What is nH Predict and what did it do?

nH Predict is an algorithm used by UnitedHealth subsidiary NaviHealth. A 2023 federal class action alleges it was used to deny post-acute Medicare Advantage care, overriding treating physicians, with roughly 90 percent of appealed denials reversed. The allegations remain unproven and UnitedHealth has contested the claims.

What governance control was missing from the nH Predict system?

The complaint alleges there was no hard checkpoint requiring a named clinician to sign each denial before it committed against a patient, and no per-decision audit record of who approved it and on what basis.

Would an AI approval gate have fixed the nH Predict problem?

An approval gate would have required a named clinician to sign each denial and created a durable audit record for every decision. It would not have fixed the underlying model accuracy; if a clinician rubber-stamped a flawed recommendation, the patient would still have been denied care.

UnitedHealth nH Predict: AI Medicare Denials Lawsuit

UnitedHealth's nH Predict lawsuit centers on a core allegation: that an algorithm, not a physician, became the effective decision-maker for Medicare Advantage post-acute care denials, with roughly 90 percent of appealed decisions later reversed.

In November 2023, a class action was filed in federal court in Minnesota against UnitedHealth and its subsidiary NaviHealth. The complaint alleges that the companies used the nH Predict algorithm to override treating physicians and deny post-acute care to elderly Medicare Advantage patients, according to CBS News.

The suit alleges the model carried an error rate of roughly 90 percent, a figure the plaintiffs tie to the share of denials that were reversed on appeal. The same filing alleges that only about 0.2 percent of patients appealed at all, which the plaintiffs say the companies relied on. Most denied patients absorbed the decision rather than contesting it.

In 2025 and 2026 the litigation advanced. The court allowed breach of contract claims to proceed, as reported by Healthcare Finance News, and a federal judge ordered broad discovery into the model itself, including how it was built and how its outputs drove coverage decisions. The allegations remain unproven, and UnitedHealth has contested the claims.

The governance gap: what actually failed

The core allegation is not that a model produced a prediction. It is that the prediction was allowed to become a binding coverage denial without a named clinician standing behind it. A denial of post-acute care is a consequential, hard-to-reverse action against a patient. The complaint alleges the algorithm functioned as the effective decision-maker, and that physician judgment was displaced rather than confirmed.

Two structural gaps follow from that. The first is the absence of a hard checkpoint between a generated recommendation and a committed denial. Nothing required a qualified human to sign each denial before it took effect against the patient's coverage. The second is the absence of a clean per-decision record of who approved each denial and on what basis. When the court ordered discovery, the plaintiffs had to fight for visibility into the model and its outputs, rather than reading an account that already existed.

This is the same shape that appears across automated decisioning. A system that can commit an irreversible decision against a person, with no gate and no durable record of the responsible human, concentrates harm in the small fraction who appeal and leaves the rest with no recourse.

How MakerChecker changes the outcome

MakerChecker governs the action an automated actor is allowed to take. It does not judge whether a denial is medically correct. It governs whether a denial may commit, who must sign it, and what is recorded.

Model the committed denial as a separate high-risk skill that is forced through an approval gate. The model may produce a recommendation, but the action that binds against the patient, coverage.deny, cannot run until a named clinician signs it:

skill: coverage.deny
risk_tier: high
gate:
  approvals_required: 1        # a named clinician must sign each denial
  forbid_requester: true       # the proposing system cannot finalize

Segregation of duties through forbid_requester enforces the separation the complaint says was missing. The system that proposes a denial cannot also be the party that finalizes it. A human in a clinical role has to take the consequential step, on the record.

Least privilege keeps the proposing role scoped to recommendation only. A model-proposer role is granted the skill that drafts a recommendation, and is not granted coverage.deny at all. The dangerous action lives behind the gate, not in the model's reach:

role: model-proposer
grants:
  - skill: coverage.recommend
    version: "1.0"
    risk_tier: low
  # coverage.deny is NOT granted to this role

Every grant check, every gate, and every signed approval is written to the tamper-evident, Ed25519-signed, hash-chained audit. The reviewer and the basis for each denial are recorded as the decision is made. That per-decision record is precisely the evidence the court had to force into discovery. With it, the question of who approved a given denial and why is answered from a signed log that can be verified offline, not reconstructed years later.

What MakerChecker would not fix

MakerChecker would not fix the alleged 90 percent error rate. It does not make the model more accurate, and it does not evaluate the clinical merits of a recommendation. If a clinician reviews a flawed recommendation and signs the denial anyway, the patient is still denied care and the harm still occurs.

The control delivers accountability, not accuracy. It guarantees that a named human took the consequential step, that the proposing system could not finalize its own output, and that the basis was recorded. It does not guarantee the decision was right. A gate that a reviewer rubber-stamps is a weaker control than one a reviewer reads, and MakerChecker cannot compel genuine review. What it removes is the structural condition the complaint describes, in which an automated output became a binding denial with no responsible human and no durable record behind it.

See the configuration: examples/rogue-ai/unitedhealth-nhpredict-ai-medicare-denials

Frequently asked

What is nH Predict and what did it do?: nH Predict is an algorithm used by UnitedHealth subsidiary NaviHealth. A 2023 federal class action alleges it was used to deny post-acute Medicare Advantage care, overriding treating physicians, with roughly 90 percent of appealed denials reversed. The allegations remain unproven and UnitedHealth has contested the claims.
What governance control was missing from the nH Predict system?: The complaint alleges there was no hard checkpoint requiring a named clinician to sign each denial before it committed against a patient, and no per-decision audit record of who approved it and on what basis.
Would an AI approval gate have fixed the nH Predict problem?: An approval gate would have required a named clinician to sign each denial and created a durable audit record for every decision. It would not have fixed the underlying model accuracy; if a clinician rubber-stamped a flawed recommendation, the patient would still have been denied care.