Tabular transformer generative adversarial network for heterogeneous distribution in healthcare

Area of Science:

• Medical Artificial Intelligence (AI)
• Data Science
• Bioinformatics

Background:

• Healthcare tabular data (HTD) is crucial for medical AI but faces privacy challenges.
• Generating realistic synthetic HTD is complex due to intricate variable interdependencies and sensitive information.
• Existing synthetic data generation methods struggle with the complexity of healthcare datasets.

Purpose of the Study:

• To propose a Tabular Transformer Generative Adversarial Network (TT-GAN) for generating privacy-preserving synthetic healthcare tabular data.
• To effectively capture inter-variable relationships within HTD using a multi-attention mechanism.
• To ensure data privacy through an implicit-based algorithm within a generative adversarial network (GAN) architecture.

Main Methods:

• Development of a TT-GAN incorporating Transformer architecture for multi-attention mechanisms to model column relationships.
• Application of discretization and converter methodologies to handle heterogeneous continuous variables in HTD.
• Comparison of TT-GAN performance against Conditional Tabular GAN (CTGAN) and copula GAN.

Main Results:

• TT-GAN demonstrated superior performance in generating synthetic data that closely resembles real healthcare datasets compared to CTGAN and copula GAN.
• The proposed Transformer algorithm, combined with discretization and converters, proved effective for HTD synthesis.
• Transformer-based models without discretization and converters showed significantly inferior performance, highlighting the importance of the proposed methodology.

Conclusions:

• TT-GAN shows significant potential for healthcare applications, offering a robust solution for generating realistic and privacy-preserving synthetic tabular data.
• The TT-GAN's ability to handle mixed variable types (polynomial, discrete, continuous) underscores its versatility in health research and data synthesis.
• The study validates the effectiveness of discretization and converter methodologies when integrated with Transformer-based generative models for HTD.