ClinicalBrief runs on generated data from the Synthea project. Nothing in the demo belongs to a real person. Here is what the dataset contains, how it gets in, and who to credit for it.
The demo dataset
What is in the database
409
Synthetic patients
109 Synthea sample + 300 Synthea Coherent
35,103
Synthetic clinical notes
All searchable in the database
824
Notes run through the AI pipeline
Findings wait for human review
0
Real patients
No real patient data anywhere
The notes are templated by the generator, so the AI pipeline's results will look better here than they would on real clinical notes.
What synthetic means
Realistic records, fictional people
Synthea is an open-source synthetic patient generator. It simulates whole lives of fictional patients, from birth through encounters, diagnoses, medications and procedures, and writes them out as standard FHIR records.
The records look and behave like real ones, but no one in them exists. Names such as Magali989 are generated (the digits are part of the generated name), and clinical histories are simulated.
That is what lets the demo be public: you can open any patient, read their notes and inspect the AI output without seeing anyone's private data.
Provenanceexample
patient
Magali989 (synthetic)
source_system
synthea
source_id
8f3c…e21a
provenance
imported
Illustrative values. Every imported record carries these fields, so any row can be traced back to the source file it came from.
What is imported
FHIR R4 in, one canonical model out
The Synthea adapter reads FHIR R4 bundles and maps them into one canonical clinical model:
patients
encounters
notes
diagnoses
medications
allergies
procedures
observations
Every record keeps source_system, source_id and provenance.
01
Dry run
See what a folder contains and what would be imported or skipped, without writing anything.
02
Per-record validation
Invalid source records are skipped and written to a rejected-record log with a reason; the rest continues.
03
Idempotent re-import
Record ids are derived from source system, table and source id, so re-importing updates rows instead of duplicating them.
04
All or nothing on failure
An unexpected failure rolls back every data write and marks the run as failed.
05
Run history
Each run is recorded with its counts, rejected records and skipped resource types.
06
Nothing invented
Unknown birth dates, statuses and codes stay empty rather than being guessed.
Not yet imported: Synthea immunisations, diagnostic reports, care plans and claims. From the Coherent dataset only the FHIR part is used; imaging, genomics and ECG data are not downloaded.
Licences and attribution
Credit where it is due
Synthea sample, FHIR R4
109 patients
Sample data published by the Synthea project, generated with Synthea.
Walonoski J, et al. The “Coherent Data Set”: Combining Patient Data and Imaging in a Comprehensive, Synthetic Health Record. Electronics. 2022;11(8):1199. doi.org/10.3390/electronics11081199
Why not MIMIC
A real dataset was tried, then removed
MIMIC-IV demo
The MIMIC-IV demo was imported earlier and later removed. It publishes no names, birth dates or notes, and its dates are shifted into the 2100s, which did not suit a patient-centred workspace built around reading notes.
Where the data lives
One database, in Sydney
Supabase Postgres
Hosted in Sydney (ap-southeast-2)
The canonical record, notes, AI output, review decisions and audit log are stored there. Browsers can only read through row-level security; all writes go through the API.
Local model
Evidence stays on the same machine
The Copilot uses a local model through Ollama. Only loopback model addresses are accepted and cloud models are refused, so the evidence it is given never leaves the machine running it.