Abstract
BACKGROUND: Due to organ shortage, liver transplantation (LT) in hepatocellular carcinoma (HCC) patients can only be offered subsidiary to other curative treatments, including liver resection (LR). We aimed at developing and validating a machine-learning algorithm (ML) to predict which patients are sufficiently treated by LR. METHODS: Twenty-six preoperatively available routine laboratory values along with standard clinical-pathological parameters [including the modified Glascow Prognostic Score (mGPS), the Kings Score (KS) and the Model of Endstage Liver Disease (MELD)] were retrieved from 181 patients who underwent partial LR due to HCC in non-cirrhosis or compensated cirrhosis from January 2007 through March 2018 at our institution. These data were processed using a Random Forest (RF)-based workflow, which included preprocessing, recursive feature elimination (RFE), resampling, training and cross-validation of the RF model. A subset of untouched patient data was used as a test cohort. Basing on the RF prediction, test data could be stratified according to high (HR) or low risk (LR) profile characteristics. RESULTS: RFE analysis provided 6 relevant outcome predictors: mGPS, aPTT, CRP, largest tumor size, number of lesions and age at time of operation. After down-sampling, the predictive value of our model was 0.788 (0.658-0.919) for early DFS. 16.7% of HR and 74.2% of LR patients survived 2 years of follow-up (P<0.001). CONCLUSIONS: Our RF model, based solely on clinical parameters, proved to be a powerful predictor of DFS. These results warrant a prospective study to improve the model for selection of suitable candidates for LR as alternative to transplantation. The predictive model is available online: tiny.cc/hcc_model.