AI-driven prediction of post-transplant survival and stratification of HCC recurrence risk.
Year: 2025
Type: conference-abstract
Abstract: 4114 Background: Traditional selection criteria (e.g., Milan, UCSF) and scoring systems (MELD/PELD) for liver transplant eligibility in hepatocellular carcinoma (HCC) often fail to capture the complex interplay of tumor biology, patient factors, and bridging therapies. Although Milan and UCSF yield similar 1-, 3-, and 5-year survival rates, the debate over expanded criteria highlights the need for refined, individualized risk-stratification tools. Methods: We retrospectively analyzed 21,182 HCC patients from the UNOS database to develop deep learning and Cox regression models for overall survival (OS). Models incorporated demographic (e.g., age, race), clinical (e.g., diabetes, MELD differences), and tumor-specific variables (e.g., tumor count, size tiers). Performance was compared to standard MELD-based calculations using 5-fold cross-validation, with primary endpoints of 1-, 3-, and 5-year survival. For recurrence risk, we used 15,801 records to train gradient boosting (XGBoost) and Cox models. Key variables included tumor characteristics (size levels, vascular invasion), recipient factors (insurance type, functional status, initial MELD/PELD), and alpha-fetoprotein (when available). Model performance was evaluated via area under the curve (AUC) and concordance index (c-index); external validation was performed for the recurrence model. Results: Cox Regression (time-to-event): Final multivariable models achieved c-indices of 0.611 for OS and 0.601 for progression-free survival (PFS). Stepwise Logistic Regression (mortality): Mean AUCs were 0.664 (1-year), 0.705 (3-year), and 0.758 (5-year). Random Forest Classifier: Slightly higher AUCs than logistic regression (0.663 at 1-year, 0.714 at 3-year, 0.762 at 5-year). Gradient Boosting (recurrence): 1-year recurrence predictions achieved AUC > 0.80, with microvascular invasion emerging as a key risk factor (p<0.001). Across approaches, incorporating multiple clinical and tumor-specific factors outperformed MELD-based models, consistently showing improved predictive accuracy. Conclusions: Machine learning–based models, including deep learning, random forests, and gradient boosting, offer enhanced risk prediction for post-transplant survival and HCC recurrence beyond traditional scoring criteria. These advanced tools enable more nuanced transplant selection, surveillance, and early intervention strategies, potentially improving long-term outcomes for HCC patients undergoing liver transplantation.
Source: Journal of Clinical Oncology
Authors: Junho Song, Jihyun Kim, CholMin Kang, Chaehyeon Kim, Won Jong Yang, Donghoon Shin, Jongwoo Kim, Hyungjune KU, Amy Choi, Hyung Hwan Moon
Language: English
Subfield: Hepatology
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4410815401