Risk Prediction of Early-Onset Hepatocellular Carcinoma: Derivation and Validation in a Nationwide Young Adult Cohort

The global incidence of early-onset hepatocellular carcinoma (eHCC) is increasing significantly; however, specific risk prediction tools for young adults remain scarce. This study aimed to develop and validate predictive models for eHCC using both conventional statistical and machine learning techniques.

We included 1,756,593 young adults aged 20-39 years who underwent health screenings in South Korea between 2013 and 2014 with follow-up until January 31, 2022. Participants were randomly divided into training and validation sets (1:1 ratio). Prediction models were constructed using cause-specific Cox proportional hazards regression, Fine-Gray subdistribution hazards regression, generalized boosted model (GBM), and generalized boosted survival model (GBM-survival).

Key predictors of eHCC in the training cohort included viral hepatitis, history of non-HCC cancer, liver cirrhosis, and gamma-glutamyl transferase. In the validation cohort, GBM-survival demonstrated best predictive performance for eHCC in young adults, with area under the receiver operating characteristic curves (auROCs) of 0.945 (1-year) and 0.825 (5-year). The most significant features in GBM-survival were aspartate aminotransferase level, history of non-HCC cancer, viral hepatitis, gamma-glutamyl transferase, and serum creatinine. A global surrogate model for the GBM-survival model enhanced interpretability (RMSE = 0.287; Pearson r = 0.887). Using the K-group method, GBM-survival achieved risk ratios of 2092.3 and 166.2 for 1-year and 5-year eHCC prediction, respectively.

We developed and validated robust risk prediction models for eHCC that integrated established risk factors with emerging metabolic indicators. This study provides actionable tools for risk stratification in young adults to address the growing burden of eHCC.