medGAN

Unlike standard GANs that generate directly in data space, medGAN first trains an autoencoder to learn a compressed, continuous latent representation of discrete EHR variables. The GAN then operates in this lower-dimensional manifold, avoiding the instability of generating high-dimensional binary and categorical data directly. This two-stage approach dramatically improves convergence stability and output quality.

medGAN replaces the standard discriminator with a minibatch averaging variant. Instead of scoring individual samples, the discriminator evaluates the statistical distance between averaged minibatches of real and generated data. This prevents the generator from memorizing individual patient records and provides an inherent membership inference defense, as the discriminator cannot isolate single training examples.

EHR data contains mixed variable types: binary (diagnosis present/absent), categorical (medication codes), and continuous (lab values). medGAN employs tailored output activations—sigmoid for binary, softmax for categorical, and linear for continuous variables—within a unified architecture. This preserves column-wise statistical distributions without requiring separate generation pipelines for each data type.

Generating discrete binary codes with standard GANs causes vanishing gradients. medGAN uses a straight-through estimator during training: the generator outputs continuous probabilities, but during the forward pass, values are thresholded to 0 or 1. Gradients flow through the continuous representation, enabling effective backpropagation while producing valid discrete EHR codes at inference time.

Training GANs on sparse, high-dimensional clinical data frequently causes mode collapse, where the generator produces limited varieties of samples. medGAN integrates batch normalization throughout both generator and discriminator networks, stabilizing training dynamics and ensuring the model captures the full diversity of patient phenotypes present in the original dataset.

medGAN evaluates synthetic data quality using the Train-Synthetic-Test-Real (TSTR) paradigm. A downstream predictive model is trained exclusively on generated data and tested on real held-out records. If the model achieves comparable performance to one trained on real data, the synthetic data preserves clinically relevant statistical relationships, not just superficial distributions.

A quantitative measure of how well medGAN outputs preserve the statistical relationships and predictive performance of the original data. Utility is validated using the Train-Synthetic-Test-Real (TSTR) paradigm.

• Propensity Score Matching: Evaluates if treatment effects are preserved in the synthetic cohort.
• Column Correlation Preservation: Ensures clinical logic (e.g., diagnosis-to-medication mappings) remains intact in the generated discrete variables.

An adversarial technique used to audit medGAN's privacy leakage. An attacker model is trained to determine if a specific patient's record was part of the generator's training set. Robust medGAN implementations must demonstrate resistance to black-box membership inference.

• Overfitting Detection: High attack success indicates the generator has memorized training data.
• Mitigation: Achieved through differential privacy, dropout, and strict early stopping during the GAN training loop.