n=93). Primary outcomes were acute kidney injury (AKI) and waitlist mortality over 365 days. Results: Three phenotypes emerged: “CKD-Metabolic” (30.4%, lowest eGFR 76.5 mL/min/1.73m²), “Vasodilatory-Synthetic Dysfunction” (38.9%, highest bilirubin 7.03 mg/dL, lowest blood pressure 117.5 mmHg), and “PVT-Intermediate” (30.7%, severe thrombocytopenia 68.1×10³/μL, 24.0% PVT). In validation cohorts (n=305), phenotypes demonstrated robust classification (mean posterior probability 0.823-0.855). Vasodilatory-Synthetic Dysfunction showed increased AKI risk versus CKD-Metabolic in derivation (MELD 3.0-adjusted HR 2.95, 95% CI 2.07-4.22, p <0.001) and validation cohorts (HR 3.39, 95% CI 1.93-5.94, p <0.001). Competing risk analysis revealed no mortality differences, but Vasodilatory-Synthetic Dysfunction had higher transplantation rates (validation: HR 1.63, 95% CI 1.19-2.22, p =0.002). Conclusions: Three reproducible clinical phenotypes of ascites were identified with distinct risk profiles. The Vasodilatory-Synthetic Dysfunction phenotype was associated with a 3-fold increased risk of AKI that persisted after adjustment for MELD 3.0. No differences in waitlist mortality were observed, likely reflecting effective MELD-based transplant allocation that preferentially removes the highest-acuity patients from the waitlist.