Comparative performance of pre-interventional prognostic scores in predicting 30-day mortality and ICU admission in gastrointestinal bleeding.

Comprehensive intra-cohort comparisons of pre-interventional prognostic scores for gastrointestinal bleeding (GIB) are lacking. We compared 12 pre-interventional scores [Glasgow-Blatchford Score (GBS), modified GBS (mGBS), AIMS65, Pre-Rockall, ABC, Harbinger, T-Score, Canuka, SHA2PE, Oakland, NOBLADS, and Modified Early Warning Score (MEWS)] for predicting 30-day mortality and intensive care unit (ICU) admission in emergency department patients with upper GIB (UGIB) and lower GIB (LGIB). This retrospective, single-center cohort study evaluated 2020 patients with GIB between January 2014 and March 2025, categorized into UGIB (n = 1524) and LGIB (n = 496) groups. Primary outcomes were 30-day mortality and ICU admission. Scores were evaluated in two ways: an indication-restricted analysis, in which each score was applied only to the bleeding location for which it was developed, and a comprehensive analysis, in which all 12 scores were applied to both subgroups. Discriminative performance was assessed using the area under the curve (AUC); the highest-performing scores per outcome were compared using DeLong's method. For 30-day mortality, AIMS65 showed the highest discrimination in UGIB (AUC = 0.850) and NOBLADS in LGIB (0.809) in the indication-restricted analysis. In the comprehensive analysis, AIMS65 again showed the highest discrimination (AUC = 0.850) in UGIB, outperforming NOBLADS (AUC = 0.812; p = 0.035), ABC (AUC = 0.801; p = 0.023), and Pre-Rockall (AUC = 0.798; p = 0.033), but not Canuka (AUC = 0.821; p = 0.124). For LGIB, GBS achieved the highest AUC (0.851), followed by MEWS (0.841) and mGBS (0.832), with no significant pairwise differences among the top five scores. For ICU admission, the highest-performing scores demonstrated moderate discrimination in the comprehensive analysis (AUC range, 0.755-0.788 in UGIB; 0.746-0.768 in LGIB). AIMS65 showed the highest discrimination for 30-day mortality in UGIB in both analyses, whereas no score demonstrated clear superiority in LGIB. NOBLADS ranked highest within its original indication in LGIB and retained strong discrimination when applied to UGIB, supporting its potential role as a location-independent mortality predictor. No score reliably predicted ICU admission, indicating a persistent gap in pre-interventional risk stratification.