Explainable Mortality Prediction for Liver Transplant Candidates with Hepatocellular Carcinoma: A Supervised Clustering Approach

Background: Accurate mortality prediction for liver transplant candidates with hepatocellular carcinoma (HCC) remains a critical challenge. Traditional scoring systems, including Child–Pugh, Albumin–Bilirubin, Model for End-Stage Liver Disease (MELD), MELD-Na, MELD 3.0, and Alpha-fetoprotein scores, are widely used but often fail to provide precise risk assessments. This limitation arises from the dual burden of liver dysfunction and tumor progression, which complicates prognosis. Consequently, there is a need for a comprehensive approach addressing both considerations to better manage HCC patients. Methods: We propose an advanced machine learning-based scoring system exploiting Ensemble Learning and SHapley Additive exPlanations (SHAP) for a better understanding of key mortality risk factors. SHAP offers valuable insights into the decision-making process by providing both global and local explanations. By embedding SHAP values in the Uniform Manifold Approximation and Projection space, we perform supervised clustering to infer latent subgroups, providing a higher granularity on the contribution of key variables for mortality risk assessment. Results: Our system based on LightGBM outperforms conventional scores leveraging only 8 relevant variables selected by SHAP analysis. These variables respond to the challenging dual risk problem set in this work. With supervised clustering, we uncover 7 subgroups showing an increasing mortality risk level and a fine assessment of risk factors’ contribution. Conclusion: By contrast to existing studies, our approach offers an integrative data-driven framework for handling the dual risk challenge set by HCC patients with liver dysfunction. Also, it provides a valuable tool for a more precise risk evaluation that may guide treatment decisions and help monitoring patient progression.