Anonymization Through Data Synthesis Using Generative Adversarial Networks (ADS-GAN).

Area of Science:

• Medical Informatics
• Machine Learning
• Data Privacy

Background:

• Artificial intelligence (AI) promises to revolutionize medicine, but progress is hindered by challenges in accessing sensitive electronic health records (EHR).
• Legal and ethical concerns regarding patient data privacy and consent create significant barriers to data sharing for machine learning research.
• Existing anonymization methods lack clear definitions, impeding the development of AI-driven medical solutions.

Purpose of the Study:

• To propose a novel framework for generating synthetic EHR data that preserves data utility while ensuring patient privacy.
• To establish a quantifiable definition of "identifiability" to guide the anonymization process.
• To enable broader and more secure data sharing for advancing AI in healthcare.

Main Methods:

• Developed a conditional generative adversarial network (GAN) framework, termed ADS-GAN, to generate synthetic EHR data.
• Introduced a mathematical definition of "identifiability" based on re-identification probability to minimize patient identifiability.
• Evaluated the synthetic data by comparing models trained on it versus models trained on real EHR data across four independent datasets.

Main Results:

• ADS-GAN demonstrated superior performance compared to state-of-the-art methods in generating reliable synthetic EHR data.
• The generated synthetic data closely approximated the joint distribution of variables in the original EHR datasets.
• Model performance using synthetic data showed high similarity to models trained on real patient data.

Conclusions:

• The proposed ADS-GAN framework offers a viable solution for creating privacy-preserving synthetic EHR datasets.
• This approach facilitates open data sharing, accelerating the development and deployment of AI in medicine.
• Publicly available datasets generated with this method can significantly reduce the risk of patient confidentiality breaches.