Peningkatan Sensitivitas Model Boosting untuk Deteksi Diabetes Menggunakan SMOTE pada Imbalanced Dataset

Setyawan Wibisono (Universitas Stikubank)
Eko Nur Wahyudi (Universitas Stikubank)
Imam Husni Al Amin (Universitas Stikubank)

Dinamik
Vol 31 No 2 (2026)

This study aims to analyze the effect of SMOTE on the sensitivity of boosting models for diabetes detection using the BRFSS 2015 dataset. The dataset consists of 253,680 instances with 21 features and a binary target, namely diabetes and non-diabetes. The primary issue in the dataset is class imbalance, causing the models to be more biased toward recognizing the non-diabetes class. The algorithms employed in this study include AdaBoost, XGBoost, and Gradient Boosting, evaluated under two scenarios: without SMOTE and with SMOTE. Model performance was assessed using accuracy, precision, recall, F1-score, ROC-AUC, confusion matrix, and 10-fold cross validation. The results demonstrate that SMOTE improves recall across all models. The most significant improvement occurred in AdaBoost, where recall increased from 0.016551 to 0.711840. The cross-validation results also showed that AdaBoost + SMOTE achieved a recall value of 0.721384. Although accuracy and precision decreased, AdaBoost + SMOTE became the most sensitive model for detecting diabetes. Therefore, this model has potential to be utilized as an early diabetes screening support tool.