Learning debiased graph representations from the OMOP common data model for synthetic data generation.

Area of Science:

• Medical Informatics
• Computational Biology
• Health Data Science

Background:

• Current synthetic patient data generation methods rely on black-box models, limiting expert verification and intervention.
• There is a need for privacy-preserving, compliant, interpretable, and verifiable synthetic data generation techniques.

Purpose of the Study:

• To develop and evaluate a novel method for generating synthetic patient data that addresses the limitations of existing approaches.
• To enable expert intervention and verification in the synthetic data generation process.

Main Methods:

• The proposed approach integrates OMOP (Observational Medical Outcomes Partnership) data standard with the Synthea data synthetization tool.
• Data pipelines were constructed to extract OMOP data, convert it to time series, and apply statistical (Markov chain, TARM) and causal discovery (DYNOTEARS, J-PCMCI+, LiNGAM) algorithms.
• Learned temporal rules were mapped into Synthea graphs, which were then evaluated quantitatively and qualitatively by medical experts.

Main Results:

• Different algorithms produced distinct graph representations; Markov chains resulted in large graphs, while TARM, DYNOTEARS, and J-PCMCI+ reduced data dimensions.
• The MultiGroupDirect LiNGAM algorithm proved unsuitable for this specific application.
• Quantitative and qualitative evaluations highlighted the varying effectiveness of the algorithms.

Conclusions:

• TARM and DYNOTEARS emerged as the most practical algorithms for real-world synthetic patient data generation in this context.
• DYNOTEARS, a gradient-based causal discovery algorithm, was deemed most suitable due to its ability to debias statistical relationships.