Generative Adversarial Networks for Synthetic Data Generation in Diabetic Patient Research: Techniques, Applications,...

Area of Science:

• Medical Informatics
• Computational Biology
• Data Science

Background:

• Clinical data for type 2 diabetes mellitus (T2DM) research is scarce and difficult to obtain.
• Ethical and privacy concerns restrict the use of real patient data.
• T2DM research requires high-quality data for improved diagnostics, treatments, and personalized care.

Purpose of the Study:

• To evaluate four synthetic data generation techniques for T2DM research.
• To assess the quality and utility of synthetically generated data.
• To determine the feasibility of using synthetic data for developing diagnostic tools.

Main Methods:

• Gaussian Mixture Models (GMM)
• Generative Adversarial Networks (GAN)
• Wasserstein GAN (WGAN)
• Variational Autoencoders (VAE)
• Statistical divergence metrics (Jensen-Shannon Divergence - JSD, Kullback-Leibler Divergence - KLD)
• Classification performance analysis

Main Results:

• GMM achieved the lowest JSD, indicating superior distributional similarity.
• WGAN yielded the lowest KLD, suggesting better information content alignment with real data.
• GAN and WGAN demonstrated the highest predictive performance in classification tasks, preserving key data relationships.

Conclusions:

• Synthetic data generation is a viable strategy to overcome limitations in clinical data acquisition for T2DM research.
• Generative models like GAN and WGAN can enhance diagnostic tool development without compromising patient confidentiality.
• The choice of synthetic data generation method should align with specific research objectives, balancing statistical similarity and predictive performance.