Explainable Dynamic Ensemble Framework for Classification Based on the Late Fusion of Heterogeneous Multimodal Data

Abstract

Ensuring precise and reliable classification effectiveness holds paramount importance in essential sectors such as medicine, industry, and healthcare. Machine learning (ML) techniques have evolved in recent years to address the performance, efficiency, and robustness of the applied models. Recent advancements in Machine Learning (ML) techniques have aimed to enhance the performance, efficiency, and robustness of applied models. Ensemble learning has been shown to exhibit superior accuracy, robustness, and generalization capability over classical single ML models in most classification problems. Although dynamic ensembles have extended the performance of static ensembles such as random forest and boosting, current literature on dynamic ensembles primarily focuses on the early fusion of multimodal data. In this study, we present a novel framework that combines dynamic ensemble selection (DES) with a late fusion of heterogeneous multimodal data and model explainability. We evaluated our approach on a classification task of hospital mortality prediction, and our approach achieves a testing accuracy of 90.16%, surpassing existing techniques and providing physicians with case-based reasoning and deep-based classifiers contributions explanations to support their decision-making. We compare our proposed framework against nine widely used ML techniques, including static and dynamic ensemble models with early fusion and static ensemble models with late fusion on a dataset of 6,600 patients from MIT’s GOSSIS dataset. The dynamic ensemble model with early fusion achieves a testing accuracy of 86.89%, the LightGBM model achieves a test accuracy of 87.72%, and the soft voting model reaches 87.97% and 89.45% using early and late fusion, respectively. Our proposed framework not only improves the accuracy and robustness of in-hospital mortality prediction models but also offers explainability and potential for further optimization to achieve even higher performance.