<![CDATA[Diabetes mellitus is a metabolic disease that is spreading rapidly and has the potential to be life-threatening worldwide. This condition occurs when the body experiences a decline in its ability to process glucose, triggering metabolic disorders. The use of machine learning algorithms is one effective approach to predicting or detecting diabetes based on the severity of a patient's symptoms. This study uses the Diabetes dataset from Kaggle and compares the performance of several classification algorithms in unbalanced data conditions and after data balancing using the SMOTE, Random Under Sampling, Random Over Sampling, and Near Miss resampling techniques. The results show that model performance is greatly influenced by data balance conditions and the resampling method used. In the original unbalanced data condition, Artificial Neural Network (ANN) provided the best results with the highest accuracy of 96.98%, indicating that ANN is the most adaptive to class imbalance. After resampling, the performance pattern changed: with SMOTE, Random Under Sampling, and Random Over Sampling, the Random Forest algorithm consistently produced the highest accuracy of 96.52%, 89.84%, and 96.26%, respectively, demonstrating its superiority in utilizing balanced data. Meanwhile, in the Near Miss method, the best performance was achieved by Logistic Regression with an accuracy of 94.41%, indicating that minority sample selection based on proximity is more suitable for linear models. Therefore, selecting the right combination of resampling methods and machine learning algorithms is an important factor in obtaining optimal diabetes predictions.]]>
