Development of an AI-driven predictive model and decision support system for managing acute-on-chronic liver failure: insights from the KACLiF cohort

초록

Background and aims: Acute-on-Chronic Liver Failure (ACLF) is associated with high mortality due to multisystem organ failure. Accurate prediction of mortality is crucial for guiding dialysis therapies and determining the optimal timing for liver transplantation. We hypothesized that artificial intelligence (AI) models are more precise than standard models in predicting outcomes in ACLF. Method: A novel AI-based model was developed using data collected from patients with MELD scores ≥21. Data were prospectively collected from July 2015 to August 2018 and retrospectively from January 2013 to December 2013 from the Korean Acute-on-Chronic Liver Failure (KACLiF) cohort. The prospective data were split into training and validation sets in a 7:3 ratio and used as the derivation cohort (n = 294), while the retrospective data served as an independent validation cohort (n = 177). Logistic Regression, Random Forest Classifier, XGBoost Classifier, Decision Tree and Elastic Net were employed to refine the selection of significant features. Random Forest was used to evaluate the selected features’ ability to predict survival. Results: The mean MELD scores for the derivation and validation cohorts were 30.1 and 29.9, respectively. The 30-day mortality rates in the derivation and validation cohorts were 26.2% and 29.4%, while the 90-day mortality rates were 34.0% and 40.7%, respectively. In the validation cohort, XGBoost demonstrated the highest AUC of 0.82 for predicting 30-day mortality, and Logistic Regression achieved the highest AUC of 0.84 for predicting 90-day mortality. Baseline international normalized ratio (INR), prior acute deterioration events, circulatory failure, respiratory failure, ascites, hepatic encephalopathy, albumin levels, and systemic inflammatory response syndrome (SIRS) were identified as the top features influencing 30- and 90-day outcomes. The model demonstrated relatively high predictive power and outperformed previously reported models, including MELD (0.74 and 0.67), MELD-Na (0.62 and 0.68), MELD 3.0 (0.69 and 0.65), CLIF-ACLF score (0.63 and 0.63), and CLIF-SOFA score (0.74 and 0.74), in predicting 30- and 90-day mortality. Conclusion: The AI-based model outperformed previously established models in predicting mortality among patients with severe ACLF, providing a more accurate tool for clinical decision-making.