Improved Ensemble-Based Approaches with Stacking for Imbalanced Medical Data Classification

Abstract

Classimbalance is a prevalent challenge in medical databases, wherein minority pathology classes are under-represented, resulting in biased classification models that favor the majority class. In this study, we evaluate the efficacy of prominent ensemble-based methods (Balanced Bagging, Balanced Random Forest, RUSBoost, XGBoost, and EasyEnsemble) designed to address the challenge of imbalanced data without generating synthetic samples. Through experimentation on eight datasets with varying degrees of imbalance, no single method demonstrated consistent superiority across all cases. To address this, we propose a stacking meta-predictor that combines the outputs of these models, leveraging their strengths to create a more robust classification system. The stacked model demonstrates improved generalization and superior classification performance across diverse imbalanced medical datasets.