A method for generating synthetic longitudinal health data.

Area of Science:

• Health Informatics
• Data Science
• Biostatistics

Background:

• Accessing administrative health data for research is challenging due to strict privacy regulations.
• Synthetic datasets, which mimic real data patterns without containing identifiable information, offer a potential solution.

Purpose of the Study:

• To assess the feasibility of generating synthetic administrative health data using a recurrent deep learning model.
• To evaluate the utility and privacy risks of the generated synthetic data.

Main Methods:

• A recurrent deep learning model was used to generate synthetic data from a large administrative health database (120,000 individuals).
• Utility was assessed using distribution comparisons (Hellinger distance) for various data attributes and by replicating a Cox regression analysis.
• Privacy risks were evaluated by assessing attribution disclosure risk.

Main Results:

• Utility assessments showed small differences between real and synthetic data distributions (e.g., Hellinger distance for joint distributions: 0.352).
• Cox regression analysis on synthetic data yielded comparable results to real data, with a 68% mean confidence interval overlap for hazard ratios.
• Privacy assessment indicated attribution disclosure risk was substantially below the acceptable threshold (0.09).

Conclusions:

• The generated synthetic administrative health data demonstrates high utility and is sufficiently similar to the real data for research purposes.
• This synthetic data approach can help overcome barriers to accessing sensitive health information, facilitating broader research.
• The method provides a viable, privacy-preserving alternative for sharing health data in specific research contexts.