Abstract
Background Intensive care unit (ICU) case triage in cirrhotic-associated upper gastrointestinal bleeding (UGIB) varies greatly by local practice, while existing risk stratification data prioritizes mortality and rebleeding rather than ICU resource use. Methods We performed a retrospective cohort study of consecutive cirrhotic patients with UGIB in a tertiary referral center between January 1, 2024, and December 31, 2025. Our primary endpoint was in-hospital mortality, and a secondary endpoint was ICU triage decision. The institutional ICU criteria included persistent hypotension after initial fluid resuscitation (systolic blood pressure (SBP) <90 mmHg and/or vasopressor requirement), respiratory distress requiring advanced support (intubation or escalating oxygen requirement) or severe metabolic acidosis (pH ≤7.20 and/or lactate ≥4 mmol/L). We compared discrimination (Area under the receiver operating characteristic curve (AUROC), with bootstrap 95% confidence interval (CI)) of UGIB scores (Albumin, International Normalized Ratio, Mental status, Systolic blood pressure, age 65 (AIMS65); Rockall and Glasgow-Blatchford) and liver-specific scores (Model for End-stage Liver Disease with Sodium (MELD-Na), Child-Pugh, Albumin to Bilirubin index (ALBI), Platelet-Albumin-Bilirubin index (PALBI), International Normalized Ratio to Albumin (PTAR), Aspartate Aminotransferase (AST) to platelet ratio index (APRI)). Logistic regression models assessed incremental value of combined physiology and liver severity models. A prespecified sensitivity analysis restricted the cohort to endoscopically confirmed variceal bleeding. Results The cohort included 224 patients with a median age 58 years and 174 (77.7%) were men. Variceal bleeding was present in 212 (94.6%) cases. ICU admission occurred in 37 (16.5%) cases and in-hospital mortality occurred in 53 (23.7%) cases. For mortality, MELD-Na showed the highest discrimination (AUROC 0.898, 95% CI 0.849-0.940), followed by PTAR (0.865) and ALBI (0.852). For ICU admission, AIMS65 (0.930, 95% CI 0.896-0.960) and Rockall (0.919, 95% CI 0.885-0.949) outperformed liver-only scores. A combined MELD-Na + AIMS65 model improved mortality discrimination (AUROC 0.919) and calibration versus either score alone. Conclusions In cirrhosis-associated UGIB, liver dysfunction severity (estimated through MELD-Na and related scores) best predicts in-hospital mortality, while acute physiology (AIMS65 and Rockall scores) best predicts ICU admission. Combining physiology and liver severity improved mortality prediction. ICU analyses reflect prediction of observed triage decisions under institutional criteria and require external validation against organ-support endpoints.