Generating synthetic electronic health record data: a methodological scoping review with benchmarking on phenotype...

Area of Science:

• Health Informatics
• Data Science
• Medical Data Generation

Background:

• Synthetic Electronic Health Records (EHR) data generation is crucial for research and development.
• Existing methods for synthetic EHR data generation vary in their effectiveness and application.
• A comprehensive review and benchmarking of these methods are needed to guide practitioners.

Purpose of the Study:

• To conduct a scoping review of synthetic EHR data generation approaches.
• To benchmark major synthetic data generation methods using open-source EHR datasets.
• To provide an open-source software tool and practical recommendations for synthetic EHR data generation.

Main Methods:

• A scoping review of three academic databases identified 48 relevant studies.
• Seven state-of-the-art methods and two baseline methods were implemented and benchmarked.
• Evaluation focused on data fidelity, downstream utility, privacy protection, and computational cost using MIMIC-III/IV datasets.

Main Results:

• Forty-eight studies were classified into five categories of synthetic data generation methods.
• Generative Adversarial Network (GAN)-based methods showed competitive performance in fidelity and utility.
• Rule-based methods demonstrated superior privacy protection, with similar trends observed on MIMIC-IV data.

Conclusions:

• Method selection depends on the prioritized evaluation metrics for specific use cases.
• A decision tree and an open-source Python package (SynthEHRella) are provided to aid practitioners.
• Future research should focus on enhancing data fidelity and privacy, and benchmarking longitudinal/conditional generation methods.