Feature-based ensemble modeling for addressing diabetes data imbalance using the SMOTE, RUS, and random forest...

Area of Science:

• Machine Learning
• Medical Informatics
• Data Science

Background:

• Class imbalance is a significant challenge in machine learning for medical datasets.
• Early diabetes detection requires accurate predictive models to identify at-risk individuals.

Purpose of the Study:

• To develop and evaluate a feature-based ensemble model for early diabetes detection.
• To address class imbalance using synthetic minority oversampling technique (SMOTE) and random undersampling (RUS).
• To improve predictive performance compared to baseline models.

Main Methods:

• Utilized the Scikit-learn diabetes dataset, binarizing the target variable and employing stratified sampling.
• Balanced the training set using SMOTE (0.6) and RUS (0.66) to a 1:2 ratio.
• Constructed a feature-based ensemble model with random forest classifiers on two-feature subsets, combined via soft voting.

Main Results:

• The feature-based ensemble model achieved the highest accuracy (0.8764) and excellent AUC (0.9227).
• Demonstrated superior discriminative ability for the minority (high-risk) class.
• Outperformed 13 baseline models, including a fully connected neural network.

Conclusions:

• Integrating SMOTE, RUS, and feature-based ensemble learning enhances classification performance for imbalanced diabetes data.
• The proposed approach offers a scalable and interpretable solution for early diabetes prediction.
• This method shows potential for application in other medical prediction tasks.