Medical data sharing and synthetic clinical data generation - maximizing biomedical resource utilization and...

Area of Science:

• Biomedical Informatics
• Data Privacy
• Health Data Science

Background:

• Sharing sensitive electronic health records (EHR) and wearable data poses significant re-identification risks.
• Existing methods struggle to balance data utility with robust privacy preservation for complex datasets.

Purpose of the Study:

• To introduce and evaluate an end-to-end pipeline for generating privacy-preserving "digital twin" datasets.
• To compare the efficacy of DataSifter and Synthetic Data Vault (SDV) methods in obfuscating EHR and wearable data.
• To benchmark privacy-utility trade-offs across different obfuscation levels.

Main Methods:

• Utilized DataSifter and Synthetic Data Vault (SDV) with varying obfuscation levels (DataSifter: small, medium, large; SDV: CTGAN, Gaussian Copula).
• Generated digital twin datasets from 3029 participants' EHR and Apple Watch data.
• Assessed utility via statistical fidelity and machine learning performance; evaluated privacy using re-identification risk and detection likelihood metrics.

Main Results:

• The highest-obfuscation DataSifter digital twin demonstrated superior privacy protection (0.83) compared to SDV.
• DataSifter maintained significant statistical and predictive signal preservation (83.1% CI overlap in regression models).
• While higher obfuscation impacted machine learning performance, overall data utility was generally preserved, especially for longitudinal data.

Conclusions:

• Digital twin datasets are vital for secure biomedical data sharing.
• DataSifter provides adaptable privacy-utility trade-offs, outperforming SDV, particularly for longitudinal data.
• The developed pipeline supports secure sharing of complex health data while minimizing re-identification risks.