Abstract
Clinical trials face mounting challenges: fragmented patient populations, slow enrollment, and unsustainable costs, particularly for late phase trials in oncology and rare diseases. While external control arms built from real-world data have been explored, a promising alternative is the generation of synthetic control arms using generative AI. A central challenge is the generation of time-to-event outcomes, which constitute primary endpoints in oncology and rare disease trials, but are difficult to model under censoring and small sample sizes. Existing generative approaches, largely GAN-based, are data-hungry, unstable, and rely on strong assumptions such as independent censoring. We introduce a variational autoencoder (VAE) that jointly generates mixed-type covariates and survival outcomes within a unified latent variable framework, without assuming independent censoring. Across synthetic and real trial datasets, we evaluate our model in two realistic scenarios: (i) data sharing under privacy constraints, where synthetic controls substitute for original data, and (ii) control-arm augmentation, where synthetic patients mitigate imbalances between treated and control groups. Our method outperforms GAN baselines on fidelity, utility, and privacy metrics, while revealing systematic miscalibration of type I error and power. We propose a post-generation selection procedure that improves calibration, highlighting both progress and open challenges for generative survival modeling.
Introduction
Late-phase clinical trials, especially in oncology and rare diseases, are notoriously slow and expensive due to fragmented patient populations. One promising solution is generating synthetic control arms using AI to augment small control groups or enable privacy-preserving data sharing. However, the primary endpoints in these trials are usually "time-to-event" or survival outcomes. Modeling survival data is complex because of censoring—when a patient drops out or the trial ends before the event occurs. Existing generative models for tabular data are often data-hungry, unstable, and rely on the unrealistic assumption that censoring is entirely independent of a patient's covariates. Furthermore, current models are typically evaluated on standard machine learning (ML) fidelity metrics, leaving it unclear if the synthetic data holds up in rigorous downstream statistical testing.
Methods: A Unified VAE Framework
We propose extending the Heterogeneous and Incomplete Variational Autoencoder (HI-VAE) to jointly generate mixed-type patient covariates and survival outcomes within a unified latent space. Unlike modular pipelines that generate covariates first and then predict survival times, this VAE jointly learns the distributions of covariates, event times, and censoring times, removing the flawed "independent censoring" assumption.
We implemented two variants for the time-to-event density: HI-VAE_Weibull (parametric Weibull distribution) and HI-VAE_piecewise (discretized time axis). Because generative models struggle to natively calibrate statistical power, we introduce a post-generation selection safeguard. We generate a large pool of candidate synthetic datasets and select the "best" one by running a log-rank test to find the dataset whose survival distribution most closely matches the original control arm.
Key Results
Fidelity, Utility, and Privacy
Across both simulated and real-world Phase III clinical trial datasets (e.g., ACTG 320, NCT00119613), the HI-VAE variants consistently outperformed GAN-based baselines. They achieved lower divergence from original distributions (J-S distance) and better preserved survival curve patterns. Regarding privacy, the model achieved K-map scores of 4 to 6, providing enough privacy for internal trial augmentation or controlled-access data sharing, though falling short of the strict thresholds required for open public release.
Statistical Validity: Type I Error and Power
A critical insight of this work is that good ML metrics do not guarantee valid clinical statistics. A synthetic dataset can score perfectly on fidelity yet fail to preserve Type I error and statistical power in downstream analysis. Without intervention, all models skewed statistical power. However, applying the post-generation selection procedure allowed the HI-VAE models to closely reproduce theoretical target power, such as mimicking a 3x control arm augmentation.
Discussion and Future Work
While the proposed framework represents a significant step forward, the post-generation selection procedure is currently a necessary patch. The ultimate goal is developing "calibration-aware" training strategies that bake statistical error control directly into the model's loss function. Additionally, future work will explore stronger privacy-preserving architectures, such as calibrated differential privacy or diffusion models, tailored specifically for the nuances of survival data.
Summary updated