Background: Hepatocellular carcinoma (#HCC ) is the most common primary liver tumor. Despite efforts to mitigate risk factors and implement surveillance programs in high-risk populations, such as screening, these strategies alone appear insufficient to significantly improve prognosis at diagnosis. The identification of novel prognostic factors remains an underdeveloped field that may play a key role in guiding optimal therapeutic decisions from the initial stages of patient management.
Aims: To develop a machine-learning prognostic model to compare the prognostic performance of different MELD-based scores at the time of HCC diagnosis and to assess their relative clinical applicability in comparison with established prognostic staging systems.
Methods: A multicenter retrospective analysis including 219 patients with HCC was performed. For MELD-based score comparisons and model development, 216 patients with complete MELD, MELD-Na, and MELD 3.0 data constituted the analytic cohort. Clinical and diagnostic variables were analyzed using machine-learning approaches.
Results: In the analytic cohort, 148 all-cause deaths occurred during follow-up. Among the MELD-derived models, MELD 3.0 showed higher discrimination than MELD and MELD-Na. EXtreme Gradient Boosting (XGB) algorithm achieved the best overall performance and calibration (AUC 0.94, Brier score 0.13, calibration slope 1.02, CITL 0.03). A parsimonious reduced-feature XGB model including TNM stage, MELD 3.0, ECOG-PS, ALP, and AFP retained most of the discriminatory performance of the full model (AUC 0.91).
Conclusions: These findings suggest that updated MELD-based scores, particularly MELD 3.0, may provide complementary prognostic information at the time of HCC diagnosis. The XGB-based model may represent a feasible tool for exploratory prognostic modeling and may support more precise risk stratification and personalized, data-driven therapeutic decisions in patients with HCC. Further validation in larger, prospective cohorts is warranted before clinical implementation.
https://link.springer.com/article/10.1007/s10620-026-10032-6
2 Medical News 