Esophageal variceal bleeding (EVB) remains a major cause of morbidity in portal hypertension. Classical predictors provide limited individualized risk estimation. This study aimed to identify predictors of EVB in portal hypertensive gastropathy (PHG) using both multivariate logistic regression and an artificial intelligence (AI) based machine-learning model.

A retrospective cohort of 213 PHG patients (64 with EVB) was analyzed. Clinical, laboratory, and endoscopic variables,including ascites, encephalopathy, platelet count, MELD, albumin, creatinine, H. pylori status, variceal grade, PHG severity, GAVE, NSBB use, and intestinal metaplasia were collected. Univariable and multivariable logistic regression models were constructed. A Random Forest classifier was developed to assess non-linear interactions and determine variable importance. Model performance was evaluated using the area under the ROC curve (AUC).

Multivariate analysis identified three independent predictors of EVB: encephalopathy (OR 2.46, p=0.013), ascites (OR 1.81, p=0.018),  and low platelet count per 10,000/mm³ decrement (OR 0.98, p=0.048). The Random Forest model achieved an AUC of 0.75. AI-based feature importance analysis revealed platelet count as the strongest predictor, followed by MELD, creatinine, albumin, age, ascites, NSBB use, and high-grade varices.

In patients with PHG, encephalopathy, ascites, and thrombocytopenia independently predict EVB. AI modeling highlights additional contributors, particularly MELD-related parameters, capturing patterns not detected by classical analysis. Integrating statistical and AI-derived predictors may enhance individualized bleeding-risk stratification.