Generative models for synthetic data generation: application to pharmacokinetic/pharmacodynamic data.

Area of Science:

• Clinical Pharmacology
• Biostatistics
• Machine Learning

Background:

• Synthetic patient data generation is crucial for enabling data access and augmenting datasets, particularly for under-represented populations.
• Deep learning generative methods offer advanced solutions for creating realistic synthetic data.

Purpose of the Study:

• To benchmark state-of-the-art deep learning generative methods for synthetic patient data generation.
• To evaluate model performance across diverse clinical datasets and scenarios.

Main Methods:

• Implemented and compared Multi-layer Perceptron Conditioning Generative Adversarial Neural Network (MLP cGAN), Time-series Generative Adversarial Networks (TimeGAN), and Probabilistic Autoregressive (PAR) models.
• Evaluated performance using discriminative and predictive scores, statistical tests (Kolmogorov-Smirnov, Chi-square), and pharmacometrics-related metrics.

Main Results:

• MLP cGAN demonstrated the best overall performance across most evaluated metrics.
• The study confirmed the utility of synthetic data for augmenting and sharing proprietary clinical data.

Conclusions:

• Deep learning generative models, particularly MLP cGAN, are effective for generating high-quality synthetic patient data.
• Synthetic data generation holds significant potential for advancing clinical pharmacology research and data accessibility.