- A 98.5%-confident denial still triggered a physician-review route in our synthetic Medicare Advantage case. Individual clinical factors carried only 22.06% of the model's attribution. 🧵
- In A-4471, recovery-timeline gap accounted for about 49% of absolute attribution and prior utilization about 19%. We used exact Shapley attribution across 10 model features to make that weighting visible.
- The rule we built is narrow: for a salient denial, individual clinical factors must account for at least 35% of attribution. A-4471 fell below that floor, so the deterministic gate routed it as NEEDS_PHYSICIAN_PROOF.
- That route is pending physician review, not a final denial or a medical-necessity judgment. Confidence alone did not give the model authority to close the case.
- CMS's February 2024 guidance says Medicare Advantage coverage determinations must consider the individual patient's circumstances. An algorithm based on a larger dataset cannot substitute for that review.
- We kept the explanation separate from authority. Model-generated text can explain the result when configured; a deterministic template works without a provider. Neither changes the code gate's route.
- The local record lets us reconstruct the initial assessment, attribution and route. Its hash chain can expose a deliberately altered record. That is technical traceability, not a legal finding or independently secured custody.
- This is a synthetic, local demonstration. The physician queue is a demo state, and the 35% floor is a configurable policy choice, not a clinically validated threshold. Real review still belongs to qualified humans.
- For MA utilization management teams: should a high-confidence denial with 22% individual-factor attribution be held on that fact alone, or should a second clinical trigger be required? #MedicareAdvantage #UtilizationManagement #AIGovernance
- https://veriprajna.com/demos/medicare-advantage-ai-governance
Published on X · September 25, 2026
On social media
See this post on its original platform
In our archive