Metaheuristic Algorithms based Ensemble Models for Enhanced Chronic Disease Diagnosis

Abstract

Disease diagnosis is a fundamental aspect of modern healthcare, where accurate and timely detection can profoundly impact patient outcomes. Leveraging the power of en semble techniques has emerged as a promising avenue to enhance diagnostic accuracy. However, the intricate landscape of disease data presents challenges that necessitate inno vative solutions. Ensemble methods, which combine the predictions of multiple models, offer a means to improve disease diagnosis by leveraging diverse perspectives. Yet, ef fectively harnessing these techniques remains challenging due to class imbalance within datasets and the intricate task of configuring optimal ensembles. This thesis embarks on a comprehensive exploration, addressing these challenges through a meticulously designed sequence of objectives. In this thesis first, we intro duced a three-level stacking approach that integrates the Adaptive Synthetic Sampling (ADASYN) technique to handle class imbalance, while Particle Swarm Optimization (PSO) fine-tunes Support Vector Machine (SVM) meta-model. The resulting ensemble ex hibits exceptional performance across key metrics, including AUC, accuracy, specificity, and precision. Building on this foundation, we delve into diversity-based ensemble frameworks. In our second objective, address the challenges of diversity based classifier selection. To achieve this proposed a novel fitness function that enhances the diversity of base learners within the ensemble. By combining this with bootstrapped bags and cross-validation, we demonstrate its superiority over existing ensemble models, reinforcing the potential of diversity-driven strategies. Continuing our exploration, the third objective introduces the Bagging Approach with Teaching-Learning-Based Optimization (BA-TLBO). This dynamic ensemble optimiza tion technique strikes a delicate balance between accuracy and diversity through dynamic iv weight updation and bag size adjustments. The approach’s ability to maintain exploration while optimizing exploitation is validated through rigorous experimentation, positioning it as a robust alternative to traditional ensemble methods. Ourfinalobjective takes on the complex challenge of classifier selection and placement within an ensemble framework. We navigate the intricate landscape of classifier config urations through a dynamic three-level ensemble framework guided by a nested Genetic Algorithm (GA) and an innovative fitness function. The approach’s remarkable outcomes further underscore its potential for accurate disease diagnosis. In summary, this thesis unveils a strategic sequence of ensemble techniques that ef fectively address challenges in disease diagnosis. By systematically advancing from class imbalance solutions to diversity-driven strategies and sophisticated ensemble optimiza tion, it promises enhanced diagnostic accuracy and ultimately improved patient care