A novel pipeline for realistic synthetic longitudinal EHR data generation

Area of Science:

• Health Informatics
• Data Science
• Computational Biology

Background:

• Synthetic health data generation is crucial for patient privacy and data sharing.
• Existing methods often lack structural realism and are narrowly evaluated.
• This limits the practical application of synthetic data in downstream analytical workflows.

Purpose of the Study:

• Introduce a novel pipeline for generating realistic synthetic longitudinal electronic health record (EHR) data.
• Evaluate the pipeline's performance across three diverse datasets.
• Provide evidence-based guidance on using synthetic data to replace or augment real data.

Main Methods:

• Extended existing HALO and ConSequence frameworks with a post-processing step for continuous variables and timestamps.
• Applied the pipeline to small longitudinal, medium intensive-care, and large multi-hospital administrative datasets.
• Assessed realism and utility for machine learning, statistical modeling, and time series analysis.

Main Results:

• Generated realistic synthetic data preserving key statistical properties and relationships across all datasets.
• Machine learning models trained on synthetic data showed comparable predictive accuracy and feature importance to those trained on real data.
• Statistical modeling results closely aligned with real data, though precision for rare conditions may be limited; time series analysis was unsuitable.

Conclusions:

• The pipeline successfully produces realistic and analytically useful synthetic longitudinal EHR data across various scales.
• Synthetic data demonstrates strong utility for machine learning and statistical modeling tasks.
• Findings offer practical guidance for the judicious use of synthetic data in healthcare analytics.