A claims-denial model with a 90% reversal rate on appeal — and 0.2% of patients ever appealed
UnitedHealthcare · 2019–ongoing
What they did
UnitedHealthcare's nH Predict tool (from its naviHealth subsidiary) analyzes a Medicare Advantage patient's diagnosis, age, and living situation to predict how much post-acute nursing care they "should" need, and was allegedly used to time cutoffs for further coverage.
What happened
A federal class-action lawsuit alleges the model had a 90% error rate — the share of coverage denials overturned on internal or federal administrative appeal — and that UnitedHealth kept using it knowing only about 0.2% of denied patients would ever file an appeal. The suit alleges the AI's predictions routinely overrode treating physicians' own medically-necessary determinations. As of mid-2026 the litigation and a related court order to disclose the algorithm are ongoing; the allegations are unproven claims, not a final judgment.
The so-what
A 90%-overturned-on-appeal rate isn't a model with a bug — if the allegations hold, it describes a model whose errors were the business model, because the appeal rate needed to catch them was priced in at under 1%. Watch what a company measures a model against, not just what it reports.