Optimization of Machine Learning Models for Diabetes Prediction in Senegal

This work explores the optimization of machine learning models to improve diabetes diagnosis and monitoring in Senegal, where medical resources are limited. Using an extensive dataset of 100,000 patients, we implemented and compared several models, including decision trees, random forests, gradient boosting, SVM, KNN, Naive Bayes, and neural networks. To handle class imbalance (91.5% non-diabetic, 8.5% diabetic), we applied advanced oversampling techniques: SMOTE, ADASYN, and Borderline-SMOTE. The results show that gradient boosting and neural networks offer the best performance, with AUC-ROC reaching 0.979. The use of ADASYN significantly improved recall, crucial for early detection, reaching 92.1% for gradient boosting. Feature importance analysis revealed that blood glucose and HbAlc are the most powerful indicators of diabetes. This study also proposes integrating the optimized model into a fog computing platform to develop a smart mobile laboratory, aiming to improve screening access in rural areas. These advances could transform diabetes management in developing countries, combining technological innovation with public health needs.