Privacy-Preserving Generative Deep Neural Networks Support Clinical Data Sharing.

Area of Science:

• Biomedical informatics
• Machine learning
• Data privacy

Background:

• Sharing individual-level data accelerates scientific progress but patient privacy concerns create a barrier.
• Developing methods to share clinical data while preserving privacy is crucial for advancing research.

Purpose of the Study:

• To generate synthetic participants resembling real individuals from the Systolic Blood Pressure Trial (SPRINT) using deep neural networks.
• To train these networks with differential privacy to ensure patient confidentiality.
• To assess the utility of synthetic data for machine learning and hypothesis generation.

Main Methods:

• Utilized pairs of deep neural networks to generate synthetic participants.
• Implemented differential privacy during network training to protect individual participant information.
• Trained machine learning predictors on the synthetic data.

Main Results:

• Generated synthetic participants closely resembling those in the Systolic Blood Pressure Trial.
• Demonstrated that differential privacy can be effectively applied to these networks.
• Machine learning models trained on synthetic data generalized well to the original dataset.

Conclusions:

• Deep neural networks can generate privacy-preserving synthetic participants.
• This synthetic data facilitates secondary analyses and reproducible investigations of clinical datasets.
• Enhanced data sharing through synthetic data generation supports scientific advancement while maintaining participant privacy.