Synthetic patients. Provably private.
MirrorCohort is a privacy engine for clinical data. It learns the statistics of your patient table inside your own environment and generates a new cohort that behaves like the original, with no real individuals in it and a validation report that proves it.
stays in your environment Differential privacy layer Synthetic cohort
safe to share and analyse
Clinical research runs on data nobody is allowed to touch.
Feasibility, analysis plans, pipeline builds and partner onboarding all wait on patient-level access. Synthetic data removes the wait, but only when a privacy team can verify the claim.
Access takes months
IRB reviews, data use agreements and de-identification projects stall the work that should happen first: building and testing the analysis.
Small cohorts can't be shared
Rare-disease, pediatric and subgroup tables are too thin to release safely with conventional de-identification.
"Trust us" fails review
Most synthetic data is sold on faith. Governance teams need a stated privacy budget and a fidelity score they can reproduce.
A privacy pipeline, accounted end to end.
Every stage that touches real records spends from one privacy budget, tracked by a single accountant. The generator is chosen automatically for the shape of your data.
Source table
Your patient-level CSV, loaded on your own hardware. Columns and types are detected automatically.
Noisy statistics
Column ranges and distributions are measured with calibrated noise, inside the budget.
Learns the cohort
A DP-SGD neural generator, or a private copula for small and lab-heavy tables.
New patients
Any number of rows, same schema and relationships, with outcome-aware generation.
Proof report
Fidelity, utility on real held-out patients, and a membership attack, in one PDF.
Benchmarked in the open, on public clinical data.
Four public datasets, all at ε 8. The generator trained on 70% of each; every score below is measured on the 30% of patients it never saw.
Statistical fidelity
FIDELITY SCORE (SDMETRICS)Predictive utility
ROC AUC ON REAL HELD-OUT PATIENTSMembership attack, all four datasets. An attacker tries to tell which patients were used for training. A score of 0.50 means a coin flip, so the synthetic data gave no detectable signal about who was in the source.
Datasets: UCI diabetes hospital encounters (5,000 rows), CDC diabetes health indicators (5,000), NHANES adults vitals and labs (3,570), heart failure clinical records (299). Fidelity Score uses the open-source SDMetrics quality definitions, so anyone can rerun it. Single runs; expect a point or two of variation.
Every dataset ships with its own proof.
A stated privacy number
The ε covers the full pipeline, including preprocessing, so the guarantee isn't limited to the model alone.
Reproducible fidelity
Scored with open-source tooling your biostatisticians can rerun, not a proprietary black-box metric.
Utility tested on real patients
Models trained on synthetic data are scored on real held-out rows, so you know exactly what the data is good for.
Built for the teams who live in patient tables.
CROs and biostatistics
Build and test the analysis before the real data arrives.
- Protocol and feasibility simulation
- Statistical analysis plan dry runs
- Pipeline and vendor onboarding
Research informatics
Realistic data for investigators and trainees while full access is under review.
- Teaching and hackathon datasets
- External collaborator sharing
- Pilot analyses during IRB review
Rare-disease programs
Cohorts of a few hundred patients that conventional methods can't release.
- Small-cohort private generation
- Planning and feasibility data
- Method development
Designed for the people who say no.
Privacy officers and data governance committees are the real approvers. The engine is built around what they need to see.
Stays in your environment
Installs on-premises or in your private cloud. Source records never leave.
Math, not masking
Differential privacy bounds what any output can reveal about any one patient.
No LLMs in the data path
Rows come from dedicated statistical models with privacy accounting, never a chatbot.
Open, checkable methods
Open-source fidelity scoring and a published benchmark protocol.
What we don't claim. Differentially private synthetic data reduces re-identification risk, but it does not by itself make a dataset exempt from HIPAA or other regulations. De-identification status should be determined by a qualified expert or your privacy office, and we provide the documentation they need.
Asked by privacy and biostatistics teams.
Not answered here? Write to us. We answer technical questions directly.
What does ε 8 mean in practice?
ε (epsilon) is the privacy budget in differential privacy. Lower values add more noise and give stronger protection. Our benchmarks use ε 8, a common setting for synthetic data that keeps fidelity high. You can choose a tighter budget per run, and the report states exactly what was spent.
Does any data leave our environment?
No. The engine runs on your hardware or private cloud. We help with installation and review results with you, but source records are never sent to us.
What kinds of tables work best?
Single patient-level tables: encounters, registries, survey and risk-factor data, vitals and labs. Encounter and survey data score highest today. Lab-heavy and very small cohorts use a private copula model, and we report their scores just as plainly.
Can synthetic data be used in regulatory submissions?
MirrorCohort is for planning, development, teaching and sharing: the work that happens before or alongside real-data access. Final analyses for submissions should run on the real data.
How is this different from other vendors?
Most platforms are general-purpose or cloud-first. MirrorCohort focuses on clinical tables, runs inside your environment, and attaches a privacy budget and a fidelity score to every single run.
Bring one dataset. See the proof first.
We're running on-premises pilots with a small group of research teams. You keep the data, the synthetic cohort and the validation report.
- A 20-minute call about the dataset and the question it should answer.
- We install the engine in your environment, alongside your team.
- You receive a synthetic cohort and its validation report.
- We review fidelity and privacy results with your biostatistics and privacy leads.