Reward-Guided Generation Improves the Scientific Utility of Synthetic Biomedical Data.

Area of Science:

• Biomedical informatics
• Machine learning
• Data science

Background:

• Synthetic data generation is crucial for biomedical data sharing and augmentation.
• Existing methods often fail to preserve statistical properties essential for scientific analysis.

Purpose of the Study:

• To introduce RLSYN+REG, a reinforcement learning-driven generative model.
• To ensure regression models trained on synthetic data accurately reproduce real-data coefficients and predictions.

Main Methods:

• Developed a reinforcement learning-driven generative model, RLSYN+REG.
• Evaluated the model on MIMIC-III and American Community Survey (ACS) datasets.
• Assessed regression model reproduction, data fidelity, and privacy.

Main Results:

• RLSYN+REG significantly improved correlations between real and synthetic regression coefficients (MIMIC-III: 0.054 to 0.600; ACS: 0.160 to 0.376).
• Reduced the predictive performance gap between real and synthetic data by 81.4% (MIMIC-III) and 97.6% (ACS).
• Achieved these improvements with negligible impact on data fidelity or privacy, demonstrating robustness to reduced training data.

Conclusions:

• RLSYN+REG offers a superior approach to synthetic data generation for biomedical applications.
• The model effectively preserves statistical properties necessary for robust scientific analysis and data sharing.