Background <p>Peptic ulcer bleeding (PUB) is a common, life-threatening disorder. Although endoscopic therapy achieves hemostasis, short-term rebleeding remains frequent, increasing mortality and healthcare burden. Existing tools lack accuracy and interpretability. This study aimed to develop and validate an interpretable machine learning (ML) model to predict rebleeding risk after initial therapy in PUB.</p> Methods <p>We retrospectively analyzed patients with Forrest Ia–IIb PUB undergoing emergency hemostasis at Shenzhen Hospital, Southern Medical University (June 2019–September 2023). Among 204 patients, 25 (12.25%) rebled. Nineteen clinical and laboratory features were collected, and 11 ML algorithms were evaluated. Data were split 7:3 into training and testing sets. Model performance was assessed by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, F1 score, and Youden index. The best model was interpreted using SHapley Additive exPlanations (SHAP).</p> Results <p>Categorical boosting (CATBoost) achieved the best performance (AUC 0.980 training; 0.817 testing). Calibration confirmed stability. SHAP identified Rockall score, prothrombin time (PT), activated partial thromboplastin time (APTT), AIMS65 score, blood urea nitrogen (BUN), creatinine, and lesion location (particularly the descending duodenum) as the most important predictors. A web-based calculator was developed for individualized risk estimation.</p> Conclusion <p>We established and validated an interpretable CATBoost model using routine variables to predict short-term rebleeding in PUB. This tool may aid early identification of high-risk patients and support individualized management strategies.</p> Trial Registration <p>This study was registered at <a href="https://www.chictr.org.cn/">https://www.chictr.org.cn/</a> with the registration number ChiCTR2300079290.</p> Graphical abstract <p></p>