Machine learning improves post-transplantation HCC recurrence prediction

We aimed to enhance post-transplantation HCC recurrence prediction by evaluating additional novel risk factors and leveraging state-of-the-art machine learning (ML) algorithms. Using the United Network for Organ Sharing (UNOS) database, we identified adult HCC patients who underwent liver transplantation (LT) 2015-2018 and considered >50 available clinical, radiographic, laboratory/biomarker, and explant pathology variables to predict post-transplantation recurrence-free survival. The cohort was split 70:30 into training and test datasets. Recursive feature elimination was employed to select an optimal number of variables for each candidate ML model. Final model performance was compared to clinically used tools with the test dataset. Of the 3106 patients identified, 7.2% developed post-transplantation HCC recurrence. The Gradient Boosting Survival algorithm performed best (C-index 0.73) and included 7 variables: explant tumor burden score (TBS), alpha fetoprotein (AFP) at transplantation, maximum pre-transplantation TBS, pre-transplantation AFP slope, microvascular invasion on explant, poor tumor differentiation on explant, and change in pre-transplantation TBS normalized by the number of locoregional therapies received. This outperformed the Risk Estimation of Tumor REcurrence After Transplant (RETREAT) Score (C-Index 0.70). A Random Survival Forest model including only preoperative variables [AFP at transplantation, pre-transplantation AFP slope, change in AFP from listing to transplantation, maximum pre-transplantation TBS, and Albumin-Bilirubin (ALBI) Grade change from listing to transplantation] was also able to predict post-LT HCC recurrence (C-Index 0.69). In summary, we developed a novel ML model that outperforms a widely used post-transplantation HCC recurrence risk score. This model may be used to better risk-stratify patients following transplantation and tailor surveillance/adjuvant therapy. The pre-transplantation ML model may be used with the Milan Criteria to further risk-stratify patients being considered for transplantation.