Accurate Liver Disease Prediction using a Hierarchical Two-Level Ensemble Stacking Model

Abstract

Liver disease is a serious health problem prevalent in the world. Early diagnosis is very important for enhancing quality of patient's health and reducing mortality. Traditional diagnostic techniques tend to take quite long time, be expensive and rely heavily on the skill of the medical personnel. That's why intelligence systems are needed that can make decisions based on data. This study suggests effective framework based on machine learning which allow to make accurate prediction about the liver disease. The methodology includes data preprocessing, which involves such steps as encoding categorical data, filtering out irrelevant information, as well as dealing with missing data. Then Random Forest and Recursive Feature Elimination can be used for feature selection. We proposed a framework to evaluate the performance of various machine learning classifiers such as Logistic Regression, K-Nearest Neighbor (KNN), Decision Tree, Linear Discriminant Analysis (LDA), Multilayer Perceptron (MLP) and many ensemble learning models including Random Forest, Extra Trees, Gradient Boosting, AdaBoost, and Extreme Gradient Boosting (XGBoost). To increase the prediction accuracy, we have developed a two-level ensemble stacking model with multiple ensemble classifiers as base learners and an advanced meta-learner to make predictions. The models have been assessed with accuracy, precision, recall, specificity, F1 score, Area Under the Receiver Operating Characteristic Curve (AUC-ROC), confusion matrix and Cross validation. The experiments conducted showed that the proposed two-level ensemble stacking framework is better compared to both individual classifiers as well as conventional ensemble techniques since it attained an accuracy close to 94%.