Développement d’un modèle de prédiction et de diagnostic médical assisté par l'intelligence artificielle.

Abstract

Diabetes is currently a major global public health issue, affecting hundreds of millions of people. Early and effective detection is essential to reduce the serious complications associated with this chronic disease. In this context, artificial intelligence (AI), particularly machine learning and deep learning, offers promising solutions to predict the risk of diabetes based on medical data.This thesis aims to design, implement, and compare various intelligent models for diabetes prediction using simple clinical features. Several approaches have been explored, ranging from supervised algorithms such as Random Forest and Extra Trees, to gradient boosting models like XGBoost, LightGBM, and CatBoost, as well as neural networks including the Multilayer Perceptron (MLP) and a deep neural network. Ensemble techniques such as stacking, voting, and weighted averaging have been used to improve performance. Handling imbalanced data, hyperparameter optimization, and model explainability are also among the contributions of this work.Three datasets (PIMA, Frankfurt, and a combined dataset) were used to train and evaluate the models. The results were analyzed using various metrics (accuracy, precision, recall, F1-score, AUC-ROC) and visualized through confusion matrices and ROC curves. An interactive web application was developed to enable healthcare professionals to easily and quickly predict the risk of diabetes, with automatic generation of an explanatory report.This work thus illustrates the concrete contribution of AI to predictive medicine and highlights the potential of intelligent models in improving early diagnosis of chronic diseases.