- A federal court is now reading the internal dev docs of a Medicare Advantage denial algorithm. 90% of its denials were reversed on appeal. The court isn't treating that as a software bug. It's treating it as breach of contract. 🧵
- The algorithm is nH Predict. In Lokken v. UnitedHealth, a March 2026 discovery order opened its training data, validation reports, and development history. If your MA plan runs AI in utilization management, that's the template for what gets subpoenaed next.
- The flaw was architectural, not a glitch. nH Predict heavily weighted diagnosis-based recovery timelines and assigned minimal weight to individual clinical indicators — blood oxygen, caregiver availability, comorbidities. It optimized for population throughput.
- Medicare coverage requires individual clinical judgment. A model trained on average recovery times can't deliver that. One patient with methemoglobinemia was discharged on her diagnosis group's average timeline, not her actual status. Her family paid $16,768 to stop it.
- The "human-in-the-loop" defense won't save you. Managers narrowed the allowed variance from the algorithm's projection 3% → 1%, then disciplined clinicians who overrode it. Review became theater. Every denial now carries the full legal weight of the model.
- A contractual trap sits underneath. Evidence of Coverage documents promise coverage decisions are made by "clinical services staff." If an algorithm is the de facto decider, the gap between what you promised and what you actually ran is the lawsuit.
- The regulatory clock runs in parallel. Under CMS-0057-F, as of March 31, 2026 MA plans publicly report denial rates and turnaround at the contract level. Those numbers are now visible to regulators, media, and plaintiff attorneys at the same time.
- And the overturn rate understates the problem. Only 0.2% of MA beneficiaries ever appeal a denial. That 90% reversal is measured on the handful who fight back — the denials that stick are mostly never tested at all.
- The vendor market doesn't close the gap. Fiddler, Credo AI, Holistic AI monitor models — they don't rebuild a flawed decision architecture. Cohere Health and others optimize PA throughput, not defensibility. Monitoring a broken model better doesn't fix it.
- What we build instead: explainability middleware that integrates with Facets/QNXT, a decision record of the individual clinical factors behind each call, and controls mapped to CMS-0057-F, NIST AI RMF, and state law at once — so a denial can survive discovery.
- Honest question for MA plans and UM medical directors: if your prior-auth AI were subpoenaed tomorrow, could it show why it denied one specific patient — not the cohort? For the cohort, sure. For that one patient? That's where most plans break. #HealthcareAI
- We wrote up the full architecture — the liability mechanics, the CMS-0057-F mapping, and how to make individual decisions defensible: https://veriprajna.com/solutions/medicare-advantage-ai-governance