Augmented TimeGAN in Clinical Data Synthesis

• Augmented TimeGAN is a generative model that enhances clinical time-series synthesis by injecting Gaussian noise into real embeddings for improved discriminator stability.
• It utilizes a five-component GRU architecture and a three-phase training protocol to maintain temporal consistency and boost the realism of synthetic data.
• Empirical results demonstrate significant gains in metrics like MMD, α-precision, and downstream predictive performance over the baseline TimeGAN in healthcare applications.

Augmented TimeGAN is a modification of the original TimeGAN framework for generating realistic synthetic time-series data, specifically designed to address the challenges of modeling longitudinal clinical records. The defining innovation is a lightweight augmentation: injection of i.i.d. Gaussian noise into real embeddings before discriminator evaluation. This intervention regularizes the adversarial dynamic, dramatically stabilizing training and yielding synthetic outputs superior in faithfulness and diversity by several statistical metrics. Augmented TimeGAN's best-of-breed performance stands out in healthcare settings, where strict privacy regimes and the need for data realism constrain generative modeling options (Ballyk et al., 29 Nov 2025).

1. Model Architecture

Augmented TimeGAN retains the canonical five-component RNN backbone of TimeGAN—Embedding network ($\mathcal{E}$), Recovery network ($\mathcal{R}$), Supervisor ($\mathcal{S}$), Generator ($G$), and Discriminator ($D$)—each realized as multi-layer gated recurrent units (GRUs). The augmentation consists of a Gaussian noise injection applied exclusively at the input to the discriminator for real embeddings. The architectural pathways can be summarized as follows:

• Encoding–Decoding (Autoencoder Path): $$\mathbf{x}_{1:T} \xrightarrow{\mathcal{E}} \mathbf{h}_{1:T} \xrightarrow{\mathcal{R}} \tilde{\mathbf{x}}_{1:T}$$
• Temporal Consistency (Supervisory Path): $$\mathbf{h}_{1:T} \xrightarrow{\mathcal{S}} \mathbf{\hat{h}}_{2:T+1}$$
• Synthesis: $$\mathbf{z}_{1:T} \xrightarrow{G} \mathbf{e}_{1:T} \xrightarrow{\mathcal{S}} \hat{\mathbf{h}}_{2:T+1} \xrightarrow{\mathcal{R}} \tilde{\mathbf{x}}_{1:T}$$
• Discrimination: Discriminator $D$ receives either $\mathbf{h}_{1:T}^\text{noisy}$ (noise-perturbed real embedding) or $\mathcal{S}(G(\mathbf{z}_{1:T}))$ (synthetic embedding).

The Gaussian augmentation is the sole architectural modification: for real embeddings, noise $\mathbf{n}_t\sim\mathcal{N}(0,\sigma^2 I)$ is added at each time step so that $$\mathbf{h}^{\text{noisy}}_{1:T} = \mathbf{h}_{1:T} + \mathbf{n}_{1:T}$$, with standard deviation $\sigma=0.2$.

2. Mathematical Formulation

Objective functions are closely aligned with the original TimeGAN losses, with the only addition being the regularization induced by noise injection. All loss terms in the joint min–max optimization are precisely defined:

• Reconstruction Loss:

$$\mathcal{L}_\mathrm{rec} = \mathbb{E}\left[\|\mathbf{x}_{1:T} - \mathcal{R}(\mathcal{E}(\mathbf{x}_{1:T}))\|_2^2\right]$$

• Supervised Loss:

$$\mathcal{L}_\mathrm{sup} = \mathbb{E}\left[\|\mathcal{E}(\mathbf{x}_{2:T+1}) - \mathcal{S}(\mathcal{E}(\mathbf{x}_{1:T}))\|_2^2\right]$$

• Adversarial Loss:

$$\mathcal{L}_\mathrm{adv} = \mathbb{E}\left[ \log D(\mathbf{h}_{1:T} + \mathbf{n}_{1:T}) \right] + \mathbb{E}\left[ \log\big(1 - D(\mathcal{S}(G(\mathbf{z}_{1:T})))\big)\right]$$

• Unsupervised Adversarial Loss (Optional):

$$\mathcal{L}_\mathrm{uns} = \mathbb{E}\left[ \log D(\mathcal{E}(\mathbf{x}_{1:T})) \right] + \mathbb{E}\left[ \log\big(1 - D(G(\mathbf{z}_{1:T}))\big)\right]$$

• Augmentation Loss:

$\mathcal{L}_\mathrm{aug} = 0$

(The noise injection itself does not introduce an explicit loss term.)

The composite objective is

$$\min_{G, \mathcal{E}, \mathcal{S}}\;\max_D \left\{ \mathcal{L}_\mathrm{adv} + \beta \mathcal{L}_\mathrm{uns} + \alpha \mathcal{L}_\mathrm{rec} + \gamma \mathcal{L}_\mathrm{sup} \right\}$$

with $\alpha = \beta = \gamma = 1$ in practice.

3. Training Protocol and Hyperparameters

Training employs a three-phase regime:

1. Phase I: $\mathcal{E},\mathcal{R}$ optimize $\mathcal{L}_\mathrm{rec}$ (1,000 epochs).
2. Phase II: $\mathcal{S},\mathcal{E}$ optimize $\mathcal{L}_\mathrm{sup}$ (1,000 epochs).
3. Phase III: $G, D$ (with fine-tuning of $\mathcal{E},\mathcal{S}$) optimize adversarial losses ($$\mathcal{L}_\mathrm{adv} + \mathcal{L}_\mathrm{uns}$$), for the remainder of training.

Key hyperparameters are:

• Optimizer: Adam (default settings)
• Learning rate: $1 \times 10^{-3}$
• Batch size: 32
• Latent dimension: $p = 24$
• Network depth: 3 GRU layers per module
• Noise standard deviation: $\sigma = 0.2$
• Total epochs: 7,000

Dataset-specific sequence lengths:

• Sines: $T = 24$, 5 features
• eICU: $T = 12$, 5 features
• CKD: $T = 7$, 7 features

4. Evaluation Metrics and Empirical Performance

Performance is benchmarked on five “faithfulness/diversity/privacy” metrics:

Model	MMD	DS	α-Precision	β-Recall	Authenticity
TimeGAN (baseline)	0.061	0.341	0.848	0.841	0.613
Augmented TimeGAN	0.049±.013	0.231±.049	0.925±.030	0.936±.019	0.604±.083
DP-TimeGAN (private)	–	–	–	–	0.778±.053

On the chronic kidney disease (CKD) dataset, Augmented TimeGAN achieves improved MMD (0.049 vs 0.061), lower discriminative score (0.231 vs 0.341), and substantial gains in both α-precision and β-recall over the original TimeGAN. Authenticity is calculated as the fraction of generated samples whose nearest-neighbor distance to any real training point exceeds a given threshold.

Train-on-Synthetic-Test-on-Real (TSTR) evaluations show that Augmented TimeGAN achieves a predictive score of $0.381 \pm 0.050$ and AUC-ROC of $0.615 \pm 0.046$ for downstream diabetes classification in CKD, outperforming the base TimeGAN in AUC-ROC (0.615 vs 0.564). Blinded clinical review reports deception rates of approximately 96% in classifying CKD patient time-series as real or synthetic.

5. Comparative Analysis and Ablation Studies

Ablation experiments on a synthetic sinusoid toy dataset demonstrate that noise injection yields the largest improvement among tested modifications (e.g., replacing GRUs with xLSTM). Specifically, noise-injected models achieve MMD reduction from 0.008 to 0.002, discriminative score reduction from 0.269 to 0.089, and marked increases in α-precision (0.648→0.951) and β-recall (0.657→0.963).

In comparison with other model families—transformer-based (TransFusion) and flow-based models (SeriesGAN, DP Normalizing Flows)—Augmented TimeGAN consistently outperforms on statistical fidelity and diversity while retaining comparable TSTR utility and clinical realism.

6. Interpretation and Practical Implications

The core insight of Augmented TimeGAN is that minor Gaussian noise regularization in the discriminator input is sufficient to prevent discriminator overfitting early in training, which otherwise results in generator collapse or poor diversity. This stabilization maintains a strong gradient signal to the generator throughout adversarial cycles, enabling the synthesis of high-fidelity, diverse, and non-memorized temporal samples, especially in irregular and noisy EHR sequences.

A plausible implication is that similar lightweight regularization mechanisms could be beneficial for adversarial models dealing with tabular or multivariate sequential data, particularly in low-sample-size or highly privacy-constrained domains.

7. Summary of Contributions and Outlook

Augmented TimeGAN introduces a single-point regularization into a well-calibrated RNN GAN framework, validating its impact empirically on both public and real-world clinical datasets. This approach succeeds in generating temporally consistent, diverse, and non-memorized synthetic health records while maintaining or improving upon the fidelity and utility of existing state-of-the-art generative models. Its deployment in privacy-sensitive applications such as synthetic EHR generation is supported by favorable authenticity and deception metrics, as well as rigorous statistical comparisons across baselines (Ballyk et al., 29 Nov 2025).