Predicting severe renal dysfunction in alcohol-associated cirrhosis: Comparative performance of liver function scores and machine learning models

Abstract

Background Renal dysfunction is a frequent and clinically relevant complication of cirrhosis, yet chronic kidney disease (CKD) often remains underrecognized, particularly in non-acute settings. Early identification of at-risk patients is essential to guide timely interventions. Although MELD, Child-Pugh Score (CPS), APRI, and FIB-4 are widely used to assess hepatic disease severity, their predictive value for advanced renal dysfunction is uncertain. Methods In this retrospective cohort study (2014–2021, Klinikum Stuttgart), we evaluated the ability of MELD, CPS, APRI, and FIB-4 to predict severe renal dysfunction (chronic kidney disease [CKD] stage ≥ 3, according to Kidney Disease: Improving Global Outcomes [KDIGO] classification) in patients with alcoholic cirrhosis. Logistic and linear regression analyses were performed. In addition, machine learning (ML) models were trained to identify non-renal predictors of CKD stage ≥ 3. Results Among 131 patients (mean age 62.8 ± 11.3 years; 71% male), 33% met criteria for KDIGO stage ≥ 3. MELD was significantly associated with advanced CKD (OR = 1.379, p < 0.001), with prevalence increasing from 17% (MELD ≤ 9) to 80% (MELD ≥ 20). CPS showed an inverse association (p = 0.002), while APRI and FIB-4 were not predictive. The optimized Random Forest model, refined through ROSE oversampling and feature selection, achieved an AUC of 0.757, with 76% accuracy, 82% sensitivity (KDIGO < 3), and 63% specificity (KDIGO ≥ 3). Conclusion MELD was the most reliable conventional score for identifying advanced renal dysfunction in alcoholic cirrhosis. ML-based models incorporating routinely available clinical parameters further improved predictive performance and may support risk stratification in this high-risk population.