HOLD-B: A Practical Score for Early Risk Stratification in Upper Gastrointestinal Bleeding

Background/Objective: Early identification of high-risk patients with acute upper gastrointestinal bleeding (UGIB) is essential for appropriate triage, timely intervention, and optimal resource allocation in the emergency department (ED). Existing risk scores have limitations in practicality and predictive performance. We aimed to develop and internally validate a novel pre-endoscopic score based solely on readily available ED variables to predict clinically significant high-risk outcomes in UGIB. To develop a novel pre-endoscopic risk score using routinely available emergency department parameters and to evaluate its predictive performance for clinically significant high-risk outcomes in patients with acute upper gastrointestinal bleeding. Methods: This retrospective single-center prognostic model development study with internal validation included 312 adults with endoscopy-confirmed non-variceal UGIB presenting to a tertiary ED between January 2023 and January 2025. High-risk status was defined as the occurrence of at least one of the following during hospitalization: blood transfusion, endoscopic hemostatic intervention, intensive care unit admission, rebleeding within 30 days, or in-hospital mortality. This composite endpoint included both clinical outcomes and management-related interventions. Independent predictors were identified using multivariable logistic regression. The HOLD-B score (Hemoglobin, Lactate, Diastolic blood pressure, and Blood urea nitrogen) was derived from significant variables. Discriminative performance was evaluated using receiver operating characteristic (ROC) analysis and compared with established scores. Results: Of the study population, 51.6% were classified as high risk. Hemoglobin ≤8.05 g/dL, blood urea nitrogen >44.5 mg/dL, lactate >2.15 mmol/L, and diastolic blood pressure ≤67.5 mmHg were independent predictors of high-risk status. The HOLD-B score (0-10 points) demonstrated good discriminative performance (AUC 0.798; 95% CI 0.749-0.846), outperforming Rockall (0.687), GBS (0.733), AIMS65 (0.576), ABC (0.626), ABL (0.716), and Pre-RS (0.599). At a cutoff ≥4, the HOLD-B score yielded 71.4% sensitivity and 77.5% specificity. However, in a sensitivity analysis excluding blood transfusion from the composite endpoint, the AUC decreased from 0.798 to 0.676, and hemoglobin was no longer an independent predictor, indicating that part of the observed model performance may have been influenced by the inclusion of transfusion in the composite endpoint. DeLong analysis demonstrated statistically significant differences compared with several established scores (p < 0.01). Conclusions: In this endoscopy-confirmed non-variceal UGIB cohort, the HOLD-B score demonstrated promising discriminative performance for predicting a clinically significant composite outcome using only emergency department variables. Because this composite outcome included management-related interventions in addition to clinical events, the model should not be interpreted as predicting individual clinical endpoints. Prospective external validation in broader emergency department populations is required before routine clinical implementation.