The Lawsuit Alleging an Algorithm, Not a Doctor, Decided When Grandma's Care Ended

This account draws on reporting by the Star Tribune, Becker's Payer Issues, Digital Healthcare Law, and public federal court filings. The HealthTech Signal has reviewed the underlying court order referenced below.

Families of deceased Medicare Advantage patients allege that UnitedHealth Group used an algorithm to cut off coverage for skilled nursing and rehabilitation care, overriding treating physicians' recommendations that patients needed to stay longer. The case, consolidated in Minnesota federal court, has survived a motion to dismiss and, as Becker's Payer Issues reported in September 2025, a federal judge denied UnitedHealth's attempt to narrow the scope of discovery, meaning the case is now digging into the internal data behind the algorithm itself.

UnitedHealth disputes the core allegation directly. The Star Tribune quoted the company saying the lawsuit is based on "unfounded allegations" and insisting that medical directors, not artificial intelligence, make coverage decisions. That is precisely the factual dispute now being tested through discovery: whether the algorithm was a decision support tool that human reviewers meaningfully considered, or whether it functioned, in practice, as the decision.

What the algorithm is alleged to do

The tool at the center of the case, developed by a company called naviHealth and known as nH Predict, is designed to estimate how many days of post-acute care, such as skilled nursing facility stays after a hospitalization, a patient with a given diagnosis is likely to need, based on a database of prior patients. The plaintiffs allege that UnitedHealth used these statistical predictions as an effective cap on coverage, denying continued care once a patient hit the algorithm's predicted length of stay, even when the patient's own doctors said more time was clinically necessary.

The plaintiffs' complaint, according to the court's own opinion dismissing certain claims while allowing others to proceed, alleges an internal error rate for the algorithm's predictions high enough that human reviewers who deviated from it faced pressure not to. In February 2025, the federal judge dismissed several state law claims but allowed breach of contract and good faith and fair dealing claims to proceed, meaning the core allegation, that UnitedHealth breached its coverage obligations by relying on the algorithm's output rather than individualized medical necessity review, is now headed toward further fact-finding rather than an early dismissal.

The clinical stakes of a length of stay prediction

A statistical model that predicts, on average, how many days of rehab a hip fracture patient with a given set of comorbidities typically needs is not inherently unreasonable as an input into a coverage decision. Population-level data genuinely can help identify overutilization and inform benchmarks. The clinical problem arises when an average, built from a population, gets applied to an individual patient whose actual recovery trajectory diverges from it, which by definition happens to roughly half of any population sitting above or below a median prediction.

Medical necessity determinations under Medicare Advantage are supposed to be individualized: a physician assessing this specific patient's functional status, comorbidities, and recovery progress. The allegation here is that a population average became a de facto ceiling applied to individuals regardless of their actual clinical trajectory, and that human review, while nominally present, was not meaningfully independent of the algorithm's output.

Why this case matters beyond one insurer

This is not an isolated dispute. UnitedHealth is the largest Medicare Advantage insurer in the country, and naviHealth's tool, or close analogues from other vendors, are used widely across the Medicare Advantage industry to manage post-acute care utilization. If discovery in this case surfaces evidence that the algorithm functioned as a hard denial trigger rather than a decision aid, it will not just be evidence against one company. It will be evidence about how an entire category of AI-assisted utilization management tools actually operates inside the insurance industry, versus how it is described in public compliance documents.

Medicare Advantage plans are already under scrutiny from CMS for elevated denial rates on post-acute care compared with traditional Medicare, and this litigation is unfolding alongside that broader regulatory pressure. A ruling against UnitedHealth, or a damaging discovery revelation, would land at a moment when regulators are already primed to act on AI-driven utilization management more broadly.

The systemic tension

There is a legitimate case for using predictive tools in utilization management: healthcare spending on unnecessary extended stays is real, and insurers manage risk pools that require some form of utilization control to remain solvent. The tension is not whether insurers can use data to inform coverage decisions. It is whether an automated prediction, built for population-level accuracy, can be deployed in a way that preserves meaningful, individualized clinical judgment for each patient, or whether the economic incentive to control costs inevitably pulls the "human review" step toward rubber-stamping the algorithm's output.

That question cannot be answered by press releases from either side. It requires exactly the kind of internal data, error rates, override rates, and reviewer time-per-decision metrics, that this litigation's discovery process is now compelling UnitedHealth to produce.

The takeaway

This case has not been decided, and UnitedHealth's denial that AI drove these determinations deserves to be taken seriously until the evidence says otherwise. But the fact that a federal judge has allowed the core contract claims to proceed, and specifically ordered broader discovery into the algorithm's internals, tells you this is not a frivolous dispute. It is the first major test of a question every health system and insurer deploying predictive AI in coverage decisions will eventually have to answer under oath: was the human in the loop actually deciding, or just signing off.