Machine Learning-Based Severity Classification in Decompensated Liver Cirrhosis: Incremental Value of Oxidative Stress Biomarkers for Predicting Ascites and Hepatic Encephalopathy

Oxidative stress biomarkers are elevated in liver cirrhosis, but their clinical utility for severity staging and complication prediction remains uncertain. This retrospective single-centre study enrolled 90 patients with decompensated cirrhosis (Child–Pugh classes B and C) to evaluate serum malondialdehyde (MDA) and 8-isoprostane (8-isoPGF2α) as predictors of Child–Pugh severity, severe ascites, and severe hepatic encephalopathy, and to quantify their incremental value within supervised machine learning models. Four algorithms—logistic regression, Random Forest, Gradient Boosting, and Support Vector Machine—were evaluated using stratified 10-fold cross-validation; logistic regression models with and without oxidative stress biomarkers were compared for the prediction of ascites and encephalopathy. Routine biochemical parameters effectively discriminated Child–Pugh class B from C, with machine learning models achieving AUC-ROC values of 0.921–0.972. Neither MDA nor 8-isoPGF2α differed between Child–Pugh classes or across ascites categories, and both failed to improve ascites prediction (ΔAUC = −0.015). For severe hepatic encephalopathy, the extended model showed modest but consistent improvements in accuracy (+3.4 percentage points), sensitivity (+6.4%), and model fit, suggesting an outcome-specific complementary role consistent with the established involvement of lipid peroxidation in ammonia neurotoxicity. These findings support the use of machine learning for automated cirrhosis severity classification and indicate that oxidative stress biomarkers hold selective relevance for hepatic encephalopathy rather than global disease staging.