Personalized medicine in hematology requires vast amounts of real-world patient data - clinical features, genomic profiles, treatment details, and survival outcomes. Collecting this information across large populations is difficult, especially for rare blood cancers with complex molecular backgrounds.
Real patient datasets often have critical limitations: small sample sizes, missing or incomplete data, and strict privacy regulations that prevent sharing across institutions. These barriers slow down research and delay the development of better treatments.
Synthetic data offers a potential solution. Synthetic data are artificially generated records produced by an algorithm trained on real patient information. The goal is to create data that captures all the statistical properties of the original population without containing any actual patient information.
This study aimed to apply generative artificial intelligence to hematologic cancers - specifically myelodysplastic syndromes (MDS) and acute myeloid leukemia (AML) - to generate synthetic patient cohorts, validate their quality, and test whether they can accelerate medical research.
Generative adversarial networks (GANs) are a class of artificial intelligence that learn to produce realistic synthetic data. A GAN consists of two competing neural networks: a generator that creates fake data, and a discriminator that tries to distinguish real from fake.
Through repeated competition, both networks improve until the generator produces data so realistic that the discriminator cannot reliably tell it apart from real data. The researchers used a specialized type called a conditional Wasserstein GAN with gradient penalty, which handles complex, mixed data types particularly well.
The study included 7,133 patients with MDS or AML from multiple European cohorts. The input data for each synthetic patient included demographics, blood counts, genetic mutations, chromosomal abnormalities, treatments, and survival outcomes.
The researchers also developed a Synthetic Validation Framework (SVF) to rigorously assess two key properties: fidelity (how closely synthetic data resembles real data) and privacy preservability (how little information about real individuals can be recovered from the synthetic data).
The Synthetic Validation Framework evaluated synthetic data quality across multiple dimensions. Clinical synthetic fidelity (CSF) measured how accurately distributions, correlations, and patterns of clinical features were reproduced; genomic synthetic fidelity (GSF) did the same for genetic mutation data. A score of 85% or higher was considered the minimum acceptable threshold for both.
Privacy was assessed using two main metrics. Identical match share (IMS) checked whether any synthetic patient was an exact copy of a real patient. The nearest neighbor distance ratio (NNDR) measured how close synthetic records were to real ones, with optimal values between 0.60 and 0.85 - far enough to protect privacy, close enough to remain statistically useful.
Explainability analysis using SHAP values (Shapley Additive Explanations) allowed researchers to confirm that the same genomic features driving patient classification in real data were also driving classification in synthetic data - a critical check that the AI learned meaningful biology rather than superficial patterns.
Survival analysis using Kaplan-Meier curves and Cox proportional hazards models compared the survival trajectories of real versus synthetic patients across multiple risk categories, ensuring that synthetic patients replicated real-world prognosis accurately.
In the primary experiment (Setting A), 2,043 synthetic MDS patients were generated from an equal number of real patients. The resulting synthetic cohort achieved excellent fidelity scores - CSF of 93% and GSF of 90% - well above the 85% threshold, confirming accurate reproduction of clinical and genomic features.
Privacy was rigorously preserved: the IMS analysis showed zero exact copies of real patients in the synthetic dataset. The NNDR of 0.64 indicated that synthetic records were appropriately distant from real ones, balancing privacy protection with statistical fidelity.
The model generalized well across different data sizes and cancer types. When trained on incomplete data with 30% missing values, or on smaller training sets, it still achieved fidelity scores above 90%. Similar performance was observed in 1,002 synthetic AML patients, demonstrating the model's versatility across different blood cancers.
Survival curves of synthetic MDS patients were statistically indistinguishable from those of real patients across all five IPSS-R risk categories. A Cox model comparing real and synthetic cohorts found no significant difference (p = 0.742), confirming that synthetic patients replicate real prognosis patterns.
In Setting C, the researchers tested whether synthetic data could accelerate translational research. Starting from a 944-patient MDS cohort available in 2014, they generated a 300% augmented synthetic dataset of 2,832 patients.
Using this augmented synthetic cohort, the researchers were able to recapitulate the molecular MDS classification that was only officially defined in 2022 using 2,043 real patients, and the Molecular International Prognostic Scoring System (IPSS-M) that required nearly 3,000 real patients to develop. In other words, synthetic data from 2014 could have predicted discoveries that took years and thousands of patients to confirm.
The same eight disease subgroups identified in the 2022 real-patient study emerged from the synthetic 2014 cohort using Dirichlet process clustering. SHAP explainability analyses confirmed the same genes were driving patient classification in both the real and synthetic cohorts.
A synthetic molecular prognostic score derived from the augmented data identified six risk categories with survival probabilities closely matching those from the real-data IPSS-M. This demonstrates that synthetic data augmentation could have cut years off the timeline for developing precision medicine tools in blood cancer.
In Setting D, the researchers investigated whether synthetic patients could serve as a control arm in clinical trials. They replicated a real multicenter study of 187 MDS patients treated with luspatercept, a drug that stimulates red blood cell production.
A synthetic cohort of 187 patients was generated from the original study population and compared on all primary and secondary trial endpoints, including transfusion independence rates over 8 and 12-week periods and overall survival.
The synthetic cohort matched the real cohort on all key clinical endpoints with no statistically significant differences. The primary endpoint of transfusion independence for 8 weeks during weeks 1-24 was 31.5% in both the real and synthetic cohorts. Overall survival curves were nearly identical (p = 0.86).
This finding suggests that synthetic control arms could potentially reduce the need for placebo groups in future clinical trials, lowering costs and ensuring all enrolled patients receive active treatment - an important ethical improvement over traditional trial designs.
To make synthetic data generation accessible to clinicians and researchers without deep technical expertise, the team developed a prototype web portal built on the IWG-PM cohort of 2,957 real MDS patients.
The portal allows users to generate synthetic cohorts of custom sizes, evaluate the fidelity and privacy performance of the generated data, and download the synthetic dataset for research use - all without requiring any programming knowledge.
This accessibility is critical because one of the major barriers to adopting synthetic data in medicine is the technical skill required. By providing a user-friendly interface, the researchers aim to enable any hospital or biobank to create shareable, privacy-safe synthetic patient populations from their existing data.
The development of such tools, combined with the establishment of regulatory frameworks and quality standards for synthetic data, represents the key steps needed before synthetic patients can be widely adopted as standard research tools in hematology.
Synthetic data technology has the potential to transform how medical research is conducted. By enabling data augmentation from small real-world cohorts, it allows rare disease research to reach statistical power that would otherwise require decades of patient collection.
For privacy protection, synthetic data offers significant advantages over traditional anonymization. Simple removal of identifying information can be reversed by combining datasets, but synthetic data - which contains no real patient records - fundamentally eliminates that risk while maintaining scientific utility.
The technology also opens new possibilities for international data sharing. Research consortia across countries can share synthetic datasets without legal barriers, enabling larger and more diverse studies than any single institution could achieve alone.
As generative AI continues to improve - including newer approaches like time-series GANs that can model treatment response over time - the accuracy and applicability of synthetic patient data in hematology will expand, ultimately accelerating the path from research discovery to patient benefit.