ScholarWorks@성균관대학교

초록

Diabetes represents a significant metabolic disorder marked by elevated glucose levels due to suboptimal insulin production or function. Early diagnosis and effective diabetes management are crucial to reducing related health complications. This study introduces a robust approach for predicting diabetes through advanced machine learning methods. Utilizing the diabetes dataset from the University of California Irvine (UCI) machine learning repository, we performed extensive preprocessing to guarantee data quality and integrity. To counteract class imbalance, we employed the synthetic minority over-sampling technique, which improved the representation of minority classes. We explored several machine learning (ML) models, including Random Forest (RF), logistic regression (LR), and K-nearest neighbors (KNN), while optimizing hyperparameters through grid search and randomized search techniques. Additionally, we introduced a stacking ensemble method paired with a tab transformer model, effectively harnessing the advantages of both techniques for efficient handling of tabular data. The outcomes from the stacking and tab transformer models were later aggregated using a meta learner, specifically extreme gradient boosting (XGBoost), to create a robust ensemble model. Our comprehensive methodology yielded an impressive accuracy rate of 99%, significantly outperforming individual models. Unlike previous studies that rely solely on individual models, our approach fills the gap by combining deep learning with ensemble methods to enhance generalization and interpretability in diabetes prediction. We have validated the model's performance using ablation studies and paired statistical significance tests. These results highlight the efficacy of integrating diverse ML strategies to enhance both the accuracy and reliability of diabetes prediction.

키워드