🤖 machine learning

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 privacy restrictions, significantly improving the accuracy of both population- and patient-level treatment effect estimation compared to existing methods.

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

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://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

Imagine you are a chef trying to perfect a new recipe for a life-saving medicine. To do this, you need to test it on thousands of patients to see who gets better and who doesn't. However, there's a huge problem: you can't actually use real patients for these tests because their medical records are private, locked away for legal and ethical reasons. It's like trying to learn how to bake a cake without ever being allowed to touch the real ingredients or see the real bakers.

This is where OncoSynth comes in. Think of OncoSynth as a "digital twin" factory that builds a perfect, fake version of a patient population. But here's the catch: most existing factories make "fake" patients that look real on the surface (they have the right age, weight, and tumor size) but fail to understand the story of how they got sick and how they were treated. They mix up the timeline, like a movie where the ending is shown before the beginning, which leads to wrong conclusions about whether a treatment actually works.

The Problem with Old Methods

Imagine you are trying to learn how a specific type of rain (treatment) affects a garden (patient outcome).

Old methods are like taking a snapshot of the garden, the rain, and the flowers all at once and shoving them into a blender. The blender spits out a smoothie that looks like a garden, but it has lost the logic that rain causes flowers to grow. Because the blender mixed everything together, it might accidentally learn that "flowers cause rain," which is impossible. In medical terms, this leads to biased results where a drug might look effective when it's not, or vice versa.

How OncoSynth Works

OncoSynth is different because it acts like a chronological storyteller rather than a blender. It builds the fake patients step-by-step, strictly following the order of real life:

Step 1: The Patient Profile. First, it generates a patient's background (age, genetics, tumor size). This is the "before" picture.
Step 2: The Doctor's Decision. Next, it decides what treatment that specific patient would have received based on their profile. It asks, "Given this patient's age and tumor, what would a real doctor have prescribed?"
Step 3: The Outcome. Finally, it simulates what happens to that patient after receiving that specific treatment. It asks, "Given this patient and this treatment, how long do they survive?"

By keeping these steps separate and in order, OncoSynth ensures that the "fake" patients have the same logical cause-and-effect relationships as real people. The treatment doesn't magically change the patient's age, and the outcome doesn't influence the doctor's past decision.

The Results: A Better "Fake" World

The researchers tested this system using real data from two massive groups of patients: one with lung cancer (over 37,000 people) and one with breast cancer (over 17,000 people). They compared OncoSynth against the best existing "blender" methods.

The Look: OncoSynth created fake patients that looked statistically identical to real ones. The distribution of ages, tumor sizes, and survival times matched the real world almost perfectly.
The Logic (The Big Win): The real magic happened when they tried to measure how well the fake data could predict treatment success.
- When estimating the average benefit of a treatment for a whole population, OncoSynth reduced the error by up to 66% compared to other methods.
- When estimating the personalized benefit for a specific individual (precision medicine), it reduced the error by up to 58%.

Why This Matters

In the world of oncology (cancer research), knowing exactly how a treatment works for different types of people is crucial. But because real data is so hard to share, researchers often can't run these tests.

OncoSynth provides a solution: a safe, synthetic playground. It allows researchers to generate a "fake" dataset that is so faithful to reality—not just in appearance, but in the causal logic of how treatments work—that they can run their experiments, figure out which drugs work best for whom, and generate reliable evidence, all without ever seeing a single real patient's private record.

In short, OncoSynth doesn't just make fake patients; it makes fake patients that think and react like real ones, allowing scientists to learn how to save lives without breaking privacy rules.

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 these data is frequently restricted due to governance, legal, and ethical constraints, as well as the risk of patient re-identification. While synthetic data generation offers a potential solution by creating artificial datasets that reproduce the statistical properties of real patient data, 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., Variational Autoencoders, GANs, and diffusion models like TabDiff) typically model covariates, treatment assignment, and outcomes jointly. This approach ignores the chronological and causal ordering of clinical events. Consequently, these models may learn spurious dependencies where future outcomes influence past treatment assignments (information leakage), leading to systematically biased estimates of treatment effects. While such data may suffice for predictive tasks like risk scoring, it is unreliable for causal inference tasks, such as estimating Average Treatment Effects (ATE) or Individualized Treatment Effects (ITE), which are essential for evaluating therapeutic strategies.

Methodology: OncoSynth Framework

To address these limitations, the authors introduce OncoSynth, a generative, causally-aware machine learning framework designed specifically for oncology. OncoSynth employs a diffusion-based sequential approach that mirrors the clinical data-generating process by decomposing the generation into three distinct, chronologically ordered steps:

Patient Covariate Generation: The framework first generates patient characteristics (e.g., age, tumor size, receptor status) using a state-of-the-art tabular diffusion model (TabDiff). This step captures the joint distribution of patient covariates ( $\hat{P}_X$ ).
Treatment Assignment Modeling: Conditional on the generated covariates, the framework models the treatment assignment mechanism ( $\hat{P}_{W|X}$ ). This is implemented using a binary classifier (logistic regression with isotonic regression calibration) to predict the probability of treatment based on patient characteristics, ensuring the synthetic data reflects realistic treatment allocation patterns.
Outcome Generation: Finally, survival outcomes are generated conditional on both the covariates and the assigned treatment ( $\hat{P}_{Y|X,W}$ ). This step utilizes Random Survival Forests (RSFs) within a T-learner framework to model time-to-event data, accounting for both the observed survival time and censoring status.

By factorizing the joint distribution as $\hat{P}_{X,W,Y} = \hat{P}_X \cdot \hat{P}_{W|X} \cdot \hat{P}_{Y|X,W}$ , OncoSynth ensures that information flows only from past to future, preventing the temporal leakage that plagues joint generation models.

Experimental Setup

The framework was evaluated using two large, real-world oncology cohorts from the Surveillance, Epidemiology, and End Results (SEER) cancer registry:

Lung Cancer Cohort: $N = 37,128$ patients diagnosed between 2004 and 2015. Treatment was defined by the receipt of radiotherapy.
Breast Cancer Cohort: $N = 17,046$ female patients with Stage IIIA-IIIB breast cancer diagnosed between 2007 and 2015. Treatment was defined by therapy sequencing (adjuvant vs. neoadjuvant chemotherapy).

The evaluation compared OncoSynth against two baselines: CTGAN (a GAN-based method) and TabDiff (a diffusion-based method). Notably, OncoSynth shares the same underlying diffusion architecture as TabDiff but differs in its sequential, causally-aware generation strategy. Performance was assessed across five independent runs with different train-test splits.

Key Results

1. Statistical Fidelity

OncoSynth demonstrated superior fidelity in reproducing the statistical properties of the original data compared to baselines:

Patient Characteristics: It achieved the lowest distribution distances for univariate covariates (e.g., lung cancer $\Delta X = 0.031$ vs. TabDiff $0.046$) and covariate co-dependencies.
Treatment and Event Prevalence: OncoSynth closely matched the treatment prevalence and event/censoring rates of the original data, significantly outperforming CTGAN and TabDiff.
Survival Outcomes: Kaplan-Meier curves generated by OncoSynth showed near-perfect overlap with original survival trajectories. It achieved the lowest Jensen-Shannon distance for survival time distributions and the smallest deviations in Restricted Mean Survival Time (RMST) across multiple horizons (3, 5, 7, and 10 years).

2. Clinical Utility (Treatment Effect Estimation)

The primary contribution of OncoSynth is its ability to support accurate downstream causal inference:

Treatment Assignment Mechanism: Propensity scores learned from OncoSynth data closely matched those from the original data, with significantly lower distributional deviation and patient-level error compared to baselines.
Population-Level Treatment Effect (ATE): OncoSynth reduced the error in ATE estimation by up to 66% compared to existing methods. For instance, in lung cancer, the ATE error was $0.246$ for OncoSynth versus $1.883$ for TabDiff.
Patient-Level Treatment Effect (ITE): OncoSynth improved the precision of estimating heterogeneous effects (PEHE) by up to 58%. It successfully captured patient-specific variations in treatment benefit, whereas baselines exhibited significant deviations.
Treatment Allocation Policies: Policies derived from OncoSynth data achieved higher Area Under the Qini Curve (AUQC) values, indicating more effective prioritization of patients who would benefit most from treatment.

Significance and Claims

The paper claims that OncoSynth provides a causally-aware ML framework that overcomes the limitations of existing synthetic data generators in oncology. By explicitly modeling the causal chain of clinical events, OncoSynth generates high-fidelity synthetic cohorts that preserve not only statistical distributions but also the underlying causal mechanisms of treatment assignment and outcomes.

The authors assert that this framework enables reliable evidence generation for precision oncology in settings where data sharing is restricted. Specifically, OncoSynth supports:

Accurate Treatment Effect Estimation: Providing unbiased estimates of ATE and ITE that align with real-world data.
Clinical Decision Support: Enabling the derivation of treatment allocation policies that effectively target patient subgroups.
Reproducibility: Offering a method to generate synthetic datasets that can be used for independent validation and reproducibility without exposing raw patient information.

The study concludes that while synthetic data cannot replace the need for rigorous data curation and may inherit biases from the source data, OncoSynth offers a practical foundation for generating real-world evidence when direct access to patient-level records is limited.