Approche basée IA pour un système de prédiction du diabète

Abstract

This thesis explores the potential of artificial intelligence in early diabetes prediction, highlighting innovative contributions while utilizing various databases. The first two contributions use the "Pima Indian Diabetes Database". An ensemble model based on the three best methods obtained from supervised machine learning classification achieves an accuracy of 90.62%, surpassing other state-of-the-art methods. The contribution continues by focusing on handling missing values, introducing an innovative imputation approach. The Random Forest model, with an accuracy of 92%, demonstrates the effectiveness of this method in managing incomplete data. The subsequent contributions leverage the "Early Stage Diabetes Risk Prediction Dataset". In the third contribution, an evaluation of seven major classification techniques reveals that the XGBoost algorithm outperforms its counterparts with a remarkable F1 score of 94.74% and an accuracy of 96.15%. For the fourth contribution, the crucial challenge of early diabetes prediction is addressed by balancing data, carefully selecting features, and applying nine supervised machine learning techniques. The Extra Trees algorithm stands out with an exceptional accuracy of 97.95%, significantly surpassing other referenced models in the literature, highlighting the efficiency of the developed approaches in diabetes prediction, with remarkable precision. The last contribution is based on the "U.S. Centers for Disease Control and Prevention dataset" to enhance the performance of prediction models based on survey data. The presented machine learning approach, especially with the Random Forest model, demonstrates outstanding performance on a larger test dataset, anticipating the system's validity in a real-world context. In conclusion, this thesis proposes a performance-driven approach based on AI for early diabetes prediction while exploiting different datasets. The exceptional results obtained underscore the efficiency of the developed models, paving the way for more targeted and preventive medical interventions, as well as significant improvements to existing diabetes prediction systems