Synthetic healthcare data generation system for Medicare analytics
Umbizo built a synthetic data system for a healthcare analytics company that generates around 6 million synthetic patients for testing and developing code within the client's Medicare claims analytics platform.
The challenge
The client needed synthetic data to support code development and testing on its Medicare claims platform. The technical problem involved generating roughly 6 million synthetic patients while keeping statistical alignment with real-world distributions, applying billing patterns across different encounter settings, and creating clinically plausible patient journeys with realistic progression.
What we set out to do
- Keep patient identities consistent, with demographic and clinical characteristics preserved across monthly data generations
- Produce epidemiologically reasonable population distributions
- Build realistic patient journeys across multiple disease areas with appropriate clinical progression
- Model encounters for outpatient, inpatient, clinic, and other care settings
- Create a validation framework measuring alignment between synthetic and real-world data
Our approach
Data architecture analysis
Evaluated the existing Databricks environment and synthetic data scripts, reviewed data structures and reference pricing tables for procedures and diagnoses, identified requirements for a SNOMED-to-ICD-10 and CPT code mapping system for Medicare compatibility, and designed patient consistency logic for persistent identities across time periods.
Technical implementation
Configured and optimised Databricks for the existing infrastructure, implemented deterministic algorithms for persistent patient IDs across monthly datasets, built a code mapping system translating SNOMED clinical terms to ICD-10 and CPT codes, and configured encounter generation by setting (outpatient, inpatient, clinic, other).
Validation framework
Developed a statistical validation framework with automated tests against population norms, measuring alignment across prevalence rates, comorbidity patterns, treatment sequences, and billing distributions.
Solution delivery
Built disease models for therapeutic areas incorporating prevalence rates, comorbidity patterns and treatment pathways, implemented treatment sequences reflecting temporal care patterns, produced technical documentation (architecture diagrams, module specifications, integration guides), and ran a hands-on knowledge transfer programme with the client's CareSet engineering team.
Results and impact
The system was designed to generate 6 million synthetic patients while maintaining persistent patient identity and demographic consistency across an unlimited number of monthly generations. The validation framework measures alignment across prevalence rates, comorbidity patterns, treatment sequences, and billing distributions. The client gained a more realistic test environment reflecting population complexity beyond single-disease scenarios, supporting development of analytics algorithms and analysis of correlations between National Provider Identifiers, hospitals, and patient diagnoses.
“Umbizo solved the challenge of creating realistic test data that actually reflects the complexity of Medicare populations. Our previous synthetic data had every patient with diabetes; this approach moves us to epidemiologically reasonable distributions across multiple disease areas.”
Timeline
- Assessment and planning: weeks 1-2
- Core framework implementation: weeks 3-5
- Initial disease implementation (2 areas): weeks 6-9
- Additional disease areas (2 areas): weeks 10-13
- Knowledge transfer and deployment: week 14
- Total initial implementation: 14 weeks
Looking forward
Exploring advanced comorbidity interaction models, more sophisticated temporal progression for chronic disease management, integration of social determinants of health, machine learning approaches for more realistic clinical decision patterns, and continuous evolution of the validation framework as disease coverage expands.
