Make patient data shareable. Without sharing patient data.
The model learns from real records inside the custodian's environment. The data never leaves.
Trained under differential privacy, it carries the patterns, not the people. It cannot leak any individual, so the model itself can safely leave.
The model is evaluated against real data on three pillars: statistical similarity, consistency with clinical rules and downstream utility.
When does a differentially-private synthetic dataset stop being useful? Across six tabular datasets we map a sharp viability boundary for DP-SGD with MLP variational autoencoders, set by the ratio of training samples to encoded dimensions. Marginal-based DP methods can be viable two orders of magnitude lower.
On a CITE-seq PBMC dataset, grouping single-cell features into biological modules substantially beats flat baselines at matched parameter budgets. Preliminary, one dataset.
→Three autoregressive conditional mean models on PhysioNet 2019 ICU data. Marginal similarity improves with complexity. Cross-feature clinical rules do not. The pattern points to architecture, not missing signal.
→A position paper. We propose a 16-item reporting checklist extending STROBE from epidemiology into healthcare GenAI: data, privacy, bias, clinical relevance, regulatory compliance.
→A reinforcement-learning gym for medical-reasoning agents, on multi-modal patient trajectories. DP-trained.
Coming soonOutputs as FHIR R4 and R5, or formats tailored for research.
Cohorts, edge cases and markets real data cannot give you. Generated on demand, PHI-free.
Plan and simulate trials before the first enrolment.
Run internal studies without the data-access wait.
Hand the same cohort to collaborators, reproducible by seed.
Build and test products against realistic data, no PHI to handle.
Privacy guarantees come from the mathematics, not from policy. The generator is built to slot into regulated healthcare workflows.
Based in Iceland, inside the EU/EEA. No real records are stored, and differential privacy means none can be recovered.
Synthetic data carries no personal data, which eases the Act's data-governance burden on the medical AI you build and test.
Synthetic data carries no PHI, with proof.
Your own cloud, with SSO, RBAC and audit logs. The data never leaves your infrastructure, and we never see it.
Machine learning and clinical inference. Previously Samsung AI, Evidation Health.
Computer vision and large-scale synthetic data for ML.