Synthetic cohort¶

1,000 synthetic patients were generated with Synthea (master-branch-latest, released 2026-08-18): the original 100 (seed 618), used by the hand-written scenario cases, plus 900 more (seed 619) for the generated cases, note search and feature scenarios. No real patient data was used.

java -jar synthea-with-dependencies.jar -s 618 -cs 618 -p 100 -a 25-90 \
  --exporter.years_of_history 5 --generate.only_alive_patients true Massachusetts
# +900: seed 619, -p 900, otherwise the same flags


Age	median 59, range 25–89
Sex	509 female, 491 male
Active disorders	mean 4.3 per patient
Active medications	mean 3.9 per patient
Recorded allergies	154 patients
Most common active disorders	Anemia (449), Essential hypertension (355), Ischemic heart disease (337), Chronic sinusitis (265), Metabolic syndrome X (209), Loss of teeth (205)

Storage and FHIR versions¶

Path	Content	FHIR
`data/synthea-r4/`	Complete Synthea bundles for the original 100, gzipped, including claims	R4 4.0.1 + US Core, as exported
Release `cohort-1000`	Complete Synthea bundles for the other 900 (too large to commit)	R4 4.0.1 + US Core, as exported
`data/fhir-r5/`	Clinical snapshot per patient, including the 10 most recent clinical notes: Condition 12,445, DocumentReference 9,843, Observation 7,978, Encounter 4,993, MedicationRequest 3,884, Patient 1,000, AllergyIntolerance 679	R5 5.0.0, validated
`data/fhir-r5-scenarios/`	Hand-written scenario inputs: vital signs (`Observation`), nursing notes and discharge text (`DocumentReference`), patient messages (`Communication`)	R5 5.0.0, validated
`data/scenarios/`	Case definitions and reference labels, each set committed before its first model run	–

The R4 → R5 mapping (src/explore_typesafe/fhir.py) handles the breaking changes these resources hit: MedicationRequest.medication[x] → medication (CodeableReference), reasonReference → reason, Dosage.asNeededBoolean → asNeeded, AllergyIntolerance.type code → CodeableConcept, reaction.manifestation → CodeableReference, Encounter.class → list, period → actualPeriod, DocumentReference.context.period → period, and status finished → completed. Every bundle is validated against the R5 models in fhir.resources 8.3.

Jev never sees raw FHIR. Code builds a small, named JSON state for each question set, with only the fields that decision needs. The jaggedness notes warn that irrelevant state costs accuracy.