SurvivalGAN: GAN-Based Survival Data Generation

• SurvivalGAN is a family of GAN-based techniques that integrate censoring and time-to-event structures to generate synthetic survival data.
• It employs modular architectures combining conditional generation, survival function estimation, and time regression to accurately mimic risk-set dynamics.
• Empirical evaluations on diverse clinical datasets demonstrate that SurvivalGAN preserves survival curves and outperforms baselines in calibration and discrimination.

SurvivalGAN denotes GAN-based approaches that incorporate censoring and time-to-event structure into generative or representation learning for survival analysis. The term most specifically names the modular synthetic survival-data generator introduced in "SurvivalGAN: Generating Time-to-Event Data for Survival Analysis" (Norcliffe et al., 2023), but adjacent literature also uses it in a broader sense for adversarial conditional survival estimation from right-censored data (Zhou et al., 2022), for StyleGAN2 latent embeddings coupled to CoxPH and DeepSurv for cancer prognosis from portrait photographs (Hagag et al., 2023), and for a Synthcity implementation used to synthesize heart-failure cohorts with right-censored outcomes (Puttanawarut et al., 4 Sep 2025). Across these usages, the common statistical objects are the event time $T$, censoring time $C$, observed time $Y = \min(T, C)$, and event indicator $\delta = 1[T \le C]$.

1. Conceptual scope and survival-analysis setting

In the original SurvivalGAN formulation, the motivating problem is synthetic generation of censored survival data rather than direct estimation of a single predictive risk score. Survival data are defined by an event time $T$, a censoring time $C$, the observed time $Y = \min(T, C)$, and the event indicator $\delta = 1[T \le C]$; the risk set at time $t$ is $R(t) = \{i : Y_i \ge t\}$, the survival function is $C$0, and the hazard satisfies $C$1 with $C$2 (Norcliffe et al., 2023). The stated difficulty is that censoring and skewed time horizons induce failure modes that generic tabular generators do not address: misrepresentation of the censored-versus-event proportion, collapse toward short horizons, and distortion of survival-curve shape.

A persistent source of ambiguity is that the label does not denote a single invariant architecture across the literature. In one line of work, SurvivalGAN is a modular generator of the full joint distribution over covariates, time, and event status. In another, it is a "SurvivalGAN-style" adversarial estimator of the conditional observable distribution $C$3, from which Kaplan-Meier and Nelson-Aalen estimators are recovered. In a third, it refers to survival modeling on GAN-derived latent codes rather than direct generation of time-to-event outcomes. This suggests that the stable semantic core of the term is not a particular network topology, but the use of adversarial or GAN-derived machinery to preserve survival structure under censoring.

2. Original SurvivalGAN architecture and training mechanics

The architecture introduced in (Norcliffe et al., 2023) is explicitly modular. It combines a conditional tabular GAN for covariates, a survival-function estimator, and a dedicated time-to-event/censoring regressor. The method models the full joint $C$4 by handling covariates $C$5 separately from the time-event pair $C$6, assumes independent censoring, and uses the empirical censoring ratio during generation via a user-specified or learned sampler.

The tabular component begins with an encoder. Continuous features are represented by a per-feature Gaussian Mixture Model with $C$7 components; each value is encoded as a one-hot component identity together with a normalized residual $C$8 within the selected component. Categorical features are one-hot encoded. A ClassEncoder provides a conditioning variable $C$9 built from one-hot component identities and related cohort conditions. The generator $Y = \min(T, C)$0 receives $Y = \min(T, C)$1 with $Y = \min(T, C)$2, while the discriminator $Y = \min(T, C)$3 receives $Y = \min(T, C)$4 for encoded covariates $Y = \min(T, C)$5. Training uses WGAN-GP with

$Y = \min(T, C)$6

and

$Y = \min(T, C)$7

where $Y = \min(T, C)$8 and $Y = \min(T, C)$9.

The survival-function component is implemented with DeepHit, used to estimate $\delta = 1[T \le C]$0 over $\delta = 1[T \le C]$1 discrete horizons. The time module is separate: an XGBoost regressor takes covariates $\delta = 1[T \le C]$2, the event label $\delta = 1[T \le C]$3, and the vector $\delta = 1[T \le C]$4, and predicts a continuous time $\delta = 1[T \le C]$5. Its training loss is mean squared error on $\delta = 1[T \le C]$6. This separation is central to the paper’s argument: the GAN is responsible for high-fidelity covariates, the survival model for global temporal structure, and the regressor for translating survival probabilities into a continuous event or censoring time.

Generation follows the same decomposition. The model samples $\delta = 1[T \le C]$7 with an imbalanced sampler matched to empirical censoring ratios and optionally reweighted horizon bins, draws $\delta = 1[T \le C]$8, generates encoded covariates, decodes them to $\delta = 1[T \le C]$9, evaluates $T$0 across the horizon grid, and then predicts $T$1 with the regressor. The output is a synthetic triplet $T$2. Recommended hyperparameters in the appendix include a 3-layer generator of width about $T$3 with $T$4, a 2-layer discriminator of width about $T$5 with Leaky ReLU, dropout around $T$6, learning rate around $T$7, weight decay around $T$8, batch size around $T$9, WGAN-GP $C$0, $C$1 durations for DeepHit, and about $C$2 estimators with depth around $C$3 for the XGBoost time regressor.

3. Survival-specific failure modes, metrics, and empirical results

A defining contribution of the original paper is the formalization of three survival-specific failure modes: generating too few at-risk members, generating too many at-risk members, and censoring too early (Norcliffe et al., 2023). These are operationalized through the relationship between synthetic and real Kaplan-Meier curves. Over-optimism corresponds to $C$4 for many $C$5, over-pessimism to $C$6, and short-sightedness to synthetic horizons ending early, $C$7.

The paper introduces three metrics derived from these failures. Optimism is the signed area difference between curves,

$C$8

with $C$9. Short-Sightedness is the relative horizon deficit,

$Y = \min(T, C)$0

and KM Divergence is the mean absolute curve discrepancy,

$Y = \min(T, C)$1

The paper further gives bounds linking Optimism to total variation, KL divergence, and Hellinger distance. In interpretation, Optimism measures aggregate risk-set inflation or deflation, Short-Sightedness measures premature truncation of follow-up, and KM Divergence measures overall survival-curve fidelity even when signed bias is small.

Empirical evaluation spans AIDS, CUTRACT, PHEART, SEER, and METABRIC, against ADS-GAN, CTGAN, TVAE, PrivBayes, and Normalizing Flows. The reported downstream setting is train on synthetic, test on real, with CoxPH, Survival XGBoost, Random Survival Forest, and DeepHit as evaluators. SurvivalGAN is reported to match or outperform baselines in Jensen-Shannon and Wasserstein distance on covariates, to exhibit near-zero Optimism and low KM Divergence with no extreme Short-Sightedness, and to improve downstream discrimination and calibration relative to the synthetic baselines. Representative C-index results are $Y = \min(T, C)$2 on AIDS versus $Y = \min(T, C)$3 on original data, $Y = \min(T, C)$4 on CUTRACT versus $Y = \min(T, C)$5, $Y = \min(T, C)$6 on PHEART versus $Y = \min(T, C)$7, $Y = \min(T, C)$8 on SEER versus $Y = \min(T, C)$9, and $\delta = 1[T \le C]$0 on METABRIC versus $\delta = 1[T \le C]$1. Ablations attribute specific roles to each module: removing the time regressor increases Short-Sightedness and lowers C-index, removing imbalanced sampling degrades C-index and Optimism, removing temporal sampling induces strong pessimism, and removing the conditional GAN worsens Brier scores and Optimism.

4. Adversarial nonparametric estimation of conditional survival

A broader "SurvivalGAN-style" construction appears in "Deep Generative Survival Analysis: Nonparametric Estimation of Conditional Survival Function" (Zhou et al., 2022). Here the goal is not synthetic cohort release but nonparametric estimation of the conditional survival function $\delta = 1[T \le C]$2, the hazard $\delta = 1[T \le C]$3, and the cumulative hazard $\delta = 1[T \le C]$4 with right-censored data, without semiparametric structure such as Cox proportional hazards.

The key move is to learn a conditional generator for the observable pair $\delta = 1[T \le C]$5 given covariates $\delta = 1[T \le C]$6 by adversarial training. By the noise outsourcing lemma, there exists a measurable map

$\delta = 1[T \le C]$7

with $\delta = 1[T \le C]$8 independent of $\delta = 1[T \le C]$9, such that $t$0. In practice, $t$1 is a feedforward neural network with two outputs: $t$2 for time and $t$3 for the censoring indicator via a sigmoid, with $t$4 at inference. Training uses a WGAN objective over triples $t$5,

$t$6

with an empirical gradient penalty to enforce the $t$7-Lipschitz critic.

Once the conditional generator is learned, the method draws $t$8 i.i.d. synthetic samples $t$9 at a fixed covariate value $R(t) = \{i : Y_i \ge t\}$0 and then applies classical nonparametric estimators to the synthetic cohort. The generated-sample Kaplan-Meier estimator is

$R(t) = \{i : Y_i \ge t\}$1

and the Nelson-Aalen estimator is

$R(t) = \{i : Y_i \ge t\}$2

An optional smoothed hazard is obtained by kernel smoothing of increments of $R(t) = \{i : Y_i \ge t\}$3. The paper’s theoretical contribution is a plug-in consistency result: if the learned conditional generator converges to the true joint law in bounded-Lipschitz distance, then the induced conditional hazard and survival estimators converge uniformly on $R(t) = \{i : Y_i \ge t\}$4; with synthetic cohort size $R(t) = \{i : Y_i \ge t\}$5, the sample-based KM and NA estimators attain the uniform rate $R(t) = \{i : Y_i \ge t\}$6.

The reported experiments include four simulation models covering proportional hazards and accelerated failure time regimes with covariate-independent and covariate-dependent censoring, using $R(t) = \{i : Y_i \ge t\}$7 and $R(t) = \{i : Y_i \ge t\}$8 replications. The method generally matches ground-truth survival and quantiles across nonlinear and AFT scenarios; Cox PH is superior in correctly specified proportional-hazards settings but degrades under nonlinear or AFT violations. Applications to PBC and SUPPORT emphasize prediction intervals rather than C-index or integrated Brier score. On PBC, with $R(t) = \{i : Y_i \ge t\}$9 subjects after cleaning and censoring rate about $C$00, intervals for uncensored subjects covered $C$01 true survival times. On SUPPORT, with $C$02 patients and $C$03 predictors, the method yielded narrower, more calibrated prediction intervals for shorter survival times and avoided the lower-bound issues encountered by PH under censoring for long survival times.

5. StyleGAN-derived latent prognosis from portrait photographs

A distinct usage of the term appears in "Deep Learning for Cancer Prognosis Prediction Using Portrait Photos by StyleGAN Embedding" (Hagag et al., 2023). The task is overall survival prediction for cancer patients from routine $C$04D portrait photographs taken at radiotherapy admission. The cohort comprises $C$05 cancer patients with $C$06 portrait photos; the demographic distribution is $C$07 male and $C$08 female, survival times for uncensored patients range from $C$09 to $C$10 days, and $C$11 of cases are right-censored. Faces are detected and aligned with dlib and cropped to $C$12 resolution to minimize confounding from clothing and background.

The generative component is a StyleGAN2 generator pre-trained on FFHQ at $C$13 and then fine-tuned on the institutional portrait dataset using adaptive discriminator augmentation. Fine-tuning on a single NVIDIA Quadro RTX 8000 (48 GB) GPU for $C$14 hours reaches an FID of $C$15 after $C$16k images. The adapted model reproduces domain-specific accessories such as surgical masks and bouffant caps that the original FFHQ model failed to synthesize. Each photograph is then inverted into the native StyleGAN2 latent space $C$17 by solving

$C$18

where $C$19 is Euclidean distance in a pretrained VGG-16 feature space. Optimization starts from the average face latent vector $C$20 and uses $C$21 Adam steps with learning rate $C$22, $C$23, $C$24, and $C$25. The result is a single $C$26-dimensional latent vector per face.

These latent vectors are used directly as covariates in survival models. The paper studies CoxPH,

$C$27

and DeepSurv,

$C$28

with a 6-layer MLP, batch normalization, dropout of $C$29, Adam with learning rate $C$30, and batch size $C$31. Performance is reported with the concordance index and Brier score, with the censoring distribution estimated by Kaplan-Meier for time-dependent metrics. StyleGAN2 latent embeddings plus DeepSurv achieve a C-index of $C$32 and a mean Brier score of $C$33. End-to-end CNN baselines on portrait pixels are markedly worse, with C-indices $C$34 for ResNet-18 and $C$35 for VGG-16, both with Brier scores of $C$36. Clinical baselines remain strong: CoxPH on clinical covariates reaches C-index $C$37 with Brier $C$38, and DeepSurv on clinical covariates reaches $C$39 with Brier $C$40. The best performance is obtained by early fusion of clinical covariates and latent vectors, yielding C-index $C$41 and Brier $C$42.

Interpretability is built around a Cox-derived "health attribute" in latent space. If $C$43 is the CoxPH coefficient vector on $C$44, then the log-risk is $C$45, the scalar health score is $C$46, and latent traversal is defined by $C$47. The syntheses associated with negative $C$48 show fuller cheeks, reddish skin colorit, increased muscle tone, and more positive affect; positive $C$49 yields loss of subcutaneous fat, paler skin, drooping of the nasolabial fold, reduced muscle tone, and "tippiness" of the nose. The authors report that these changes are localized to facial regions rather than clothing or background. They also construct an age attribute by linear regression from $C$50 to age and report that its traversals differ from the health direction. Confidence intervals and formal statistical significance testing are not reported, and ethnicity and fairness analyses are described as open concerns.

6. Heart-failure data synthesis, privacy assessment, and recurrent limitations

A later deployment appears in "Synthetic Survival Data Generation for Heart Failure Prognosis Using Deep Generative Models" (Puttanawarut et al., 4 Sep 2025), where SurvivalGAN is one of five generators used to synthesize a heart-failure cohort of $C$51 patients. In that study, SurvivalGAN is "a GAN architecture tailored for survival data" implemented through the Synthcity library with default hyperparameters, and it generates complete synthetic rows containing covariates, time-to-event $C$52 ("Days"), and right-censoring indicator $C$53 ("dead"). Preprocessing includes adding missing-value indicator columns, MICE imputation, clipping continuous variables to clinically plausible ranges, post-generation logical and clinical constraint filtering, and reintroduction of the original missingness patterns.

The reported evaluation emphasizes fidelity, predictive utility, and privacy. All models exceed $C$54 in dimension-wise and column-wise correlation metrics, with SurvivalGAN among the top group. Before post-processing, SurvivalGAN shows divergence in KM curves because of mismatch in the synthetic "Days" distribution. An ablation then histogram-equalizes synthetic "Days" to match the empirical distribution of the real training set; this improves KM-curve alignment and IBS while preserving C-index and privacy metrics. SurvivalGAN achieves the highest recall of significant univariate CoxPH features with consistent direction at $C$55, with precision $C$56. In train-on-synthetic, test-on-real evaluation, its C-indices range from $C$57 to $C$58 across CoxPH, DeepHit, DeepSurv, and RSF, against real-data baselines of $C$59 to $C$60; after equalization, CoxPH TSTR IBS improves to $C$61, and overall calibration becomes comparable to real and synthetic baselines while discrimination remains about $C$62 to $C$63. Privacy evaluation reports zero exact matches, membership inference attack accuracy around $C$64, attribute inference near $C$65, and nearest-neighbor adversarial accuracy near zero. No formal differential privacy guarantee is claimed.

Across the cited literature, several assumptions recur. Independent censoring is essential for identifiability in the conditional nonparametric estimator and is also assumed in the original synthetic-data generator (Zhou et al., 2022). The original SurvivalGAN is not differentially private, and the heart-failure deployment likewise relies on empirical privacy auditing rather than formal privacy mechanisms (Norcliffe et al., 2023). The portrait-based prognosis model is retrospective, uses an $C$66 split without external validation, and leaves ethnicity, fairness, and multicenter generalization unresolved (Hagag et al., 2023). Taken together, these papers suggest that SurvivalGAN is best understood as a family of survival-aware generative strategies whose value depends on how well they preserve censoring structure, risk-set dynamics, calibration, and privacy under the constraints of a specific application.