UnitedHealth used an algorithm with a 90% error rate to deny post-acute care to elderly Medicare Advantage patients

What happened

In May 2022, Gene B. Lokken, a 91-year-old man in Wisconsin enrolled in a UnitedHealthcare Medicare Advantage plan, fell at home and fractured his leg and ankle. His doctors admitted him to a skilled nursing facility for rehabilitation. He was elderly, frail, recovering from significant orthopaedic injury. The care his physicians prescribed was the kind of post-acute care Medicare Advantage was designed to pay for.

On or about 20 July 2022, UnitedHealthcare terminated his coverage. The decision was supported by a prediction from a piece of software called nH Predict, which UnitedHealth had acquired in its 2020 purchase of naviHealth. nH Predict had been built on a database of approximately six million patients, and produced estimates of how long any given patient should require in a skilled nursing facility before discharge. The software predicted that Lokken's stay should be shorter than what his treating physicians had prescribed. The coverage was cut.

Lokken's family did not stop the care. They paid out of pocket, between US$12,000 and US$14,000 a month, for almost a year. He died on 17 July 2023.

In November 2023, his estate, along with the estate of another deceased patient named Dale Henry Tetzloff, filed a class-action complaint in the United States District Court for the District of Minnesota. The complaint named UnitedHealth Group, UnitedHealthcare and naviHealth as defendants. The central allegation: UnitedHealth was using nH Predict to override the clinical judgment of treating physicians and deny post-acute care to elderly Medicare Advantage members, knowing that the model's predictions were unreliable.

The complaint cited an alleged error rate of approximately 90 percent. Nine out of ten denials, when patients or families appealed, were overturned. The model was wrong nine times out of ten about a question on which its prediction was the basis for cutting off coverage for the most vulnerable patients in the system. The complaint further alleged that UnitedHealth's internal data showed that only 0.2 percent of policyholders appeal a coverage denial, and that this asymmetry, between the error rate and the appeal rate, was both known to UnitedHealth and the operational reason the model continued to be used.

UnitedHealth's public response framed nH Predict as a guide rather than a decision-maker. Coverage determinations, the company has maintained, are made on the basis of Centers for Medicare and Medicaid Services (CMS) coverage criteria and the terms of each member's plan; the algorithm informs providers and families about what assistance the patient may need, in the facility and after discharge. The complaint alleges that in practice, denials follow the model's predictions and override the treating physicians who disagree. The case has survived motions to dismiss. As of March 2026, the court ordered UnitedHealth to disclose the algorithm itself for forensic review. The disclosure process is ongoing.

What an auditable version would have shown

The case turns on whether the model's prediction or the treating physician's clinical judgment is the basis on which coverage was actually denied. UnitedHealth's position is that the physician's judgment governs. The complaint's position is that the model's prediction governs. The fact that the parties disagree on the basic question, whose decision is this, is the symptom of an inadequate record.

An auditable coverage-denial workflow would produce, for each decision, a structured record naming the inputs the algorithm consulted, the model version that ran, the output it produced, the human reviewer who saw the output, the clinical evidence the reviewer also considered, and the reasoned justification for the final decision. The record would distinguish, in plain text, between cases where the human reviewer concurred with the algorithm and cases where the human reviewer overrode it. The pattern across the population would be queryable: how often does the reviewer concur, how often does the reviewer override, how often does the appeal succeed, segmented by reviewer, by region, by model version.

Either version of the truth would be visible in the structured record. If the plaintiffs are right and the model effectively determines denials with the human reviewer as a rubber stamp, the notes are sparse, the clinical evidence cited is a copy of the algorithm's output, and concurrence rates run close to 100 percent. If UnitedHealth is right and the reviewer makes the decision with the algorithm as one input, the notes are substantive and the reviewers and the algorithm visibly disagree on some non-trivial fraction of cases. The data answers the question.

The data was not kept that way. The case is being litigated, instead, through expert testimony reconstructing what the algorithm did from training data and discovery materials. That reconstruction is slow, expensive and ambiguous, and none of it is good for the patients.

Where the gap was

The gap is in three places.

The clinical-decision layer. The treating physician's judgment is the medical-legal record of what care the patient needs. If a coverage decision overrides that judgment, the record of the override should name the algorithm consulted, the criteria applied, and the reasoning. CMS regulations require coverage decisions to be reasoned and individualised. The complaint's allegation is that nH Predict outputs were treated as the reasoning, in substance, with the documentation framing them as something else.

The appeal-rate asymmetry. A model with a 90 percent error rate at point of denial, combined with a 0.2 percent appeal rate, produces a system in which the model is effectively making decisions for the vast majority of affected patients. The asymmetry is structural, most Medicare Advantage members are elderly, often cognitively impaired, frequently without family help, and the appeal process is opaque. CMS oversight depends on the patterns the regulator can see. Without per-decision structured records aggregated for population-level review, the asymmetry is not visible to the regulator until a lawsuit surfaces it.

The disclosure layer. UnitedHealth has resisted disclosing nH Predict's internals. The March 2026 court order compelling disclosure is the precedent the litigation will turn on. Algorithms used to make or substantially inform coverage decisions need to be inspectable by the patients they affect and the regulators that licence the insurers. Trade-secret protection does not stop being a real interest at the point a decision affects whether someone receives medical care, but it cannot trump that interest either.

What governance should have looked like

Every coverage decision that overrides a treating physician's judgment is written to a signed conduct record at the moment of the decision. The record names the algorithm consulted, the version that ran, the output it produced, the human reviewer's notes, the CMS criteria applied, and the rationale for overriding the physician. The record is appended to the member's file and is inspectable on appeal.

from headlights import ConductRecord, ConstraintGate, sign, chain
from datetime import datetime, timezone

# Coverage denial that overrides a treating physician requires a recorded
# review path. The gate refuses to release the denial without one.
gate = ConstraintGate(
    overrides_treating_physician=True,
    requires_named_human_reviewer=True,
    requires_cms_criteria_citation=True,
)

record = ConductRecord(
    decision_type="post_acute_care_coverage_denial",
    member_id_hash=sha256(member_id),
    treating_physician_recommendation="continue_skilled_nursing_facility",
    treating_physician_id=physician_id,
    algorithm_consulted="nh_predict",
    algorithm_version="v4.7",
    algorithm_output={"recommended_discharge_in_days": 0,
                      "confidence": 0.62,
                      "patient_cluster_id": "C-2104-A"},
    human_reviewer_id=reviewer_id,
    human_reviewer_notes=reviewer_notes_blob,
    cms_criteria_cited=["42 CFR 409.32"],
    rationale_plain_english=(
        "Algorithm predicts discharge readiness. Reviewer concurs based on "
        "the absence of documented physical therapy progress in the last 7 days."
    ),
    member_appeal_rights_communicated=True,
    timestamp=datetime.now(timezone.utc),
    previous_record_hash=last_record.hash(),
)

# Gate refuses to release if any required field is empty.
gate.check(record)

signed = sign(record, key=insurer_private_key)
chain.append(signed)

For every coverage decision that overrides a treating physician, the insurer should be producing a structured record at the moment the decision is made. The record names the algorithm consulted, the inputs it ran on, the version that ran, the human reviewer's notes, the CMS criteria applied, and the reasoning for overriding the physician. The record is appended to the member's file. It is available to the patient on appeal without a court order. The record exists whether or not the patient ever appeals.

That alone changes the case. CMS can monitor the population of records and see, in routine analytics, when an algorithm and its human reviewer concur at a rate too high to be plausible. The pattern between a 90 percent error rate on appeal and a 0.2 percent appeal rate stops being a courtroom revelation and becomes a dashboard metric that the regulator watches in close to real time. Whistleblower cases and class actions stop being the only mechanism through which the public learns what its insurers are doing.

The algorithm itself also has to be inspectable. Not the training data, that can stay proprietary. The architecture of the inputs and the outputs the patient was subjected to, the version pinning, the override logic. The court's March 2026 disclosure order against UnitedHealth is, for that reason, the precedent the rest of the industry will be measured against. Every insurer using AI in coverage decisions now knows it might be required to hand the algorithm over to plaintiff counsel. The ones who have built the structured record can show what the algorithm actually did. The ones who have not are arguing, after the fact, that whatever happened in those decisions was reasonable. Lokken is the first major test of whether that argument will hold.