TLDR: RareGraph-Synth is a novel knowledge-guided diffusion model that generates highly realistic and privacy-preserving synthetic electronic health records (EHR) for ultra-rare diseases. By integrating five public biomedical knowledge resources into its noise schedule, the model achieves 40% better data fidelity than unguided diffusion and strong resistance to membership inference attacks (AUROC ≈ 0.53), enabling safer data sharing for critical rare disease research.
Research into ultra-rare diseases faces a significant challenge: a severe lack of patient data. With some disorders affecting fewer than 1 in 50,000 people, obtaining enough longitudinal health records for meaningful study is incredibly difficult. This data scarcity hinders critical advancements in understanding disease progression, mapping genetic factors to symptoms, and developing effective treatments. Synthetic electronic health records (EHRs) offer a promising solution, as they can be shared across different research institutions without compromising patient privacy.
However, existing methods for generating synthetic EHRs have their own limitations. Early rule-based systems struggled to accurately reflect statistical patterns, while more advanced adversarial models (like GANs) often failed to generate realistic data for rare medical codes, a phenomenon known as ‘mode collapse.’ Even newer diffusion models, while generally robust, have treated EHR data as unstructured, overlooking the rich biological and medical knowledge embedded in ontologies that describe disease-gene-phenotype-drug relationships. This oversight can lead to the creation of biologically implausible patient trajectories.
Introducing RareGraph-Synth
A new framework called RareGraph-Synth addresses these challenges by proposing a knowledge-guided, continuous-time diffusion model designed to generate realistic, yet privacy-preserving, synthetic EHR trajectories specifically for ultra-rare diseases. Developed by Khartik Uppalapati, Shakeel Abdulkareem, and Bora Yimenicioglu from the RareGen Youth Network, this innovative approach integrates vast amounts of biomedical knowledge directly into the data generation process.
RareGraph-Synth unifies five major public resources into a massive, heterogeneous knowledge graph containing approximately 8 million typed connections. These resources include Orphanet/Orphadata (for rare disorder information), the Human Phenotype Ontology (HPO) (for standardized phenotype terms), the GARD rare-disease Knowledge Graph (linking diseases to genes, phenotypes, and drugs), PrimeKG (a multimodal knowledge graph), and the FDA Adverse Event Reporting System (FAERS) (for drug-adverse event reports). This comprehensive knowledge graph provides a rich context for understanding complex medical relationships.
How it Works: Knowledge-Guided Diffusion
The core innovation of RareGraph-Synth lies in how it uses this knowledge graph. Instead of simply feeding context into the denoising process (as many existing diffusion models do), RareGraph-Synth conditions the ‘forward’ stochastic differential equation. This means that meta-path scores, extracted from the 8-million-edge knowledge graph, are used to modulate the noise schedule during the initial phase of data generation. Essentially, tokens (like lab codes, medications, or adverse events) that are strongly connected to a patient’s anchor disease within the knowledge graph receive less noise. This ‘knowledge-guided’ noise injection steers the generation process towards creating biologically plausible co-occurrences of lab results, medications, and adverse events, ensuring the synthetic trajectories are clinically meaningful.
The reverse denoising process then refines this noisy data into timestamped sequences of medical events, such as ⟨LabCode, MedCode, AEFlag⟩ triples, which contain no protected health information (PHI). This novel approach is the first EHR diffusion framework to embed ontology priors directly into the noise schedule, guiding synthesis towards realistic outcomes while reducing the risk of simply copying existing records.
Enhanced Fidelity and Robust Privacy
The results of RareGraph-Synth are highly promising. On simulated ultra-rare-disease cohorts, the model significantly lowers the categorical Maximum Mean Discrepancy (a measure of how different two distributions are) by 40% compared to an unguided diffusion baseline. It also performs over 60% better than traditional GAN-based models, all without sacrificing the utility of the data for downstream predictive tasks. This means the synthetic data is highly realistic and can be used effectively for research and analysis.
Privacy is a paramount concern for synthetic data. RareGraph-Synth was rigorously evaluated against black-box membership-inference attacks (MIAs), specifically using the DOMIAS attacker. The model achieved an AUROC (Area Under the Receiver Operating Characteristic curve) of approximately 0.53. An AUROC of 0.50 indicates a random guess, while 0.55 is often considered a “safe-release” threshold. RareGraph-Synth’s score is well below this threshold and substantially better than the approximately 0.61 observed for non-knowledge-graph baselines, demonstrating strong resistance to re-identification. This indicates that it is extremely difficult for an attacker to determine if a specific real patient’s record was used to train the model.
Also Read:
- Preventing Document Linkage: A New Method for De-identified Texts
- RareAgent: A New AI System for Discovering Drug Uses in Rare Diseases
Conclusion
The findings suggest that directly integrating biomedical knowledge graphs into diffusion noise schedules can simultaneously improve both the fidelity (realism) and privacy of synthetic EHR data. This breakthrough enables safer and more effective data sharing for ultra-rare disease research, potentially accelerating discoveries and therapeutic developments for millions of affected individuals. The full research paper can be found here.
