An XGBoost-Based Multicenter Model for Predicting HBV-Related Hepatocellular Carcinoma: Development and Validation

Area of Science:

• Hepatology
• Oncology
• Machine Learning in Medicine

Background:

• Hepatocellular carcinoma (HCC) survival rates are stage-dependent, but current prediction models struggle with accuracy for Hepatitis B virus-associated HCC (HBV-HCC).
• Accurate risk stratification is crucial for timely intervention and improved outcomes in HBV-HCC patients.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for HBV-HCC risk stratification.
• To integrate multidimensional biomarkers for enhanced predictive accuracy.

Main Methods:

• A retrospective multicenter study involving 3568 participants (1872 HBV-infected, 1696 HBV-HCC).
• Identification of five key predictors (log10DCP, log10HBVDNA, log10ALT, AFP-L3%, log10AFP) using random forest, LASSO, and XGBoost.
• Evaluation of seven ML models using AUC, sensitivity, specificity, accuracy, and F1-score, with comparison to existing models.

Main Results:

• The XGBoost model demonstrated high performance with AUCs of 0.985 (training), 0.978 (validation), and 0.942 (external validation).
• XGBoost significantly outperformed previous models (GALAD, ASAP) in accuracy and individualized risk prediction.
• Key predictors identified include log10DCP, log10HBVDNA, log10ALT, AFP-L3%, and log10AFP.

Conclusions:

• A novel, highly accurate diagnostic model for HBV-HCC was developed using machine learning.
• The model offers superior risk stratification compared to existing methods, facilitating clinical implementation via a web tool.