Development and validation of a prognostic model for acute-on-chronic liver failure.

Background: Prognostic assessment in acute-on-chronic liver failure (ACLF), particularly in HBV-endemic regions, remains challenging due to the limited accuracy of conventional models. We aimed to develop and validate a novel, machine learning-based model incorporating liver reserve function to improve individualized prediction of short-term outcomes in HBV-ACLF. Methods: Baseline demographics, clinical features, laboratory findings, and 90-day follow-up data were retrospectively collected from 496 patients (training/internal subgroups) and 52 patients (external validation) with HBV-ACLF. Twelve machine learning algorithms were systematically evaluated for prognostic performance. The optimal model was established using the LASSO-RF approach, with key variables identified by SHAP values. Model accuracy was assessed by ROC analysis and compared with MELD and CTP scores. An interactive web calculator (https://syx123.shinyapps.io/deploy_shiny/) was developed to facilitate clinical use. Results: We initially screened 23 potential clinical risk factors for predicting ACLF prognosis. Subsequently, using the LASSO-RF model, 15 key variables were selected for model construction. The final LASSO-RF model achieved an AUC of 0.99 in the training cohort and 0.98 in the validation cohort for predicting 90-day mortality, outperforming conventional scoring systems such as MELD and CTP. To facilitate clinical application, an online tool (https://syx123.shinyapps.io/deploy_shiny/) was developed to provide real-time risk scores and 90-day mortality predictions for individual patients. Conclusions: Liver reserve function indicators, particularly EHBF and ICG-R15, play a pivotal role in prognosticating HBV-ACLF outcomes. The developed model and its accompanying online tool enable accurate risk stratification and have the potential to guide timely and individualized clinical management.