Decision-support system for emergency endoscopic surgery in nonvariceal upper gastrointestinal bleeding

Nonvariceal upper gastrointestinal bleeding (NVUGIB) is a critical condition requiring timely endoscopic intervention to prevent severe complications. Although widely used, existing risk-scoring systems, such as the Glasgow-Blatchford score (GBS) and Nscore, often exhibit limited predictive accuracy and adaptability across diverse patient populations. This study developed a machine-learning-based clinical decision support system, utilizing tabular data, to predict the need for endoscopic intervention in NVUGIB patients, with the potential to support clinical decision-making and resource utilization, pending further validation. A retrospective dataset comprising 1,989 NVUGIB patients was collected from three Japanese medical institutions between January 2016 and December 2020 and partitioned into training (2016–2019; n = 1,467) and validation (2020; n = 358) sets. For model development, extreme gradient boosting (XGBoost), a decision-tree-based ensemble algorithm, was employed. The Fβ score was used to optimize thresholds, balancing recall and precision. Model performance was assessed using metrics such as the area under the receiver operating characteristic curve (AUROC), precision, and recall. A comparative analysis was performed against established risk-scoring systems. The XGBoost model achieved an AUROC of 0.822 (95% CI: 0.777–0.863), achieving higher discrimination than Nscore alone (AUROC: 0.776), which was itself included as an input feature, and than GBS (0.576). At a β = 1.5 threshold, an 89.5% sensitivity and F1-score of 0.723 were achieved. Key predictors, including the blood urea nitrogen to creatinine (BUN/Cr) ratio and Nscore, were identified through feature-importance analysis, demonstrating their substantial contributions to model performance. The XGBoost-based system demonstrated enhanced predictive accuracy regarding the need for endoscopic intervention, suggesting a potential for optimized resource allocation. However, external validation in diverse populations is essential for clinical implementation.