Actuarial Analyst
Longitudinal Episode Leakage & Variation Decomposition
The Business Problem
An actuarial analyst must explain a $24M annual spend variance in cardiometabolic ETG episodes. The variance could stem from patient severity, supply costs, labor overhead, provider practice patterns, or network leakage. The analysis must be defensible under payer scrutiny.
The integration complexity:
- EHR: Episode-level drill-down into Low Back Pain, Diabetes Management, CHF episodes - correlating early imaging with physician notes
- ERP: Labor costs and clinic overhead to calculate cost-per-episode at facility vs. provider level
- RCM: Claims occurring outside preferred provider network (leakage detection)
- Unstructured (100+ docs): Data dictionaries, release notes, ETG Use Case Methodology documents for proper severity and risk adjustment
The analytical complexity:
- Parameterized queries against partitioned datasets spanning 3 years
- Variation decomposition: How much variance is severity (EHR) vs. supply costs (ERP) vs. billing inefficiency (RCM)?
- Source-to-target mapping validation using data dictionary documentation
- Statistical confidence intervals for each decomposition component
Why RAG + LangChain fails:
| Requirement | Failure Mode |
|---|---|
| Structured queries | Cannot execute parameterized queries against structured data |
| ETG logic | Cannot apply ETG severity adjustment logic (requires reading methodology docs and generating code) |
| Variance math | Cannot partition variance attribution with statistical rigor |
| Data quality | Cannot validate data quality against evolving data dictionary definitions |
How Zenera Solves It
Zenera's agentic execution
- 1Data dictionary comprehension: Agent parses 100+ data dictionary PDFs, extracts field definitions, validates source-to-target mappings
- 2Self-coding ETG logic: Agent reads ETG Methodology documentation, generates severity adjustment SQL, applies to the patient population
- 3Multi-source aggregation: Agent joins EHR episodes to ERP costs to RCM claims within a transactional sandbox
- 4Statistical decomposition: Agent applies variance decomposition with bootstrap confidence intervals
- 5Continuous validation: Agent monitors for schema changes in source systems, alerts when data dictionary updates affect mappings
Time to Insight
18 minutes (vs. 8-12 weeks manual analysis)