FlagshipIn use with clients

Like a real dataset,just without real patients.

The Patient Synthesizer learns from real longitudinal data and generates synthetic patients that are statistically robust without representing any real person. It runs where the data lives. Whether you hold data yourself or are looking for data for your project: we bring both sides together.

What for

The value of synthetic cohorts.

Analyse

Distributions, frequencies and associations can be examined directly on the synthetic cohort. The utility report shows for which quantities the structure holds.

Share and publish

The cohort may leave the building. Partners, contractors and reviewers work on the same material, and results stay reproducible.

Combine

Multiple organisations can contribute synthetic cohorts and analyse them jointly, without any real data leaving any building.

How a synthesis runs

Our software travels, not the data.

The Synthesizer is installed at the data holder and generates a synthetic cohort there. Once released, the cohort may leave the building, and the data user works with it. When a result is needed on the real data, the finished code goes back to the data holder, not the data to the user.

Flow of (synthetic) dataFlow of code and results
Software1
MedModelsPatient SynthesizerDelivered as a licence and installed at the data holder. Runs without an internet connection.
Source data2
Data holderReal longitudinal dataThe existing data, general or tailored to a cohort. It never leaves the building.
Generated in house3
Data holderSynthetic cohortLearned from the real data, statistically consistent, without real persons. With a utility and a privacy report.
Handover4
Data userSynthetic cohort at the userThe data holder reviews and releases the cohort. Only then does it leave the building.
Development5
Data userDevelop the analysisThe data user develops analyses and models on the synthetic data.
Run on real data6
Data holderCode runs at the data holderThe finished code can be returned to the data holder and run on the real data there.
Result7
Data userAnalysis returnedOnly the result goes back, not the data.

Schematic representation. Roles may coincide within one organisation.

Utility

What the cohort preserves.

Distributions are compared, not individuals. The report states the deviation between the real and the synthetic cohort for every quantity checked.

MetricReal cohortSyntheticDeviation
Patients48,21048,210identical
Median age67 years66 years1 year
Hypertension prevalence (I10)61.4%60.8%0.6 pp
Encounters per patient (median)99identical
Days between encounters (median)74713 days
Observation period per patient (median)4.2 years4.1 years0.1 years
First diagnosis to intervention (median)541 days528 days13 days
Inpatient readmission within 30 days12.7%13.4%0.7 pp
Histories with no encounter for 12 months18.3%17.1%1.2 pp
Sequence I10 before I2171.2%68.4%2.8 pp
Green: deviation within the defined tolerance.

Illustrative example of a report. Values differ per dataset.

Data protection

Anonymity is tested, not claimed.

Every synthesis comes with a privacy report. It checks whether individual persons can be recovered from the synthetic data. The report is versioned and machine-readable, so it can go straight into your release documentation.

Request a sample report
Privacy reportData protection
Re-identification of individual historiesno match
Nearest-neighbour distanceabove threshold
Membership inference testnear chance
Rare attribute combinationssuppressed
Operation

Runs where the data lives.

Installation

Built for offline operation: after a one-time licence check during setup, before any data is loaded, the software runs without an internet connection. Delivered as a container or Python package, compiled, for your own hardware or your cloud.

Connection

Input and output run through MedRecords, the open data container shared by all products. Common tabular formats are read directly by the open-source library; where needed, we assist with the integration. Output is delivered in whatever format your downstream processing requires.

Requirements

GPU recommended; CPU operation is possible. Actual requirements depend on cohort size and training scope. Reference figures are in the documentation; we agree on the sizing with your IT upfront.

Usage

Versioned software with support and regular releases. Model weights trained on your data stay with you and can be reused for further runs without retraining.

More in the portfolio

More tools on the same core.

Analytics, cohort selection, clinical system integration and the EHDS Tool for regulated releases share the same foundation. Whatever you set up here can be reused.