Development and Validation of a Machine Learning Model for Predicting 30-Day Mortality and Early Endoscopic Hemostatic Intervention in Acute Upper Gastrointestinal Bleeding.

OBJECTIVE

Accurate pre-endoscopic risk stratification is essential in acute upper gastrointestinal bleeding (UGIB). Established scores predict mortality but have limited ability to identify patients requiring early therapeutic endoscopy. We developed and temporally validated a pre-endoscopic machine learning model to predict 30-day mortality and early (≤24 h) endoscopic hemostatic intervention and assessed model interpretability.

METHODS

This retrospective cohort study included consecutive adults admitted with UGIB between November 2023 and November 2025. Only pre-endoscopic variables were considered. A gradient-boosted decision tree model was developed in a derivation cohort (n=762) and temporally validated in a subsequent cohort (n=353). Discrimination, calibration, decision curve analysis, SHAP-based explainability, and complete-case sensitivity analyses were assessed and compared with Glasgow-Blatchford, AIMS65, and ABC scores.

RESULTS

Among 1115 patients (median age: 69 y; 63.5% male; 19.0% variceal bleeding), 30-day mortality was 7.8% and early hemostatic intervention was 29.8%. In the validation cohort, the model achieved AUROCs of 0.874 for mortality and 0.848 for early intervention, with good calibration. Discrimination was significantly higher than established scores for both endpoints (all P<0.05). At a 20% intervention threshold, sensitivity was 84.9% and negative predictive value was 92.3%. Net benefit was greater across clinically relevant thresholds. SHAP analyses identified clinically plausible contributors, and complete-case analysis yielded similar performance.

CONCLUSIONS

A pre-endoscopic machine learning model accurately predicted short-term mortality and early therapeutic endoscopy in UGIB, outperformed established risk scores, and showed interpretable feature-contribution patterns. External validation and prospective evaluation are warranted before clinical implementation.