Table 2 A summary of the GAN models applied for benchmarking

From: A Multifaceted benchmarking of synthetic electronic health record generation models

Model

Distance measure (loss function)

Auto-encoder for discrete data generation

Normalization

Additional privacy components

medGAN25

Jensen-Shannon divergence

Yes

BatchNorm for generator

No

medBGAN35

f-divergence

Yes

BatchNorm for generator

No

EMR-WGAN20

Wasserstein divergence

No

BatchNorm for generator; LayerNorm for discriminator

No

WGAN35

Wasserstein divergence

Yes

BatchNorm for generator

No

DPGAN36

Wasserstein divergence

Yes

BatchNorm for generator

Yes (differentially private stochastic gradient descent)

  1. All models share the generator-discriminator architecture for EHR data synthesis, but differ in their specializations to enhance either utility or privacy.
Back to article page