Construction of a Predictive Model for Rebleeding Risk in Upper Gastrointestinal Bleeding Patients Based on Clinical Indicators such as Helicobacter pylori Infection

Wei Zang¹Ze Lin¹Yanduo Zhao¹Tianshi Jia¹Xinglong Zhang^2*

• ¹China Medical University, Shenyang, China
• ²Fourth Affiliated Hospital of China Medical University, Shenyang, Liaoning Province, China

Background: Upper gastrointestinal hemorrhage (UGIB) has an annual incidence of 60 cases per 100,000 people, with 40% of cases linked to hemorrhagic ulcers. Ulcer formation is often caused by Helicobacter pylori (H.pylori) infection, NSAIDs use, and other factors, making ulcers a primary cause of UGIB. H.pylori induces chronic gastritis with mucosal neutrophil infiltration, suggesting it underlies bleeding lesions. Globally, 50% of the population carries H.pylori. UGIB patients face higher in-hospital mortality due to rebleeding, which increases death risk and can lead to complications like shock and organ failure. Common clinical scores for UGIB include Rockall (RS), AIMS65, and Glasgow-Blatchford (GBS). However, many hospitals lack timely endoscopic equipment, hindering accurate diagnosis.Methods: This study included 254 UGIB patients from Shengjing Hospital, China Medical University. Clinical data on H.pylori infection, age, shock, comorbidities, systolic blood pressure, blood urea nitrogen, hemoglobin, pulse, black stool, syncope, and liver disease were collected. A deep learning model was developed using Transformer as a feature extractor and KAN as a classifier, with data categorized into clinical information, vital signs, lab tests, and stool exams. The model employed five-fold cross-validation and was compared to machine learning methods (decision tree, random forest, logistic regression, K-nearest neighbor) and clinical scores (RS, AIMS65, GBS).Results: Rebleeding risk was higher in men (62.5%) than women (43.47%) and in patients with comorbidities (60.37%). Hemoglobin levels varied with infection severity (P<0.05). Rebleeding detection rates for RS, AIMS65, and GBS were 16.14%, 0%, and 77.17%, respectively. Among machine learning models, random forest achieved the highest test set accuracy (0.68). The deep learning model outperformed others, with validation and test set accuracies of 0.9750 and 0.9615, respectively.Conclusion: This study developed a clinical model using H.pylori infection and clinical data to predict rebleeding in UGIB patients, offering a non-endoscopic, efficient tool for small and medium-sized hospitals to guide clinical decisions.