Answers are grounded in this paper and cite their sources.
Introduction: Hepatocellular carcinoma (HCC) remains a major cause of cancer-related mortality, highlighting the need for more accurate risk stratification approaches. Objective: To develop and validate an integrated clinical risk scoring (CRS) and machine learning (ML) system to accurately identify HCC, overcoming the limitations of traditional alpha-fetoprotein (AFP) screening, especially in AFP-negative individuals. Methods: A retrospective cross-sectional diagnostic study utilized data from two institutions (development cohort n = 5,000; external validation cohort n = 443) with 12 routine clinical variables. Multivariable logistic regression was used to construct the CRS, and ensemble ML models were developed. Performance was assessed via area under the receiver operating characteristic curve (AUC), quantitative calibration measures, decision curve analysis (DCA), net reclassification improvement (NRI), and integrated discrimination improvement (IDI). We applied development-trained models directly to the independent cohort for strict external validation. Results: Hepatitis C, diabetes and hepatitis B were identified as the strongest independent risk factors for HCC. In the development cohort, the CatBoost model demonstrated the best overall discriminatory ability (AUC = 0.9773), significantly outperforming the single AFP indicator (0.6377) and CRS (0.9385); notably, in the independent validation cohort, CRS demonstrated superior stability in terms of AUC (0.9849) and risk reclassification ability (NRI/IDI), performing slightly better than CatBoost (AUC = 0.9816). DCA, NRI, and IDI confirmed significant incremental clinical net benefit and risk reclassification over AFP-only screening (p < 0.001). Conclusion: This integrated HCC risk stratification system effectively overcomes traditional AFP screening limitations, particularly in AFP-negative populations. The observed bias-variance trade-off supports a tiered deployment strategy: CRS for robust screening in resource-limited settings, and ML for precision assessment in well-resourced centers.
No takes yet. Share an insight, caveat, or question.
Luo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: