SynTEG: a framework for temporal structured electronic health data simulation.

Area of Science:

• Health Informatics
• Computational Biology
• Data Privacy

Background:

• Electronic health record (EHR) data sharing is crucial for research but poses privacy challenges.
• Existing generative models often overlook the temporal dynamics inherent in healthcare trajectories.
• Balancing data utility with patient privacy remains a significant hurdle in EHR data utilization.

Purpose of the Study:

• To develop a generative framework for simulating patient diagnosis trajectories from EHR data.
• To evaluate the utility of synthetic EHR data by assessing its ability to preserve correlations and temporal patterns.
• To quantify the privacy risks associated with the generated synthetic EHR data.

Main Methods:

• A two-stage generative process was employed to simulate date-stamped diagnosis sequences.
• Temporal patterns were sequentially extracted from clinical visit data.
• Synthetic data generation was conditioned on learned temporal patterns, using phenome-wide association study (phecode) data from over 500,000 EHRs.
• Utility was measured using feature correlations and temporal patterns; privacy was assessed via membership and attribute inference attacks.

Main Results:

• Simulated temporal sequences closely mirrored real sequences across utility measures.
• Diagnosis prediction models showed minimal performance differences (1.6% relative AUC difference) between real and synthetic data.
• Key temporal characteristics like diagnosis age (4.9% difference) and visit intervals (4.2% difference) were well-preserved.
• Privacy risks, including membership and attribute inference, were found to be negligible in the synthetic data.

Conclusions:

• The developed generative framework successfully simulates temporal diagnosis code sequences.
• The approach effectively balances data utility with robust patient privacy protection.
• This method offers a promising solution for secure EHR data sharing and research.