Development and validation of a LightGBM-based diagnostic model for acute-on-chronic liver failure in patients with liver cirrhosis: a multicenter retrospective study in an HBV-predominant cohort

Background Acute-on-chronic liver failure (ACLF) is a severe and often fatal complication in patients with liver cirrhosis. Because its clinical manifestations in the early stage are often masked by the chronic liver disease, timely identification and risk stratification of ACLF are critical for improving clinical outcomes. Therefore, the present study aimed to develop a model for timely identification and risk stratification of ACLF. Methods In this multicenter retrospective study, data were collected from seven tertiary hospitals in Southwest China, including 17,567 patients with cirrhosis between 2012 and 2023. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression. Multiple machine learning algorithms, including Light Gradient Boosting Machine (LightGBM), were developed and compared for ACLF prediction. Model performance was assessed and validated using the area under the receiver operating characteristic curve (AUC–ROC), area under the precision–recall curve (AUC–PR), calibration curves, and decision curve analysis. The TreeSHAP framework was applied for model interpretation. Results A total of 13,490 eligible patients with cirrhosis were included, of whom 1,250 (9.27%) developed ACLF. The cohort was randomly divided into a training set ( n = 9,758) and a test set ( n = 4,182). Among the evaluated algorithms, the LightGBM model demonstrated the best predictive performance. In the training set, the LightGBM model yielded an AUROC of 0.960, an AUPRC of 0.727, an accuracy of 0.882, a sensitivity of 0.954, a specificity of 0.875 and an F1 score of 0.593. In the test set, the corresponding values were 0.948, 0.655, 0.866, 0.912, 0.861, and 0.549, respectively. SHAP analysis identified the nine most important predictive features, ranked by their contribution to the model: international normalized ratio, total bilirubin, prealbumin, hepatitis B virus infection, alanine aminotransferase, platelet count, total bile acid, hepatic encephalopathy and alcoholic liver disease. Conclusion A LightGBM-based diagnostic model incorporating nine key clinical variables was successfully developed. This model facilitates timely identification and risk stratification of ACLF in patients with cirrhosis, which may support clinical decision-making at presentation and targeted clinical interventions.