High-Fidelity Synthetic Data Replicates Clinical Prediction Performance in a Million-Patient Diabetes Cohort.

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Biomedical Data Science

Background:

• Synthetic patient data are crucial for privacy-preserving clinical research.
• Ensuring statistical fidelity and biomedical plausibility of synthetic data is challenging.
• Large-scale real-world clinical datasets are vital for robust research.

Purpose of the Study:

• To generate longitudinal synthetic patient datasets using a dual adversarial autoencoder.
• To evaluate the utility and biomedical plausibility of synthetic data for clinical research.
• To assess the performance of machine learning models trained on synthetic versus real-world data.

Main Methods:

• Employed a dual adversarial autoencoder to generate synthetic longitudinal data.
• Utilized real-world clinical data from nearly one million diabetes patients.
• Conducted multi-faceted evaluations including machine learning tasks and biomedical plausibility assessments.

Main Results:

• Synthetic data models achieved predictive performance comparable to real data models for chronic kidney disease onset.
• Feature importance rankings were stable, indicating clinical coherence.
• Sex-stratified analyses revealed inconsistencies not captured by standard metrics, highlighting the need for domain-specific evaluation.

Conclusions:

• Synthetic data can effectively replicate predictive performance but require robust validation.
• A comprehensive validation framework integrating machine learning utility and domain-specific biomedical evaluation is essential.
• Synthetic data generation supports large-scale, privacy-preserving research and collaborative healthcare data ecosystems.