Introduction: Diabetes Mellitus is one of the fastest-growing chronic diseases globally. Machine Learning (ML) techniques offer significant potential for identifying patterns useful for disease control. Objective: To analyze the impact of ML techniques and the use of feature selection techniques in predicting diabetes, using the “Diabetes Health Indicators” dataset. Methods: The CRISP-DM methodology was applied. The data were balanced using the NearMiss subsampling technique. Recursive Feature Elimination (RFE) and Principal Component Analysis (PCA) were used for attribute selection. Six models were tested: Random Forest, Gradient Boosting, KNN, Logistic Regression, Multilayer Perceptron (MLP), and Recurrent Neural Networks (RNN). Results: Class balancing significantly improved results. The RNN achieved the best performance, with 86.8% accuracy and an F1-score of 0.868. The combination of RFE with MLP also showed strong performance. Feature selection (RFE and PCA) reduced dimensionality without loss of accuracy Conclusion: ML and DL techniques are promising for prioritizing clinical follow-up and informing public health policies. Enhancing data representativeness, integrating Explainable AI techniques, and adjusting thresholds to reduce false negatives are essential for practical applications.
Lopes et al. (Fri,) studied this question.