[medgan-pr-prep] Overview: MedGAN PyHealth 2.0 Integration Tasks

[jalengg]

Tracking issue for integrating MedGAN into PyHealth 2.0 standards. Follows the same 9-task pattern used for previous model integrations.

Background

MedGAN is a GAN-based EHR synthetic data generator using:

• A two-phase training process: pretrain autoencoder first, then adversarial training
• Standard BCE loss (NOT WGAN — discriminator uses sigmoid + BCE)
• Minibatch averaging in the discriminator for training stability
• multi_hot schema: flat ICD code lists → binary vectors

Source: pyhealth/models/generators/medgan.py (ported from corgan-medgan-port branch)

Key differences from CorGAN

	CorGAN	MedGAN
GAN loss	Wasserstein (WGAN)	BCE (standard)
Discriminator	WGAN critic, no sigmoid	Sigmoid + minibatch averaging
Training	Single phase	Two-phase: AE pretrain then GAN
Autoencoder	No AE	Linear AE for latent space

Tasks

• T1 — Merge upstream + copy medgan.py into worktree
• T2 — Create MedGANGenerationMIMIC3(BaseTask) task function
• T3 — Refactor MedGAN to remove DummyWrapper, add train_model() / synthesize_dataset()
• T4 — Update training example
• T5 — Update generation example
• T6 — Check/remove bespoke dataset class (likely no-op)
• T7 — Update docstrings
• T8 — Integration tests (8 MECE in-memory + 4 MIMIC-III skip-gracefully)
• T9 — Final verification + push to jalengg/PyHealth:medgan-pr-integration

Branch

medgan-pr-integration (worktree at ~/.config/superpowers/worktrees/PyHealth/medgan-pr-integration)