Synthetic patient data with ground truth
Synthetic patients with answer keys
CohortLoom builds longitudinal synthetic patients: multi-visit histories, labs that trend with each condition, and the clinical notes written along the way. Every file ships with its own answer key, and every patient ships with a master key that ties the whole chart together.
Synthetic by construction. No real patient data, ever.
Test data where you know the right answer
If you build software that reads clinical documents, such as an extraction model, an autocoder, a chart summariser or an EHR integration, you need test data. Real records bring privacy work, access delays and, most of all, no ground truth: someone still has to label them by hand.
CohortLoom starts from the truth and writes the documents from it. Because each document is planned before it is written, we know exactly which diagnoses, labs, medications and facts it contains, and where in the text each one appears.
An answer key per file
Labels, ICD-10-CM codes, facts and evidence spans with exact character offsets, for every note and report.
A master key per patient
One key links a patient's files across visits: what was known at each point, and which document says it.
Checked before packaging
Each file is checked against its keys before it goes in a dataset. A file that doesn't match is revised or left out.
Longitudinal by design
Patients have several visits, with labs such as A1c and kidney function that move with their conditions over time.
Who it's for
- Health-AI teams testing clinical NLP and summarisation against labelled notes.
- Revenue-cycle and coding software teams checking an autocoder against codes with exact evidence offsets.
- EHR and integration engineers who need realistic multi-visit test patients in FHIR.
- Researchers who want a shareable dataset with no privacy constraints.
What we say, and what we don't
We say
- Synthetic by construction. Patients are simulated; documents are written from the simulated chart.
- No real patient data, including no de-identified data, anywhere in the process.
- The models that write our documents are never trained on Claude or Gemini outputs.
We don't claim
- Clinical validity. These are test and development datasets, not medical records.
- Quality numbers before they're published. A validation report and a sample are available on request.
Founding 100: our first 100 customers get 20% off their first purchase, applied automatically at checkout. See pricing. Checkout isn't open yet: join the wait-list for first access to Founding 100 pricing.