ML Predicts Dual Mortality Risk for HCC Liver Transplant Candidates (11,647 Cases)

Hepatocellular carcinoma (HCC) accounts for 70‑90% of primary liver cancers and often requires liver transplantation; however, candidates face simultaneous risks of liver failure and tumor progression, making accurate wait‑list mortality prediction critical.

The study assembled 11,647 adult HCC transplant candidates from the OPTN/UNOS STAR registry (2002‑2023), dividing them into an "on waiting list" group (11,199) and a "waitlist mortality" group (448). In addition to 25 static clinical variables, six dynamic laboratory differences (serum sodium, creatinine, albumin, bilirubin, AFP, INR) were computed, yielding a total of 31 features.

Three tree‑based ensemble models—Random Forest, XGBoost and LightGBM—were trained under two feature sets (static only and static + dynamic). To address severe class imbalance, the majority class was down‑sampled to match the minority class, and 3‑fold cross‑validation was applied. Hyper‑parameters were tuned via grid search. SHAP values were calculated for global and local interpretability, then embedded into a UMAP space and clustered with K‑medoids (supervised clustering) to reveal patient sub‑groups.

Compared with eight conventional risk scores (ALBI, Child‑Pugh, AFP, HALT‑HCC, Mehta Model, MELD, MELD‑Na, MELD 3.0), the Mehta Model achieved the best traditional AUROC (0.782) followed by HALT‑HCC (0.763). Among ensemble models, Random Forest on static features reached AUROC 0.796, while LightGBM on the full 31‑feature set achieved AUROC 0.826. Feature selection using SHAP global importance identified eight key variables; LightGBM trained on this reduced set attained AUROC 0.835, sensitivity 77.14% and specificity 75.64%, surpassing the full‑feature model.

The eight‑feature model was packaged as the ELM‑HCC score, highlighting the importance of the dynamic AFP difference (AFP_DIFF) for mortality prediction.

Embedding SHAP values into UMAP and clustering with K‑medoids produced seven distinct risk clusters. Mortality probability rose monotonically from 0.03 in cluster 1 to 0.98 in cluster 7, a trend confirmed by Kruskal‑Wallis tests. Clusters were characterized by severe liver dysfunction (high bilirubin, creatinine, ascites) and aggressive tumor activity (high AFP).

Overall, the LightGBM + SHAP framework (ELM‑HCC) delivers significantly higher predictive performance and explainability than existing scores, and the supervised clustering offers clinically actionable sub‑group stratification, filling a previously unmet need for dual‑risk assessment in HCC transplant candidates (Health Data Science, 2024).