Study proposes an explainable machine-learning framework for mortality risk in liver-transplant candidates with HCC

Researchers from Télécom SudParis at Institut Polytechnique de Paris, AP-HP Hôpital Paul Brousse, and Université Paris-Saclay have published a study in Health Data Science proposing an explainable machine-learning framework for mortality-risk assessment in liver-transplant candidates with hepatocellular carcinoma (HCC). The approach combines ensemble learning, SHapley Additive exPlanations (SHAP), and supervised clustering to address the dual risks of liver dysfunction and tumor progression.

Risk assessment while awaiting transplantation is challenging for people with HCC. Traditional scores may reflect liver-disease severity or tumor-related risk separately but may not integrate both risks and their interactions. The team therefore explored a data-driven approach designed to offer both population-level and local explanations.

The study used a LightGBM model and selected eight relevant variables through SHAP analysis. The authors report that the system outperformed the conventional scores compared in the article. They then embedded SHAP values in a Uniform Manifold Approximation and Projection space and applied supervised clustering, identifying seven subgroups with increasing mortality-risk levels and characterizing the contribution of risk factors across subgroups.

The framework offers a methodological approach to risk research and interpretable stratification for liver-transplant candidates with HCC and may support further study of patient progression. Model comparisons and clusters do not demonstrate improved treatment decisions or patient outcomes; external validation, clinical-workflow evaluation, and prospective studies remain necessary before translational use.