The Algorithm That Denied the Care

The Global Signal

UnitedHealthcare and its subsidiary naviHealth use an algorithm called nH Predict to estimate how much post-acute care, skilled nursing, inpatient rehabilitation, home health, a Medicare Advantage patient is likely to need after a hospital stay. A federal lawsuit, Estate of Lokken v. UnitedHealth Group, filed in the District of Minnesota in November 2023, alleges that UnitedHealth used nH Predict's estimates to cut off coverage for patients whose own doctors had recommended continued care, and that the company knowingly deployed an algorithm with what the complaint calls a 90 percent error rate, meaning nine in ten denials were reversed when families pursued an appeal. That figure comes from the plaintiffs' own filing; it has not been independently established by a court, and the case remains in active litigation. A February 13, 2025 ruling dismissed the plaintiffs' unjust-enrichment and bad-faith insurance claims as preempted by the federal Medicare Act, while allowing breach-of-contract and breach-of-the-implied-covenant-of-good-faith claims to proceed. A March 9, 2026 discovery order compelled UnitedHealth to turn over internal documentation describing how the algorithm was designed, trained, and reviewed.

Separately, a November 14, 2023 STAT News investigation, based on internal documents and interviews with former employees, reported that UnitedHealth set a 2023 target requiring case managers to keep patients' rehab stays within 1 percent of the algorithm's projected days, with case managers who fell outside that target facing discipline. Reporting on the same investigation describes an earlier, looser target of roughly 3 percent in 2022, tightened to that 1 percent figure the following year; this draft treats the 2023 figure as directly confirmed from the published article and the 2022 comparison point as consistently corroborated secondary reporting on the same investigation, not independently re-verified against the original article's full (subscriber-only) text.

The Hidden Signal

A length-of-stay prediction is, on its own terms, a reasonable thing to build: an estimate, drawn from population data, of how long a similar patient's recovery has typically taken. The hidden step is what happens next, when that estimate is used not to inform a case manager's judgment but to set the boundary of what will be paid for, with the model's confidence interval quietly treated as a clinical determination. Consider a hypothetical scenario, smaller than the nH Predict case but illustrative of the same mechanism: a regional health plan builds a similar length-of-stay model for knee replacement recovery, intends it purely as a planning tool for its care-management team, and never explicitly decides whether an individual patient's doctor can override it without triggering a formal appeal. Within two review cycles, case managers report that overriding the model has become the exception rather than the routine judgment call it was meant to be, because nobody above them ever named that overriding it was supposed to remain routine.

Why the Visible Metric Misleads

An internal margin target, keep actual length of stay within a small percentage of the model's prediction, measures whether the organization is following its own model faithfully. It says nothing about whether the model's prediction was the right amount of care for a given patient, and it creates no natural signal when the two diverge, because divergence from the model is exactly what the target is built to suppress. The more revealing measure sits one layer over: the rate at which a model's own denials are reversed on internal appeal, tracked as its own standing metric rather than surfacing only inside litigation years later. A 90-percent-scale reversal rate, if UnitedHealth's own case-management data in fact showed one anywhere close to that scale, is not a rounding error in an otherwise sound process; it is close to an admission that the model's initial call was usually wrong, sitting unreviewed as a decision because no one had been assigned to compare denials against appeals as a matter of course.

Implication for Operators

A standing review comparing a utilization-management model's initial denials against its own internal appeal outcomes is not an exotic capability; it is closer to the same discipline Transformidy's decision-governance lens already asks of any organization relying on a metric to stand in for a decision, does anyone routinely check whether the model's calls survive a second look, and is that check built into ongoing operations or does it only happen when a lawsuit demands it. Every month a reversal-on-appeal pattern goes unreviewed as its own metric is a month it compounds, unnoticed, the way the STAT reporting shows this one appears to have compounded across roughly a year of tightening internal margin targets before the lawsuit surfaced it publicly.