Conditional Adversarial Random Forest for Synthetic Electronic Health Record Generation

Cynthia Garcia Ybarra, Christian Gagné

Keywords: Machine Learning, Synthetic Data generation, healthcare, Adversial Conditional Random Forests

Abstract:

Synthetic Electronic Health Records (EHRs) enable privacy-preserving healthcare data sharing for machine
learning research. However, existing methods struggle with: maintaining temporal consistency across patient
visits while preserving demographic-clinical correlations. Current approaches either sacrifice temporal
fidelity or require extensive postprocessing.
We propose Conditional Adversarial Random Forest (CARF), extending Adversarial Random Forest [1]
with a two-model strategy. The first model generates patient-level demographics that remain static across
visits. The second conditional model produces visit-level clinical variables, incorporating visit rank and time
progression to create complete patient trajectories. This eliminates manual postprocessing while preserving
temporal patterns inherently.

Submission Number: 207