nH Predict sets one patient's discharge date from six million others

What nH Predict was built to compute

nH Predict is software built by naviHealth, a company that UnitedHealth's Optum unit bought in 2020 for $2.5 billion. It reads a few facts about a patient: diagnosis, age, living situation, physical function. It finds people who look similar in a database of six million past cases. Then it returns an estimate of how many days this patient will need to recover after the hospital, and a target date to send them home.¹

The care in question is post-acute care, the nursing-home and rehabilitation care a person needs after a hospital stay, such as relearning to walk after a stroke or healing a fractured leg in a skilled nursing facility. Most of the patients here are on Medicare Advantage, the private insurance plans that cover many Americans on Medicare. Under these plans the government pays the insurer a fixed sum to manage each member's care, so every extra day of nursing care the plan approves is a day the insurer itself pays for. And approval is required: before that care continues, the plan has to authorize it, a sign-off the industry calls prior authorization. Prior authorization is where the coverage decision gets made.

In November 2023, a class action filed in federal court in Minnesota, Estate of Gene B. Lokken v. UnitedHealth Group, said the company used nH Predict's target dates to make those decisions, ending coverage on or near the date the model picked and overriding the judgment of the patients' own doctors.² The suit is still being litigated. No one disputes what the tool produces. The fight is over what it was allowed to decide.

The days two patients did not get

Gene Lokken was 91. In May 2022 he fell, fractured his leg and ankle, and moved into a skilled nursing facility in Tomahawk, Wisconsin, to recover. His UnitedHealth plan paid for about nineteen days of that stay and then stopped, around July 20, 2022. His doctors wanted him to stay longer. The plan rejected the family's appeal. For the year that followed they paid roughly $12,000 to $14,000 a month out of pocket. He died on July 17, 2023.²

Dale Henry Tetzloff was 74. He had a stroke in October 2022, and his doctor referred him for at least 100 days of skilled nursing care. The plan denied continued coverage after about 20 days. On appeal, a UnitedHealth physician agreed he needed more time. He was denied again after 40 days. His family spent more than $70,000 over about ten months before he was discharged to assisted living. He died on October 11, 2023.²

One physician ordered at least 100 days. A second, reviewing the appeal on the insurer's own side, agreed the patient needed more. Coverage ended at 40.

How an average becomes a cap

The estimate nH Predict produces is an average. Match a patient to six million others and you get a number that describes the group: on average, patients who look like this go home in so many days.¹ That average can be a good guess about the group and still be wrong about the person in the bed, because a real patient is not the average of a crowd. This paper has a lesson on why a number measured across one population can mislead when it is applied to a single different case, called distribution shift. A recovery estimate used as one patient's deadline is a sharp instance of it.

An estimate becomes a cap when someone stops treating it as a guess and starts treating it as the answer. Internal targets did exactly that, the complaint and STAT News reporting say. naviHealth set a goal for 2023 to keep patients' rehab stays within 1% of the days the algorithm projected, and case managers who fell outside that narrow band risked discipline, up to being fired, whether or not the extra days were justified under Medicare's rules.³ When a reviewer is punished for departing from the model's date, the date has stopped being advice.

The complaint alleges that more than 90% of these denials, counting only the patients who appealed, were reversed on internal appeal or before a federal administrative law judge.² The figure is an allegation, pleaded "upon information and belief," with no denominator and no outside source behind it, so treat it as a claim the suit will still have to prove, not as a fact. But follow its logic. If nearly every denial that gets challenged is overturned, then most of the challenged denials were wrong. And almost no one challenges: appeals are slow, and most patients never file one. The complaint puts the share who appeal at about 0.2%, though its own footnote traces that figure to a study of a different insurance market, so lean on it lightly. A denial that is wrong and never appealed simply stands. A reversal rate that high among the few who do appeal implies a much larger body of wrong denials that no one ever touched. That is the tell that a population average was doing work it could not do, capping the care of individuals whose own records said they needed more.

CMS, the federal agency that regulates these plans, issued that guidance in February 2024, after the events in the suit, so it is the regulator's stated position and not a ruling against UnitedHealth. It states the exact move the mechanism turns on: a predicted length of stay, on its own, cannot end coverage.⁴

The dispute discovery would settle

UnitedHealth's answer is that none of this describes what nH Predict does.

The plaintiffs say the guide was decisive in practice. The 1% target and the discipline behind it, they argue, made the model's date the operative decision, and the members' policies said coverage determinations would be made by clinical staff rather than by an algorithm.³

Two things would settle which account is right, and both sit in discovery, where the court has ordered UnitedHealth to open its records. The first is the internal directions: how tightly the 1% target actually bound reviewers, and whether departing from the tool was really punished. The second is the denial-and-appeal data with its denominators: how many post-acute denials there were, how many patients appealed, and whether reversals corrected the original decision or reflected genuinely new facts. The complaint's 90% is precisely the number that data would confirm or break.

Away from the suit, a Senate investigation found the denials climbing as the automation spread. The Permanent Subcommittee on Investigations reported in October 2024, drawing on more than 280,000 pages of the insurers' own documents, that UnitedHealthcare's denial rate for post-acute prior authorization roughly doubled in three years.⁶

Those numbers do not come from the litigants, but the subcommittee stopped short of the causal claim. It described nH Predict only as an algorithm "linked in media reports" to denials, not as a proven cause.⁶ The case itself is moving. In February 2025 a judge let the core contract claims proceed while dismissing most of the other counts, according to a legal-industry summary of the order, which was the reading available; the underlying opinion could not be retrieved directly.⁷ The case remains open, with class-certification steps scheduled into September 2026.⁸ No court has yet ruled on whether nH Predict drove the denials.

The same move at Cigna

The nH Predict pattern, taking a prediction built to describe a group and using it to decide one person's case, is not unique to UnitedHealth or to nursing care. In 2023, ProPublica and The Capitol Forum reported that the insurer Cigna used a system called PXDX to deny more than 300,000 payment requests over two months, at an average of about 1.2 seconds per case. One medical director denied roughly 60,000 in a single month. A former Cigna doctor described the work plainly: "We literally click and submit. It takes all of 10 seconds to do 50 at a time." Cigna called the reporting biased and incomplete and said the system speeds payment for routine claims.⁹

The common thread runs through both. Wherever a number built to summarize a population is wired straight to a decision about one person, and the people it decides against cannot see it or easily contest it, the same failure waits. Automated prior authorization is spreading across insurers, and every system that adopts it inherits the question this suit asks: is the model a guide, or the decision.