Synthetic Tabular Data Based on Generative Adversarial Networks in Health Care: Generation and Validation Using the...

Area of Science:

• Health informatics
• Machine learning
• Data science

Background:

• Generative adversarial networks (GANs) are used for synthetic data generation (SDG) in healthcare.
• Preserving logical relationships in synthetic tabular data (STD) remains a challenge.
• Existing SDG filtering methods risk losing critical information.

Purpose of the Study:

• To propose a novel divide-and-conquer (DC) method for generating STD using GANs.
• To ensure the preservation of logical relationships within the synthetic data.
• To enhance the utility of synthetic data for healthcare applications.

Main Methods:

• The DC-based SDG strategy involves partitioning data using class-specific and Cramer V criteria.
• Subsets are processed by conditional tabular GAN and copula GAN to generate synthetic data.
• Generated data are consolidated, and performance is validated against conditional sampling (CS)-based SDG using machine learning models.

Main Results:

• The DC-based SDG method demonstrated superior performance across four classifiers (DT, RF, XGBoost, LGBM) compared to CS-based SDG.
• Specific Area Under the Curve (AUC) values indicate improved predictive accuracy for the DC method across NSCLC, breast cancer, and diabetes datasets.
• Balanced synthetic data consistently outperformed imbalanced data in model performance evaluations.

Conclusions:

• This study presents the first validation of a DC approach for STD generation, showing improved performance.
• The DC method effectively preserves logical data relationships, a significant advancement in SDG.
• The findings underscore the importance of generating balanced synthetic data for robust machine learning applications.