LAUSR

Hepatocellular Carcinoma (HCC) is the most common type of liver cancer which accounts for around 75% of all liver cancer cases. From statistical data, it has been found that fatality due to liver cancer is higher regardless of improved screening and discoveries in medicines, HCC escalate fatality rate. This paper presents an ensemble learning model for HCC survival prediction. The input predictors for the proposed model consist of geographical information, risk factors and clinical trial information of HCC patients. Fifteen different models are presented to evaluate the prediction. These models present data pre-processing, feature reduction/elimination and survival classification phase. For feature evaluation, LASSO Regression (L-1 penalization), Ridge Regression (L-2 penalization), Genetic Algorithm (GA) Optimization and Random Forest (RF) are proposed for weight valuation of features wherein features with significant weights are selected for prediction. With the aid of feature evaluators, L-1 penalized Nu-Support Vector Classification (Nu-SVC) model, L-2 penalized Nu-SVC model, GA optimized Nu-SVC model, RF-NuSVC model, L-1 penalized RidgeCV (RCV) model, L-2 penalized RCV model, GA optimized RCV model, RF-RCV model, L-1 penalized Gradient Boosting Ensemble Learning (GBEL) model, L-2 penalized GBEL model, GA optimized GBEL model and RFGBEL model are presented for survival prediction. The prediction performances of models were measured in terms of accuracy, recall/sensitivity, F-1 score, Log-Loss score, Jaccard score and Area Under Receiver Operating Curves (AUROC). The results indicate that RFGBEL model shows excellent performance in contrast to other proposed models. The proposed RFGBEL model achieves an accuracy of 93.92%, sensitivity of 94.73%, F-1 score of 0.93, Log-Loss/Cross entropy score of 5.89 and Jaccard score of 0.72. RFGBEL estimates value of area under the curve as 0.932. Comparison of RFGBEL model with other existing state of the art models are presented for performance assessment. Overall, the RFGBEL model has a capability to predict the result with more accuracy and sensitivity by means of machine learning and data mining approach.