The value of predicting intervention and rebleeding risk for patients with acute upper gastrointestinal bleeding based on machine learning

Objective: To investigate the value of machine learning (ML) in predicting blood transfusion intervention and rebleeding in patients with acute upper gastrointestinal bleeding (AUGIB). Methods: A retrospective analysis was conducted on the clinical data of 512 AUGIB patients who were admitted to the Third People’s Hospital of Yunnan Province from July 2020 to October 2023. Variable importance analysis was performed using eXtreme gradient boosting (XGBoost), and the top 10 factors in importance ranking were selected as variables in the model. Classification predictions were carried out and compared using logistic regression, XGBoost, random forest, support vector machine (SVM), and K-nearest neighbors algorithms (KNN). The best model was chosen, and interpretable analysis of the features selected by ML was performed using SHAP plots. The clinical value was assessed by comparing the best model with the commonly used AUGIB scoring systems. Results: The XGBoost algorithm model identified the top 10 risk factors for transfusion intervention as hemoglobin, international normalized ratio (INR), albumin, systolic blood pressure, urea, anesthesia risk score, pulse, creatinine, age, and presence of shock. Using these important features for modeling, the XGBoost algorithm provided the best predictive performance for transfusion intervention in AUGIB patients and it achieved the highest score, indicating its superior ability to identify patients at risk for gastrointestinal bleeding who required transfusion intervention, and outperforming the common clinical Glasgow Blatchford (GBS), AIMS65, ABC, and T scoring systems. According to the XGBoost algorithm model, the top 10 important features scores for patients with rebleeding were age, creatinine, INR, hemoglobin, anesthesia risk score, albumin, systolic blood pressure, urea, liver cirrhosis, and gender. Modeling with these top 10 risk factors, the XGBoost algorithm also showed the best predictive performance for re-bleeding in AUGIB patients, surpassing the aforementioned four scoring systems. Conclusion: In predicting the value of transfusion intervention and rebleeding in AUGIB patients, ML model is superior to GBS, AIMS65, ABC and T scoring systems. The XGBoost model algorithm is superior, with better effectiveness.