Use of machine learning models for identification of predictors of survival and tumour recurrence in liver transplant recipients with hepatocellular carcinoma.

Background: Hepatocellular carcinoma (HCC) is one of the leading indications for liver transplantation (LT) however, selection criteria remain controversial. We aimed to identify survival factors and predictors for tumour recurrence using machine learning (ML) methods. We also compared ML models to the Cox regression model. Methods: Thirty pretransplant donor and recipient general and tumour specific parameters were analysed from 170 patients who underwent orthotopic liver transplantation for HCC between March 2013 and December 2019 at the University Hospital Merkur, Zagreb. Survival rates were calculated using the Kaplan-Meier method and multivariate analysis was performed using the Cox proportional hazards regression model. Data was also processed through Coxnet (a regularized Cox regression model), Random Survival Forest (RSF), Survival Support Vector Machine (SVM) and Survival Gradient Boosting models, which included pre-processing, variable selection, imputation of missing data, training and cross-validation of the models. The cross-validated concordance index (CI) was used as an evaluation metric and to determine the best performing model. Results: Kaplan-Meier curves for 5-year survival time showed survival probability of 80% for recipient survival and 82% for graft survival. The 5-year HCC recurrence was observed in 19% of patients. The best predictive accuracy was observed in the RSF model with CI of 0.72, followed by the Survival SVM model (CI 0.70). Overall ML models outperform the Cox regression model with respect to their limitations. Random Forest analysis provided several relevant outcome predictors: alpha fetoprotein (AFP), donor C-reactive protein (CRP), recipient age and neutrophil to lymphocyte ratio (NLR). Cox multivariate analysis showed similarities with RSF models in identifying detrimental variables. Some variables such as donor age and number of transarterial chemoembolization treatments (TACE) were pointed out, but these were not influential in our RSF model. Conclusions: Using ML methods in addition to classical statistical analysis, it is possible to develop sufficient prognostic models, which, compared to established risk scores, could help us quantify survival probability and make changes in organ utilization.