Development and Validation of an Explainable Machine Learning Model for Prediction of Massive Transfusion in Upper Gastrointestinal Bleeding.

Background: Upper gastrointestinal bleeding (UGIB) is a medical emergency with high mortality, especially when massive transfusion (MT) is required. Traditional scores like Glasgow-Blatchford provide moderate accuracy but overlook complex variable interactions. We developed and validated an explainable machine learning (ML) model for MT prediction in UGIB, improving precision and interpretability. Methods: In this retrospective study, 700 UGIB patients from The Affiliated Hospital of Xuzhou Medical University (2021-2025) were divided into training (n=490) and testing (n=210) cohorts. An external validation cohort (n=300) was sourced from The Fourth Affiliated Hospital of Soochow University. From 18 clinical variables, 8 key features were selected using Boruta and LASSO regression. Seven ML algorithms were compared to identify the optimal model, which was then evaluated for discrimination, calibration, and clinical utility. SHapley Additive exPlanations (SHAP) provided interpretability. Results: The Random Forest (RF) model achieved superior performance with an AUC of 0.862 (95% CI 0.785-0.939) in training, 0.823 (95% CI 0.768-0.879) in testing, and 0.807 (95% CI 0.748-0.866) in external validation. Calibration plots showed strong agreement between predicted and observed probabilities. Decision curve analysis indicated higher net benefit than "treat all" or "treat none" strategies. SHAP analysis ranked impaired mental status, liver cirrhosis, and international normalized ratio (INR) as top predictors, aligning with clinical intuition. Conclusion: The developed machine learning model demonstrated promising performance in identifying UGIB patients at high risk of massive transfusion. While the model shows potential to assist clinicians in optimizing blood management strategies, further prospective validation is required to confirm its clinical utility in diverse settings.