Validation of Risk Models for Predicting Post-SVR HCC in Real-World Surveillance Across Global Geographic Regions.

Background & Aims Several clinical risk models have been proposed to stratify hepatocellular carcinoma (HCC) risk in patients with chronic hepatitis C virus (HCV) after sustained virologic response (SVR). However, validation efforts have focused on monocentric or country-specific cohorts, and it is unclear if clinical risk models can be broadly applied to global populations. We characterised regional variation in model performance for HCC risk stratification in post-SVR patients.Methods Four HCC clinical risk models (aMAP score, FIB-4 index, GES score, and Toronto HCC risk index [THRI]) were analysed in six real-world cohorts, which included 8796 post-SVR patients from different geographic regions globally. Model discrimination was assessed using Harrel's c-statistic index. HCC incidence rates were compared across low-, intermediate-, and high-risk groups for each model.Results Distributions of patient characteristics and HCC incidence rates varied across geographic regions. Predictive performances of models were comparable within each cohort despite the model with the highest c-statistics differing by regions. Performance was lower than those from original reports overall; c-statistics of models across most regions remained below 0.70.Conclusions There remains a continued need to improve discrimination and calibration of clinical models to stratify HCC risk in post-SVR patients. Accuracy of models may differ by geographic region, underscoring the importance of external validation to assess transportability of models and suggesting no single model can be universally applied.The risk of hepatocellular carcinoma (HCC) remains elevated in patients with hepatitis C virus (HCV) even after the eradication of HCV. Stratifying the risk of HCC development in patients with cured HCV is crucial. There have been several models for predicting HCC development based on routine clinical variables. However, the predictive performances of all four models were suboptimal in real‐world clinical settings with large variations across global geographic regions.