Abstract
ABSTRACT
Background & Aims
Existing hepatocellular carcinoma (HCC) risk scores specific to chronic hepatitis B (CHB) mostly rely solely on baseline indicators, limiting their applicability in clinical practice. By incorporating dynamic and static indicators, this study aimed to address this limitation by developing a machine learning (ML) model for HCC prediction in patients with CHB and compensated advanced chronic liver disease (cACLD).
Methods
ML models were trained to predict HCC in 1 272 patients with CHB and cACLD, who underwent at least 2 measurements from different centres and were further tested with external cohorts (
n
= 601). The first and last liver stiffness measurements were defined as the baseline and follow‐up values, respectively.
Results
A total of 1 873 individuals with CHB and cACLD were included, with 192 patients (10.3%) developing HCC after a median of 30.7 months. Based on LASSO logistic regression, baseline age, baseline albumin concentration, follow‐up albumin concentration, follow‐up platelet count, and change in the liver stiffness measurement (∆LSM) were selected, with ∆LSM identified as the most important indicator. Developed using XGBoost, the ALPS‐HCC score achieved the highest integrated AUROC (0.852), Harrell's C‐index (0.852), and F1 score (0.944), and performed significantly better than existing HCC risk scores. Based on these results, our risk stratification performed well in terms of prognosis in the validation cohort, including the identification of low‐risk (
n
= 250; reference), intermediate‐risk (
n
= 91; hazard ratio [HR] = 5.25), and high‐risk (
n
= 41; HR = 24.47) groups.
Conclusions
ALPS‐HCC score provides robust predictive power and screening capability for predicting HCC risk in patients with CHB and cACLD.