Hybrid Packet Learning Approaches for Early Diabetes Detection and Optimization of Therapeutic Strategies: A Comparative Analysis of Stack, Replacement and Reinforcement Models

Abstract

This thesis focuses on the comparative analysis of ensemble learning methods, including bagging, boosting, and stacking, in the context of early detection of diabetic retinopathy and personalized treatment. The main objective is to propose an intelligent system capable of estimating the risk of retinopathy based on a patient's clinical data and recommending a treatment tailored to their profile. Initially, the theoretical foundations of each approach are presented, along with their strengths and limitations. Particular attention is also paid to the reference model, used to evaluate the actual contribution of the ensemble learning methods. An empirical study is then conducted on a synthetic dataset representing 2,597 patients. The performance of the models, including the reference model and ensemble methods, are evaluated using several metrics (accuracy, recall, F1 score, AUC-ROC), and a comparative analysis is performed to highlight the best performing approach. The results show that combined (hybrid) methods offer better generalization capabilities and superior performance in detection and personalized recommendation compared to the basic model. This thesis thus highlights the relevance of ensemble learning in sensitive medical applications, particularly for data-driven therapeutic decisions.