Prediction of Subsequent Acute-on-Chronic Liver Failure After a 24-Hour Landmark in Cirrhotic Patients With Variceal Upper Gastrointestinal Bleeding: A Multicenter Retrospective Cohort Study
Digestive Diseases · 2026-11-07
DOI: 10.1159/ddi/acnag017 ↗Abstract
Abstract Background: Some cirrhotic patients admitted with variceal upper gastrointestinal bleeding experience clinical deterioration during hospitalization and develop acute-on-chronic liver failure (ACLF), a critical event strongly associated with short-term mortality. Since ACLF may be reversible in its early stages, risk stratification prior to overt ACLF is clinically important. Methods: We conducted a multicenter retrospective cohort study (2015–2025) involving cirrhotic patients admitted for esophagogastric variceal bleeding. A 24-hour landmark analysis was performed among patients who remained ACLF-free during the first 24 hours after admission, with subsequent ACLF before discharge as the primary outcome. Predictors were identified using LASSO and best-subset selection procedures. Logistic regression and random forest models were developed and compared with the available AIMS65 and Glasgow–Blatchford score (GBS) risk categories. Model performance was assessed in terms of discrimination (AUC), calibration (CITL and calibration slope), and clinical utility (decision curve analysis with integrated net benefit). External validation was performed using an independent cohort. Results: Of 1,787 screened patients, 555 were included in the derivation cohort and 240 in the external validation cohort after application of the 24-hour landmark. Subsequent ACLF occurred in 281 and 114 patients, respectively. Four predictors were retained in the final model: shock index (SI), neutrophil-to-lymphocyte ratio (NLR), prothrombin time (PT), and total bilirubin (TBil). In the derivation cohort, the AUCs were 0.752 for logistic regression and 0.715 for the random forest model, compared with 0.654 for the AIMS65 categories and 0.699 for the GBS categories. In the validation cohort, the corresponding AUCs were 0.723 and 0.715, compared with 0.668 for the AIMS65 categories and 0.667 for the GBS categories, respectively. Calibration was more stable for logistic regression than for random forest. Decision-curve analyses indicated positive but overlapping net benefits across models, suggesting potential complementary use rather than clear superiority over the benchmark risk categories. Conclusions: We developed and externally validated models to estimate the risk of subsequent ACLF after the first 24 hours of hospitalization following variceal UGIB. The logistic regression model demonstrated stable probability calibration and may provide complementary information for post-admission ACLF risk reassessment. Prospective studies are required to evaluate its clinical impact and broader applicability.
EUVIMED