Abstract
BACKGROUND AND AIMS: Chronic hepatitis B poses a major health risk, especially its progression to decompensated cirrhosis. Early prediction is crucial for better outcomes. This study evaluated the predictive power of Golgi protein 73 (GP73), α1-microglobulin (α1M), age, aspartate aminotransferase (AST), alanine aminotransferase (ALT), and platelet count (PLT) using machine learning models. METHODS: A total of 179 patients (69 healthy controls, 59 with decompensated cirrhosis, and 51 with nondecompensated liver disease) were analyzed. Five random forest models incorporating different combinations of variables, including GP73, α1M, age, AST, ALT, PLT, aspartate aminotransferase-to-platelet ratio index (APRI), and fibrosis-4 (FIB-4) index were assessed using area under the curve (AUC) and accuracy. Logistic regression and a decision tree were also employed. RESULTS: Random forest model 3 (age + GP73 + α1M + AST + ALT + PLT) achieved the highest AUC (0.96) and accuracy (0.90), outperforming model 4 (age + APRI + FIB-4 + GP73 + α1M, AUC: 0.96, accuracy: 0.76), and logistic regression (AUC: 0.91, accuracy: 0.86). GP73 and PLT were the most significant predictors of cirrhosis progression. There were nonlinear interactions between GP73 and α1M. When PLT levels were ≤143 × 10 9 /L, patients with GP73 > 0.168 ng/L or ≤ 0.117 ng/L indicated an increased risk of decompensated cirrhosis. CONCLUSION: The combination of GP73, α1M, age, AST, ALT, and PLT enhances prediction accuracy for the progression from nondecompensated to decompensated hepatitis B virus-related cirrhosis, with GP73 and α1M showing nonlinear interactions influenced by PLT levels.