Objective To construct a non-invasive predictive model for screening patients with hepatitis B-related high-risk esophageal gastric varices (EGV) by combining multimodal ultrasound indicators with serological indicators using the random forest algorithm.Methods Data were collected from hepatitis B patients who underwent endoscopic examination, 2D-SWE, liver and spleen ultrasound, and relevant serological tests at Mengchao Hepatobiliary Hospital of Fujian Medical University from January 2019 to March 2024. The random forest algorithm was employed to rank the importance of variables, selecting those that significantly enhance predictive performance to build a non-invasive predictive model, named USE (Ultrasound and Serological predict EGV). This model was visualized using SHAP (SHapley Additive exPlanations) and compared with other classic non-invasive predictive indicators: FIB-4, PSR, LSPS, and MELD, in terms of predictive performance.Results A total of 317 patients were included in this study, of whom 127 were high-risk EGV patients. The average age was 53.53±12.31 years. The top ten indicators in terms of importance were: spleen stiffness, platelet count, creatinine, albumin, spleen thickness, direct bilirubin, portal vein flow rate, alanine aminotransferase, liver blood flow, and spleen length. By incorporating the top seven variables, the USE predictive model achieved a high predictive performance with an AUC (Area Under the Curve) of 0.93. Compared with other non-invasive indicators, USE demonstrated a higher AUC, better improvement effect, more accurate predictive distribution, superior profit curve, and greater cumulative gain.Conclusion This model provides a non-invasive method for screening high-risk EGV patients, with significant clinical implications for prevention and treatment.