<![CDATA[Diabetes mellitus requires accurate classification systems to support early detection and clinical decision-making. Prior research has explored the use of Multilayer Perceptron combined with SMOTE, yet the methodological gap remains in evaluating its effectiveness on multiclass clinical datasets with significant class imbalance, particularly for the Prediabetes category. This study addresses that gap by examining the performance of an MLP model enhanced with SMOTE to improve overall accuracy and minority-class detection. The dataset includes age, gender, blood pressure, random blood glucose, weight, and height as clinical predictors. The preprocessing pipeline consists of label encoding for categorical variables, feature standardization, and the application of SMOTE to balance class distribution. The evaluation follows a consistent 80 10 10 split for training, validation, and testing, with three repeated experimental runs to ensure result stability. On the original imbalanced dataset, the MLP achieved an accuracy of 85 percent and showed limited capability in identifying Prediabetes. After applying SMOTE, accuracy increased to 91 percent, accompanied by notable improvements in recall and F1 score across all health status categories. These results demonstrate that SMOTE enables the model to capture non-linear patterns in minority classes and strengthens overall generalization. The proposed model can be integrated into clinical screening workflows as a decision-support tool. Its predictions can help clinicians identify at-risk individuals earlier, prioritize follow-up actions, and enhance patient management in healthcare settings.]]>
