Generative Framework

Trigger Phrases

Building Specifications

• "Build a profile for 100 Medicare members"
• "Create a journey for diabetic patients"
• "Specify a commercial cohort with CHF"
• "Define a surgical episode journey"

Executing Specifications

• "Generate patients from this profile"
• "Execute the diabetic journey for 50 patients"
• "Run batch generation for this cohort"

Using Templates

• "Use the Medicare diabetic template"
• "Start with the surgical episode journey"
• "Generate using the commercial healthy profile"

Workflow Overview

1. Build a Profile (Optional)

Build a profile for 200 Medicare Advantage members:
- Age 65-85, normal distribution centered at 74
- 40% with Type 2 diabetes
- 30% with heart failure
- San Diego County geography

2. Build a Journey (Optional)

Create a first-year diabetic journey:
- Initial diagnosis visit with labs
- Metformin prescription
- Quarterly follow-ups with A1c
- Possible titration to second agent

3. Execute Generation

Generate the cohort using this profile and journey
Save as cohort "ma-diabetic-cohort-2025"

Distribution Types

See distributions/distribution-types.md for details.

Statistical Fidelity — Matching Real Populations

Generated cohorts should use statistical distributions that match real-world population characteristics. The goal is realistic synthetic data, not random noise.

How to match a target population:

1. Identify the source benchmark. For Medicare Advantage, use CMS enrollment statistics (mean age ~72, 55% female, regional mix). For Commercial, use employer-census norms. For Medicaid, use state-level demographics.
2. Select distribution types that fit the data shape. Age in MA populations follows a roughly normal distribution centered near 72-74 with standard deviation ~7. Cost distributions are log-normal (many low-cost, few high-cost). Chronic condition prevalence uses categorical distributions with rates from CDC PLACES or CMS Chronic Conditions data.
3. Parameterize from published statistics. Do not guess distribution parameters. Use published prevalence rates (e.g., 27% diabetes in Medicare, 14% heart failure) and demographic breakdowns. PopulationSim data (skills/populationsim/) provides county- and tract-level benchmarks for prevalence and social determinants.
4. Validate output against inputs. After generation, confirm the cohort's statistical profile (mean, median, standard deviation, category proportions) matches the specification within an acceptable tolerance. Flag drift greater than 5% from target parameters.

Journey Patterns

See the journeys/ folder for pattern details.

Integration with Products

The Generative Framework orchestrates all HealthSim products:

Templates

Pre-built profiles and journeys for common use cases:

Profile Templates

• medicare-diabetic.md - Medicare T2DM cohort
• commercial-healthy.md - Commercial healthy adults
• medicaid-pediatric.md - Medicaid children

Journey Templates

• diabetic-first-year.md - New T2DM diagnosis year 1
• surgical-episode.md - Elective surgery episode
• new-member-onboarding.md - New member first 90 days

Safety Guardrails

All generated data is synthetic and fictional. HealthSim produces simulated test data only. Never present generated records as real patient data.

• No clinical advice. Generated data must not be used to make prescribing decisions, diagnoses, or treatment recommendations. If asked, remind the user this is synthetic test data, not real clinical information.
• Real codes, synthetic entities. Use real, valid medical code systems for realism (see below), but all patients, members, encounters, and claims are generated/fictional.
• No real PHI. Never pull from or reference real patient databases. All person-level data is simulated.

Standard Code Systems

Generated data should reference recognized healthcare code systems:

Real reference data (NPI registry, CMS facility files, published code sets) is safe to use. Synthetic data is generated for all patient/member-level entities.

Edge Cases and Error Handling

When generating data, handle incomplete or invalid inputs gracefully:

Missing or Partial Input

What NOT to Generate

These are common mistakes to avoid:

• Do NOT invent code systems. Use only recognized systems (ICD-10, CPT, HCPCS, LOINC, RxNorm, NDC, SNOMED CT, NPI). Never fabricate codes that look plausible but do not exist in the standard.
• Do NOT generate clinically impossible combinations. Example: a 5-year-old with Medicare Fee-for-Service, or a pregnancy diagnosis on a male patient. Cross-check age, sex, and plan type against clinical plausibility.
• Do NOT silently drop requested attributes. If the user asks for 10 fields and only 8 can be generated, surface the gap explicitly rather than returning incomplete records without explanation.
• Do NOT extrapolate beyond the specification. If the user asks for "50 patients with diabetes," generate exactly 50 with diabetes -- do not add additional unrequested conditions for "realism" unless the user has opted into realistic comorbidity modeling.
• Do NOT generate duplicate identifiers. Every patient ID, member ID, encounter ID, and claim ID must be unique within a cohort.

Validation Checklist

Before returning generated data, verify:

1. All code values exist in their respective standard (ICD-10, CPT, LOINC, etc.)
2. Date sequences are chronologically consistent (admission before discharge, prescription before fill)
3. Demographic fields are internally consistent (age matches date of birth, plan type matches age eligibility)
4. Requested counts match output counts (if 200 members requested, 200 are returned)
5. No duplicate identifiers exist across the generated dataset

Related Skills

• State Management - Save/load generated cohorts
• Identity Correlation - Cross-product entity linking
• PopulationSim - Real-world demographic data
• NetworkSim - Real provider assignments

Implementation Status: Foundation phase. See GENERATIVE-FRAMEWORK-PROGRESS.md for details.