A Comparative Assessment of Synthetic Data Augmentation Methodologies (SMOTE, Variational Autoencoders, and GANs) in Resolving Class Imbalance for Predictive Diabetes Analytics

• Mark Kennth

  Author

Diabetes mellitus remains a prevalent chronic disease with significant global health implications, where early detection is critical for effective intervention and management. Machine learning models for diabetes prediction face a persistent challenge: class imbalance in clinical datasets, where non-diabetic cases substantially outnumber diabetic cases, leading to biased predictions favoring the majority class. While Synthetic Minority Oversampling Technique (SMOTE) has been widely adopted to address this issue, emerging generative approaches including Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) offer potentially superior solutions by learning the underlying data distribution and generating more diverse synthetic samples. This study conducts a comprehensive comparative assessment of these three synthetic data augmentation methodologies using the PIMA Indian Diabetes dataset, evaluating their impact on classification performance when integrated with a Random Forest classifier. Results demonstrate that GAN-based augmentation achieves superior performance with an F1-score of 90.0%, representing a 20% improvement over SMOTE-based approaches, while VAE augmentation achieves 73.5% ROC-AUC and 67.7% F1-score. The findings establish GANs as the preferred augmentation strategy for imbalanced medical datasets, providing practitioners with evidence-based guidance for developing more reliable predictive models for diabetes diagnostics. This research contributes to the growing body of literature on generative methods in healthcare analytics and offers a replicable framework for comparative augmentation assessment.

• PDF

Copyright (c) 2026 Mark Kennth (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.