A META-LEARNING–DRIVEN HYBRID STACKING ENSEMBLE FOR ROBUST MULTI-CLASS PREDICTION OF LIVER CIRRHOSIS STATUS

Abstract

A successful clinical decision-making and patient management is highly dependent on the early and accurate prediction of the status-based of liver cirrhosis. This paper introduces a sophisticated meta-learning-based hybrid stacking model of the multi-class classification of liver cirrhosis, with the use of CatBoost, LightGBM, and Extra Trees as base learners and effective meta-learners. It used a massive extent of preprocessing pipeline with powerful scaling, K-nearest neighbor imputation, and one-hot encoding to boost the quality of data and model generalization. A number of hybrid ensemble methods such as stacking with other meta-learners and soft voting methods were systematically tested on a large-scale liver cirrhosis dataset. The experimental results show that LightGBM-based stacking model with the accuracy of 0.9922, the weighted F1-score of 0.992, and ROC-AUC of 0.9988 was the most successful in the hybrid configurations. The strong value of Cohen-Kappa and Matthews correlation coefficient indicates the strength and reliability of the proposed framework. These results affirm that meta-learning-based hybrid ensembles are an effective approach to model complex nonlinear correlations between clinical characteristics and provide a saleable and high-performance approach to predict status based of liver cirrhosis, and that they have a high potential of being integrated into clinical decision support systems.