Synthetic Data and Health Privacy

• Published in: JAMA : the journal of the American Medical Association Vol. 333; no. 7; pp. 567 - 568
• Main Authors: Abgrall, Gwénolé, Monnet, Xavier, Arora, Anmol
• Format: Journal Article
• Language: English
• Published: United States American Medical Association 18.02.2025; American Medical Association (AMA)
• Subjects: Artificial Intelligence; Computer Science; Generative adversarial networks; Generative artificial intelligence; Human health and pathology; Legislation; Life Sciences; Privacy; Risk reduction; Statistical analysis; Statistical models; Synthetic data
• ISSN: 0098-7484; 1538-3598; 1538-3598
• DOI: 10.1001/jama.2024.25821

Abgrall et al discuss generative artificial intelligence and safeguarding privacy by using synthetic data as a substitute for private health data. Synthetic data, designed to emulate real patient characteristics without revealing identifiable information, offer a potential solution to this conundrum. They can be generated by using rule-based techniques, using statistical modeling to approximate real data distributions, or training generative artificial intelligence (GAI) models such as generative adversarial networks on real data to capture underlying structures and create more complex synthetic datasets. Creating synthetic data from real personal data is considered processing under the European Union's General Data Protection Regulation. However, whether and when synthetic data remain personal data--and thus subject to the regulation--remains a complex issue. Recent legislation points toward synthetic data as not being considered personal data. Clinicians and developers should avoid using sensitive data to train, fine-tune, or use GAI models, reducing the risk of privacy breaches.