Optimization of the Decision Tree Algorithm Using SMOTE for the Classification of Diabetes Mellitus Types Based on Clinical Medical Record Data

Abstract

Diabetes mellitus is a chronic metabolic disorder characterized by elevated blood glucose levels. Accurate early classification of diabetes types is crucial for determining appropriate clinical interventions. However, clinical datasets often suffer from severe class imbalance, leading to model bias toward the majority class. This study proposes the optimization of the Decision Tree classification algorithm by integrating the Synthetic Minority Over-sampling Technique (SMOTE) using real-world clinical medical record data from the Singaparna Community Health Center. The dataset comprises 225 patient records with 38 health attributes. Preprocessing and attribute transformation techniques were applied, followed by stratified random sampling (80:20 split). To resolve the extreme class imbalance (1 Type 1 DM vs 224 Type 2 DM samples), SMOTE was applied exclusively to the training set, balancing the dataset to 358 instances (179 samples per class). The Decision Tree model was trained and evaluated using validation curves to prevent overfitting. The experimental results demonstrate that the SMOTE-optimized Decision Tree model with an optimal max depth of 3 achieved robust classification performance, yielding an accuracy of 86.67%, precision of 88.54%, recall of 84.12%, F1-score of 85.20%, and an ROC-AUC score of 0.9333. Furthermore, the decision tree visualization revealed transparent, age-based rule splits matching clinical domain knowledge. This study confirms that SMOTE effectively mitigates class imbalance in clinical decision-tree modeling, providing a reliable and interpretable decision-support tool for diabetes screening in primary healthcare settings.