Abstract
BACKGROUND AND OBJECTIVES: Wilson disease (WD), an inherited copper metabolism disorder, is a cause of acute-on-chronic liver failure (ACLF), posing life-threatening risks due to rapid progression. This study aimed to develop a machine learning (ML)-based model to predict ACLF risk in WD patients. METHODS: We retrospectively analyzed 3692 WD patients (Leipzig score ≥ 4) from The First Affiliated Hospital of Anhui University of Chinese Medicine (2014-2024), including 104 ACLF and 104 non-ACLF cases. The original data set was randomly divided into the training and test cohorts in a ratio of 7:3. Demographic, biochemical, and ultrasound data were collected. Six ML algorithms (LR, SVM, KNN, ExtraTrees, XGBoost, LightGBM) were applied to construct a predictive model, with SHAP explaining feature importance. RESULTS: The XGBoost model achieved optimal performance (AUC: 0.998, accuracy: 0.968). Key predictors included TBA, APTT, diagnosis age, onset age, Hb. Elevated TBA, APTT and diagnosis age correlated with higher ACLF risk, while reduced onset age and Hb indicated poorer outcomes. Additional parameters (TT, Cl(-), CER and hepatic imaging features) also contributed modestly to predictions. CONCLUSIONS: The ML-based model effectively predicts WD-ACLF risk, with XGBoost demonstrating superior performance. TBA, APTT, diagnosis age, onset age and Hb emerged as critical biomarkers, offering actionable insights for early clinical intervention.