Refining Diabetes Diagnosis Models: The Impact of SMOTE on SVM, Logistic Regression, and Naïve Bayes
Arief Wibowo, Anis Fitri Nur Masruriyah, Selly Rahmawati · Journal of Electronics Electromedical Engineering and Medical Informatics · 2025
The accurate classification of diabetes remains a significant challenge in medical diagnostics, particularly when dealing with imbalanced datasets. This study examines the impact of the Synthetic Minority Over-sampling Technique (SMOTE) on the performance of three machine learning algorithms: Support Vector Machine (SVM), Logsistic Regression, and Naïve Bayes. Initially, the models were evaluated on the original imbalanced dataset, with AUC values of 0.598 for SVM, 0.622 for Naïve Bayes, and 0.613 for Logistic Regression, reflecting the challenges posed by class imbalance. To address this issue, SMOTE was applied to balance the dataset by oversampling the minority class. After applying SMOTE, the models showed significant improvements in performance, with AUC values increasing to 0.991 for SVM, 0.987 for Logistic Regression, and 0.986 for Naïve Bayes. These results demonstrate SMOTE's effectiveness in enhancing model performance by mitigating the impact of imbalanced data, especially in the context of diabetes classification. Although the improvements in AUC were notable, it’s important to recognize the influence of data balancing on the models' behavior. These findings emphasize the need for careful handling of imbalanced datasets when developing machine learning models for medical diagnosis. This study provides valuable insights into the role of oversampling techniques like SMOTE in improving predictive accuracy for diabetes, a critical step toward more reliable medical diagnostics.