Artificial intelligence-based prediction of hepatocellular carcinoma in individuals without HBV or HCV infection in the MASLD era

Abstract

Abstract Current surveillance guidelines for hepatocellular carcinoma (HCC) primarily target individuals with liver cirrhosis or viral hepatitis. We developed an artificial intelligence (AI)-based model to predict HCC risk in individuals without viral hepatitis. Using the South Korean National Health Insurance Service database, we analyzed 7,961,075 individuals aged 20–79 years who underwent health screening in 2017 and were followed through 2023. Twelve AI models were compared. The best-performing model was retrained and evaluated in an unmatched validation cohort reflecting the original HCC prevalence and an independent non-selected validation cohort. Multi-Head Attention with Deep Neural Network (MHA_DNN) achieved the highest performance (AUC 0.9385, 95% confidence interval [CI]: 0.9312–0.9455 In the unmatched validation cohort, the model demonstrated excellent discrimination (AUROC 0.9450; AUPRC 0.0503), improved calibration after logistic recalibration, and superior clinical utility. Performance remained consistent in the independent non-selected validation cohort (AUROC 0.8848; AUPRC 0.0293). SHAP analysis identified fatty liver index (FLI) and triglycerides (TG) as the most influential predictors. TG increased HCC risk in individuals with FLI < 60 but showed the opposite association in those with FLI ≥ 60. MHA_DNN provides robust population-level prediction of HCC in individuals without viral hepatitis and supports personalized surveillance based on metabolic risk factors.