📄 medicine

OncoSynth: Synthetic data generation for treatment effect estimation in oncology

OncoSynth is a novel, causally-aware diffusion-based framework that generates high-fidelity synthetic oncology cohorts to overcome data access restrictions, significantly improving the accuracy of both population- and patient-level treatment effect estimation compared to existing methods.

Original authors: Stefan Feuerriegel, Octavia-Andreea Ciora, Julian Welzel, Dennis Frauen, Maresa Schröder, Marie Brockschmidt, Harry Amad, Thomas Callender, Mihaela van der Schaar

Published 2026-09-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Stefan Feuerriegel, Octavia-Andreea Ciora, Julian Welzel, Dennis Frauen, Maresa Schröder, Marie Brockschmidt, Harry Amad, Thomas Callender, Mihaela van der Schaar

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the fight against cancer, doctors increasingly rely on vast collections of patient records to understand which treatments work best for whom. These records, often called real-world data, contain details about a patient's age, tumor size, and medical history, alongside the specific therapies they received and how long they survived. However, a significant barrier stands in the way of using this information freely: strict privacy laws and ethical rules prevent researchers from sharing the actual names and details of real patients. This creates a dilemma where valuable medical insights remain locked away because the data cannot be moved between hospitals or research centers. To solve this, scientists have developed a method called synthetic data generation. This process involves using computers to create entirely new, fake patient records that look and behave statistically like the real ones. These artificial records allow researchers to run experiments and test ideas without ever exposing a single real person's private information.

The challenge with existing methods of creating these fake records is that they often treat the data like a simple list of numbers to be copied, ignoring the cause-and-effect relationships that govern real medicine. In a real hospital, a doctor decides on a treatment based on a patient's specific condition, and that treatment then influences the patient's survival. If a computer model generates the treatment and the survival outcome at the same time, it might accidentally learn that the outcome caused the treatment, a logical impossibility that leads to misleading conclusions about how well a drug works. This flaw means that while older methods can create realistic-looking patient profiles, they often fail when used to answer the most critical question: does this treatment actually help?

A team of researchers has introduced a new system called OncoSynth, designed specifically to fix this problem in the field of oncology. Instead of generating all the data at once, OncoSynth follows the natural timeline of a patient's journey. First, it creates a synthetic patient with specific characteristics, such as age and tumor type. Next, it uses those characteristics to decide which treatment that synthetic patient would logically receive, just as a real doctor would. Finally, it determines the patient's survival outcome based on the treatment they received and their initial condition. By strictly following this chronological order, the system preserves the causal chain of events, ensuring that the fake data reflects how treatments truly affect survival.

The researchers tested this approach using two massive datasets from the United States: one containing records of 37,128 patients with lung cancer and another with 17,046 patients with breast cancer. They compared OncoSynth against other leading methods for creating synthetic data. The results showed that OncoSynth produced fake patient groups that were not only statistically similar to the real ones but also far superior at preserving the relationships between treatment and survival. When the researchers used the fake data to estimate how effective a treatment was for the general population, the error rate dropped by up to 66 percent compared to other methods. When they tried to predict which specific individuals would benefit most from a therapy, the error rate fell by up to 58 percent.

This improvement is crucial because it means that researchers can now use these synthetic datasets to generate reliable evidence about cancer treatments without needing access to the original, private records. The study demonstrated that the artificial data could accurately reproduce complex medical patterns, such as how certain tumor sizes influence the choice of chemotherapy or how different treatment sequences affect long-term survival. For instance, in the lung cancer group, the system correctly identified that radiation therapy was associated with better survival outcomes, and in the breast cancer group, it accurately reflected the differences in survival between patients who received chemotherapy before surgery versus those who received it after.

The researchers emphasize that while the system is highly effective, it is not a magic replacement for real-world trials. The quality of the fake data depends entirely on the quality of the real data it learns from; if the original records are incomplete or biased, the synthetic versions will reflect those same issues. However, the study confirms that this new approach offers a powerful tool for precision medicine. By allowing scientists to analyze treatment effects in a safe, privacy-preserving environment, OncoSynth helps unlock the potential of large medical databases, enabling the generation of clinical evidence that could lead to better, more personalized care for cancer patients everywhere.

Technical Summary: OncoSynth

Problem Statement

In precision oncology, the generation of clinical evidence relies heavily on large-scale patient-level datasets, such as cancer registries and electronic health records. However, access to such data is frequently restricted due to governance, legal, and ethical concerns regarding patient re-identification. While synthetic data generation offers a potential solution by creating artificial datasets that mimic statistical properties without exposing raw patient information, existing methods suffer from a critical limitation: they fail to preserve the causal relationships between covariates, treatments, and outcomes.

Current state-of-the-art generative models (e.g., VAEs, GANs, and diffusion models like CTGAN and TabDiff) typically model covariates, treatment assignments, and outcomes jointly. This approach ignores the chronological and causal ordering of clinical events. Consequently, these models may allow information from future outcomes to influence past treatment assignments (temporal leakage), creating spurious dependencies. While such methods may suffice for predictive tasks like risk scoring, they produce systematically biased estimates of treatment effects, rendering them unreliable for causal inference and the generation of clinical evidence regarding treatment efficacy.

Methodology: OncoSynth

To address these limitations, the authors introduce OncoSynth, a causally-aware, generative machine learning framework designed specifically for treatment effect estimation in oncology. Unlike existing approaches that model all variables simultaneously, OncoSynth employs a sequential, diffusion-based approach that strictly adheres to the chronological order of clinical events.

The framework decomposes the data-generating process into three distinct, sequential steps:

Patient Covariate Generation: A state-of-the-art tabular diffusion model (TabDiff) generates patient characteristics (e.g., age, tumor size, receptor status) to capture the joint distribution of covariates ( $\hat{P}_X$ ).
Treatment Assignment Modeling: Treatment assignment is modeled conditional on the generated covariates using a binary classifier (logistic regression with isotonic regression calibration). This preserves the treatment assignment mechanism ( $\hat{P}_{W|X}$ ), ensuring that treatment decisions are influenced only by patient characteristics available at the time of decision-making.
Outcome Generation: Survival outcomes are generated conditional on both covariates and treatment using Random Survival Forests (RSFs). This step models the time-to-event outcomes, including censoring status, based on the treatment–outcome mechanism ( $\hat{P}_{Y|X,W}$ ).

Mathematically, OncoSynth factorizes the joint distribution as $\hat{P}_{X,W,Y} = \hat{P}_X \cdot \hat{P}_{W|X} \cdot \hat{P}_{Y|X,W}$ . This factorization ensures that no information from future outcomes leaks into the generation of past treatment assignments, thereby preserving the underlying causal structure.

Key Contributions

Causally-Aware Architecture: The primary contribution is the shift from joint modeling to a sequential, causally-aware generative process that explicitly separates the modeling of covariates, treatment assignment, and outcomes.
High-Fidelity Synthetic Cohorts: The framework generates synthetic patient cohorts that not only reproduce marginal and joint statistical distributions but also preserve the complex clinical mechanisms governing treatment decisions and survival.
Downstream Utility for Causal Inference: OncoSynth is specifically validated for its ability to support downstream causal machine learning tasks, including the estimation of Average Treatment Effects (ATE), Individualized Treatment Effects (ITE), and the derivation of treatment allocation policies.

Experimental Results

The authors evaluated OncoSynth using two large-scale real-world cohorts from the SEER cancer registry:

Lung Cancer: $N = 37,128$ patients (comparing radiotherapy vs. no radiotherapy).
Breast Cancer: $N = 17,046$ patients (comparing neoadjuvant vs. adjuvant chemotherapy).

The framework was benchmarked against two state-of-the-art baselines: CTGAN (GAN-based) and TabDiff (diffusion-based).

Statistical Fidelity

OncoSynth demonstrated superior statistical fidelity across multiple dimensions:

Patient Characteristics: It achieved the lowest distribution distances for univariate covariates and pairwise correlations compared to baselines.
Treatment and Event Rates: It more accurately preserved treatment prevalence and event/censoring rates.
Survival Outcomes: Kaplan-Meier curves and Restricted Mean Survival Time (RMST) deviations showed near-perfect overlap with original data, significantly outperforming baselines in both short-term (3 years) and long-term (10 years) horizons.

Clinical Utility (Treatment Effect Estimation)

The most significant findings relate to the preservation of causal relationships:

Treatment Assignment Mechanism: Propensity scores learned from OncoSynth data closely matched those from original data, with significantly lower calibration errors and distributional deviations than baselines.
Population-Level Treatment Effects (ATE): OncoSynth reduced the error in ATE estimation by up to 66% compared to existing methods. For example, in lung cancer, it correctly identified the positive survival benefit of radiotherapy, while baselines showed larger deviations.
Patient-Level Treatment Effects (ITE): OncoSynth reduced the error in estimating individualized treatment effects by up to 58%. It successfully captured the heterogeneity in treatment benefits across different patient profiles.
Treatment Allocation Policies: Policies derived from OncoSynth data achieved higher Area Under the Qini Curve (AUQC) scores, indicating more effective prioritization of patients who would benefit most from treatment.

Significance and Claims

The paper claims that OncoSynth provides a robust, causally-aware framework for generating synthetic oncology data that overcomes the limitations of current methods. By preserving the causal ordering of events, OncoSynth enables reliable estimation of treatment effects in settings where data sharing is restricted.

The authors assert that their framework supports the generation of real-world evidence for precision oncology, allowing researchers to:

Analyze treatment effectiveness at both population and patient levels without accessing raw patient data.
Identify patient subgroups that derive the most benefit from specific therapies.
Develop and validate treatment allocation policies.

The study concludes that while synthetic data cannot replace the need for rigorous clinical trials, OncoSynth offers a practical foundation for generating clinical evidence and facilitating independent validation when direct access to patient-level data is constrained. The framework is designed to be modular, allowing for the incorporation of additional clinical variables and adaptation to other clinical domains beyond oncology.