Abstract No. 527 A random forest model for predicting survival of HCC patients treated with radioembolization

Abstract

To generate a prognostic model capable of predicting survival outcomes of hepatocellular carcinoma (HCC) patients after radioembolization (Y90) using machine learning methods. With IRB approval, we included baseline characteristics and overall survival (OS) data of 519 HCC patients who had Y90 between 2004 and 2017. Inclusion criteria included patients who: a) had Y90 for HCC; b) no subsequent surgical intervention (liver transplantation or resection) and c) reached death endpoint. Each patient was given a label of “1” (> median OS) or “0” (≤ median OS). Baseline data was transformed using a factor analysis (FA) model with 13 latent components. Cases were then randomly split into training and testing subsets (95 and 5% of cohort, respectively). A random forest classifier model was then trained using the training cases. Model classification performance was evaluated on the test set by measuring the area under the curve (AUC) of the receiver operator characteristics (ROC). The optimum model threshold was determined based on the ROC curve point which optimized sensitivity and specificity. The median OS was 8.3 months. A random forest classifier model was generated. The model performance on the test set (5% of the total dataset, 26 patients) had an AUC of 0.92 for distinguishing between patients with OS greater or less than the median (8.3 months). The positive predictive value of the model for predicting survival >8.3 months was 0.81 and the negative predictive value was 0.91. The optimum model threshold was determined to be 0.49 (range: 0.0-1.0). At the optimum model threshold, the sensitivity was 0.85 and the specificity was 0.92. Machine learning models (both unsupervised and supervised techniques) can be used to develop accurate predictive models to estimate post-treatment HCC survival. When compared to BCLC (AUC for prognosticating survival >1 year has been reported at 0.71-0.75), the AUC of our method for prognosticating survival >8.3 months was 0.92. While the model in its current state is rudimentary, a model with superior ROCs for prognosticating HCC survival could be developed with more detailed imaging/clinical data.