<![CDATA[Diabetes Mellitus (DM) is a chronic disease that can lead to serious complications if not detected early; therefore, early diagnosis is highly important. One of the methods that can be applied for early diagnosis is the classification technique in data mining. However, the classification process often faces challenges due to class imbalance, which can reduce model performance. This study aims to analyze the effect of class balancing techniques on the performance of the Backpropagation Neural Network (BPNN) in classifying DM cases. BPNN is a form of Multi-Layer Perceptron (MLP) with a simple structure and the ability to solve complex problems with good accuracy. The dataset used in this study is the Pima Indians Diabetes Dataset, consisting of 768 instances, including 500 non-diabetic and 268 diabetic cases. The research was conducted using three scenarios: without balancing, Synthetic Minority Over-sampling Technique (SMOTE), and Random Under Sampling (RUS). The BPNN model was designed with two architectural variations (one hidden layer and two hidden layers), three learning rate values (0.1, 0.01, and 0.001), and a varying number of neurons. The dataset was divided using the 10-Fold Cross Validation technique. The results show that applying SMOTE achieved the best performance, with an average accuracy of 90.89%, precision of 91.22%, recall of 90.89%, and F1-score of 90.89% on the BPNN architecture with one hidden layer. Furthermore, the single hidden layer architecture proved more stable than the two hidden layers, especially when the dataset size decreased due to RUS. Therefore, the combination of SMOTE and BPNN with one hidden layer provides better performance in classifying Diabetes Mellitus cases.]]>
