OncoSynth: Synthetic data generation for treatment effect estimation in oncology

Abstract: In oncology, access to patient-level data is often restricted. Synthetic data provides an alternative for analyzing treatment effectiveness, but existing methods for synthetic data generation fail to preserve the causal relationships between covariates, treatments, and outcomes, thereby leading to biased estimates of treatment effects. Here, we introduce OncoSynth, a generative, causally-aware machine learning framework designed to produce synthetic cohorts that enable accurate estimation of population- and patient-level treatment effects. OncoSynth uses a diffusion-based sequential approach to model how covariates influence treatment assignment and how treatment affects survival. We evaluate OncoSynth using large lung (N = 37,128) and breast cancer (N = 17,046) cohorts. Our results show that OncoSynth generates high-fidelity synthetic patient cohorts that preserve real-world patient, treatment, and outcome distributions. Notably, OncoSynth improves treatment effect estimation over existing approaches, by reducing population-level treatment effect error by up to 66%, and patient-level treatment effect error by up to 58%. Thereby, OncoSynth supports reliable evidence generation for precision oncology in settings where data sharing is restricted.

• The paper introduces OncoSynth, a framework that uses causal decomposition to sequentially generate synthetic oncology patient data, reducing bias in treatment effect estimation.
• It employs a TabDiff diffusion model for covariate generation, logistic regression for treatment assignment, and random survival forests for outcomes, achieving up to 66% reduction in ATE error.
• OncoSynth outperforms models like CTGAN by preserving clinical temporal ordering, offering robust and reproducible surrogates for precision oncology research.

Synthetic Data Generation for Causal Treatment Effect Analysis in Oncology: The OncoSynth Framework

Introduction

OncoSynth proposes a causally-aware generative framework for synthetic patient data in oncology, explicitly designed to enable robust estimation of treatment effects under the constraints of limited data accessibility. Unlike prior generative models that focus on replicating population-level statistical distributions but disregard causality, OncoSynth operationalizes the causal chain of clinical events: covariate-driven treatment assignment and subsequent treatment-driven outcomes. This approach addresses inherent data-sharing and reproducibility barriers in oncology and offers reliable surrogates for real-world clinical and research tasks.

Methodological Innovations

Causal Decomposition and Sequential Generation

OncoSynth achieves its fidelity and utility by sequentially factorizing the joint data distribution. Patient covariates are generated first using a tabular diffusion model (TabDiff), followed by treatment assignment modeled via a calibrated logistic regression classifier conditioned on covariates. Finally, outcomes are synthesized using random survival forests conditioned on both covariates and treatment status, with outcomes defined in terms of event time and censoring indicators. This explicit factorization mirrors the real-world clinical temporal ordering, unambiguously preventing information leakage from outcome to treatment assignment, which is a documented source of bias in joint generative strategies.

Comparison with Standard Approaches

Baseline models, including CTGAN and conventional TabDiff, jointly learn all variables, thus implicitly violating causal ordering and enabling spurious dependencies. OncoSynth, by maintaining modularity and sequential conditioning, differentiates itself not only architecturally but also in its capacity to produce clinically reliable synthetic data for downstream causal analyses. Any improvements attained over TabDiff are attributable exclusively to OncoSynth's causal decomposition.

Empirical Evaluation

Cohort Description and Data Accessibility

The evaluation utilized two large, publicly available cohorts from the SEER cancer registry: lung cancer (N = 37,128, radiotherapy as binary treatment) and breast cancer (N = 17,046, adjuvant versus neoadjuvant chemotherapy). Both datasets included multidimensional demographic, clinical, and outcome variables, with survival measured as time-to-event under censoring. The use of SEER ensures reproducibility and standardized benchmarking.

Fidelity Metrics

OncoSynth consistently surpassed CTGAN and TabDiff across all fidelity metrics: lower univariate distribution distance ($A_X$), lower deviation in covariate pairs ($A_{X^2}$), more accurate preservation of treatment prevalence ($A_W$) and event/censoring rates ($A_C$), and closer alignment of survival time distributions (Jensen-Shannon distance, $JSD_T$) as well as restricted mean survival time (RMST) across several follow-up horizons. For instance, in the lung cohort, OncoSynth reduced distribution distance across covariates by nearly 60% (0.031 vs. 0.084 for CTGAN, 0.046 for TabDiff), and RMST deviation at 3 years was 0.23 (vs. 2.87 for CTGAN, 0.62 for TabDiff).

Clinical Utility and Causal Estimation

OncoSynth demonstrated superior recovery of treatment effect estimates, both at a population level (ATE) and at an individual level (ITE). The framework reduced ATE error by up to 66% (lung: 0.246 vs. 0.727 for CTGAN, 1.883 for TabDiff) and ITE error by up to 58% (lung: 0.447 vs. 1.075 for CTGAN, 1.864 for TabDiff). Calibration, discrimination (AUROC), and policy utility (AUQC under Qini curve) were consistently better, confirming that the synthetic data not only replicated statistical distributions but also upheld the causal mechanisms required for effective treatment allocation and personalization.

Robustness and Sensitivity

Experiments, repeated on independent randomized splits and assessed across horizons (3/5/7/10 years), established that OncoSynth's results were stable and robust, with treatment effect estimates closely matching those in real data across all settings.

Implications and Future Directions

Practical Impact

OncoSynth's capacity to induce high-fidelity, causally-structured synthetic cohorts directly supports precision oncology tasks where access to patient-level data is restricted, including independent validation, protocol development, and multi-site evidence generation. The implementation provides an open-source foundation for standardized benchmarking and adaptation to other endpoints (e.g., progression, toxicity) and domains (non-oncology clinical areas). Importantly, OncoSynth's modularity accommodates extensions for additional covariates, longitudinal settings, and non-survival outcomes.

Theoretical and Methodological Relevance

The explicit demonstration that joint generative models fail to preserve causal relationships is significant. OncoSynth's sequential conditioning represents a paradigm shift toward generative models that maintain interpretability and causal validity for downstream causal inference. Its performance gains highlight the necessity of respecting the data-generating process structure for unbiased effect estimation.

Limitations and Considerations

Dependence on underlying data quality, inherent biases, and the inability to address unmeasured confounding remain core limitations. Further, while synthetic data provides privacy gains, complementary safeguards (e.g., differential privacy overlays) may still be required in sensitive settings. The reliability of effect estimates ultimately hinges on the inclusion of relevant effect modifiers and confounders in the original data.

Conclusion

OncoSynth offers a causally-aware, modular generative framework for synthetic data in oncology, enabling accurate population- and patient-level treatment effect analyses from data-limited settings. Through rigorous empirical validation, it has been shown to outperform existing generative models in both statistical fidelity and causal utility, supporting precision oncology research and clinical evidence generation. The approach establishes new methodological standards for synthetic data-driven causal inference and is readily extensible to broader clinical applications (2606.25762).