Healthcare Quality
Control Prototype
A working model of the Nartis loop: pick a synthetic member, evaluate them against 34 simplified HEDIS and Stars measures, see what each closure does to the plan rate, rank the population by value, and ask Nartis to explain a gap. Everything here runs in your browser except the one live-model call.
What's real vs simulated
Real, deterministic code
- 34 simplified measure rules written as policy-as-code: denominator, numerator, lookbacks, exclusions for Medicare-only measures.
- Rate math from the synthetic population: each closure moves a rate by exactly 1 / denominator.
- Stars bonus formula: 5% benchmark increase at 4 stars or higher, 10% in double-bonus counties.
- Audit trail of every evaluation, simulated action and model call (in memory, cleared on reload).
Simulated
- 10,004 synthetic members generated deterministically, with dates relative to measurement year 2026. No real patient data.
- Risk scores are fixed constants per rule outcome (for example, A1c ≥ 9% scores 92). They are not model output.
- The prioritizer uses transparent heuristics: measure weight, distance to cut point, visit recency, adherence level.
- Cut points and measure weights are illustrative. "Simulate" buttons send nothing anywhere.
Live model, and where models plug in
- Ask Nartis is the one live-model feature: a hosted language model called through a serverless function, synthetic de-identified fields only, capped output, rate limited.
- Next: a closure-propensity model trained on outreach history replaces the fixed risk constants.
- Next: spec-to-code drafting from measure specifications, reviewed by a person and tested against test decks.
- With real PHI: Amazon Bedrock with Guardrails under a BAA, minimum-necessary fields, full audit.
The one live-model feature in this prototype. For the selected member and open gap, Nartis explains why the gap is open, citing the measure rule and the evidence found or missing, and drafts a short outreach note for a person to review. Only synthetic, de-identified fields are sent: no name and no member ID.
Rank every Medicare Advantage member in the synthetic plan by the expected value of closing their open gaps. Each gap is scored on measure weight, rule-based severity, how far the measure sits from its cut point, and how likely this member is to close it (visit recency, adherence level, how close a lab value is to threshold). Star Ratings apply to Medicare Advantage only, so commercial members are not ranked here.