UnitedHealth nH Predict

What They Said

UnitedHealth’s NaviHealth subsidiary marketed nH Predict as a clinical decision-support tool — a way to bring “evidence-based” precision to post-acute care planning. The algorithm, built on data from a proprietary database of 6 million patient cases, would predict how long a Medicare Advantage patient needed in a skilled nursing facility, rehab hospital, or home health program after discharge.

UnitedHealth acquired NaviHealth in 2020 for an estimated $2.5B. The pitch to investors and providers was straightforward: a data-driven tool would reduce unnecessary care, standardize length-of-stay decisions, and align post-acute spending with clinical evidence. The word “support” was load-bearing.

What Actually Happened

In November 2023, STAT News published an investigation showing that nH Predict’s recommendations were being treated as ceilings, not guidance. Case managers were pressured to keep stays within 1% of the algorithm’s predicted length, regardless of physician judgment. Internal documents showed UnitedHealth knew the model’s recommendations were overturned on appeal more than 90% of the time — and continued to use them as the basis for denials anyway.

A federal class action was filed in Minnesota that same month on behalf of the estates of Gene Lokken and Dale Tetzloff, two elderly UnitedHealth members whose post-acute care was cut off based on nH Predict outputs. The complaint alleges UnitedHealth used the algorithm specifically because it knew most patients would not appeal.

Senate Permanent Subcommittee investigators followed in 2024, publishing a report documenting that prior authorization denials at UnitedHealth’s Medicare Advantage business roughly tripled between 2019 and 2022 — coinciding with the rollout of nH Predict and similar automated tools at Humana and CVS. The Centers for Medicare and Medicaid Services issued new rules in 2024 requiring that algorithms used in coverage decisions reflect individualized clinical circumstances, not population averages.

The Root Cause

The algorithm was accurate enough to be defensible and inaccurate enough to be lethal. nH Predict was not catastrophically broken in the technical sense. It was a length-of-stay regression on a large dataset. The failure was in how its outputs were institutionalized — converted from a clinical hint into a hard contractual gate, with case managers performance-reviewed against algorithmic alignment rather than patient outcomes.

The second failure was incentive design. The same business unit that benefited financially from shorter stays controlled the tool that determined stay length. There was no independent clinical authority empowered to override the model. By the time appeals reversed the denial, the patient had already been discharged, gone home without care, fallen, and been readmitted — or died.

The Pattern to Watch For

If your AI vendor cannot tell you the appeal-reversal rate of decisions their model influences, you have no idea whether the model works. Accuracy on the training set is irrelevant. The only metric that matters for any decisioning model in a regulated environment is what happens when a human reviews the call. A 90% reversal rate is not a model — it is a litigation pipeline.

What You Should Steal

Build the reversal-rate dashboard before you ship the model. Every algorithmic decision that affects a customer, employee, patient, or claim should be tracked against its eventual human-reviewed outcome. If the gap between model output and human judgment exceeds a defined threshold, the model retrains or gets pulled. UnitedHealth had this data. They chose not to act on it. The discipline is not collecting the metric — it is wiring the metric to an automatic shutoff.