How the data is generated
I build a synthetic person, give them a clinical history, and let the records come out of it.
Records for the same person stay consistent because they come from one source. Here is the path from an archetype to a calibrated journey to the claims, labs, encounters, ADT events, and chart you get.
The five steps
From a person template to a complete data trail.
The model supplies clinical knowledge once per version. A seeded engine turns that knowledge into people, journeys, and records, so the same inputs always produce the same output.
Author the patient archetypes
A language model writes a library of patient archetypes. An archetype is a person template: a realistic bundle of co-occurring conditions, how severe each is and where it is heading, how much care the person tends to use, and which acute events they are at risk for. The model uses real clinical knowledge to keep the bundles coherent, so diabetes travels with kidney disease and hypertension the way it does in real patients.
Set the targets from real-world patterns
Before generating any member, we fix the numbers the population has to hit. These come from published sources: CMS risk-score levels, condition prevalence from the CMS Report to Congress and MedPAC, inpatient admits and ER visits per thousand, and medical loss ratio. These are the numbers an actuary checks first.
Sample and weight to the targets
Each archetype carries a prevalence weight. We calibrate those weights so the generated population matches the targets, scoring every archetype with the real CMS HCC model end to end. Members are then drawn from the weighted library and perturbed, so two members from the same archetype are never identical.
Simulate the journey
Each member lives 36 months. Conditions progress, acute episodes fire at their modeled rates, medications start and stop, and care follows realistic patterns. The output is a timeline of clinical events for that person.
Render the records
The timeline becomes the data files. Each event emits properly coded rows: eligibility, claims (with real DRGs, CPT/HCPCS, revenue codes, place of service, and ICD-10s), pharmacy, revenue with a risk score computed from the codes that actually appeared, labs whose values track the member's conditions, encounters, ADT events, and a per-member chart.
Worked example
Margaret, 74: resistant hypertension with early kidney disease.
One archetype becomes one member, and her whole data trail follows from the life she lives.
Conditions: resistant hypertension (I10), left ventricular hypertrophy, CKD stage 3 (N18.30, HCC138), hyperlipidemia, former tobacco.
Utilization: moderate, balanced, moderate adherence.
At-risk events: hypertensive urgency, acute kidney injury, new atrial fibrillation, seasonal flu and cold.
Documentation: 85% coding capture, so some true conditions go uncoded in a given year.
- Eligibility36 member-months, dual status, plan
- Medical claimsoffice visits (99214, dx I10 / N18.30), the hypertensive-urgency ER visit (99285)
- LabseGFR 47, creatinine 1.6, A1c 5.5 (normal), potassium monitored
- Pharmacylisinopril, amlodipine, atorvastatin refill chains
- Encounters / ADTthe ER visit as an encounter and an A04 register / A03 discharge pair
- Revenuemonthly capitation with a risk score built from the coded HCCs
- Journeyher pseudo-chart: true problem list, coded HCCs, and a chart-note narrative
Margaret's eGFR reads 47 because she has stage-3 kidney disease. Her A1c reads normal because she is not diabetic. The lab values match the person.
Worked example
James, 68: healthy.
The healthy tail looks healthy.
James has no chronic conditions, low utilization, an annual wellness visit, and seasonal acute care. His file is thin: a wellness visit each year, one or two minor acute visits (a winter cold, a urinary infection), a normal lab panel, no chronic medications, and a low risk score.
Because the population is weighted to published prevalence, common low-acuity members like James make up the bulk of the panel, exactly as they do in a real book of business.
Why this matters
Every file for a member comes from the same simulated life, so the claim, the lab, the medication, and the revenue agree by construction. Because the population is weighted to published targets, the aggregate looks like a real book of business. Person-level coherence plus population-level realism is the product.
Want the deeper version? Sampling, targeting, and the eval framework →