UnitedHealth nH Predict
2023–ongoing · Estate of Lokken et al. v. UnitedHealth Group, Inc. et al.
AI denied elderly patients care with 90% error rate. Two deaths documented in federal lawsuit.
What Happened
Gene B. Lokken was 72 years old. He needed post-acute care following a hospitalization. His physicians believed he required continued time in a skilled nursing facility to recover safely.
An AI model called nH Predict disagreed.
nH Predict, developed by UnitedHealth Group's NaviHealth subsidiary and deployed across its Medicare Advantage insurance plans, determined how long patients like Lokken could remain in post-acute care facilities. Its determination overrode the recommendation of the physician of record.
Lokken was discharged. He died shortly afterward.
His estate is now a plaintiff in Estate of Lokken et al. v. UnitedHealth Group, Inc., Case No. 0:23-cv-03514, in the U.S. District Court for the District of Minnesota. A second plaintiff, 83-year-old Dale Henry Tetzloff, is also documented in the lawsuit. He also died after being discharged prematurely, according to the filing.
STAT News investigative reporting documented what the lawsuit alleged: nH Predict had a 90% error rate when its determinations diverged from treating physician recommendations. It appeared to apply a statistical average length of stay rather than individual clinical assessment. Physicians documented that their recommendations were routinely overridden. Patients had no meaningful way to challenge the determinations.
The Atlas Analysis
This is what happens when the Human Agency pillar fails at Level 0 in a system making life-affecting decisions at scale.
The Human Agency pillar asks three questions: Can the people affected by this AI understand its decisions? Can they challenge them? Are they protected from harm? In the case of nH Predict: no, no, and no.
Patients could not understand the criteria — the methodology was not disclosed. Physicians could not override it in practice — a 90% physician override rate documented in litigation demonstrates that clinical assessments were routinely overridden by the model's determination. Patients could not effectively appeal — the appeals process that existed was, according to the lawsuit, structured in a way that made meaningful recourse inaccessible to elderly patients with limited time, health, and resources.
A physician who is structurally unable to override an AI determination is not a human in the loop. They are a human in the theater. Presence and agency are different things. The nH Predict case is the clearest documented example of the difference.
The stated purpose of nH Predict was efficient care coordination — matching patients to appropriate levels of post-acute care based on clinical need. The documented operational effect, as alleged in the lawsuit and supported by investigative journalism, was systematic early discharge of Medicare Advantage patients, reducing UnitedHealth's cost of providing care.
No public methodology disclosure. No independent audit of error rates. No bias testing across patient demographic groups. No public reporting on how often physician recommendations were overridden. The 90% physician override rate that became central to the lawsuit was documented by investigative journalism — not by UnitedHealth's own quality reporting.
Signals That Would Have Caught It
The system makes life-affecting decisions with no meaningful patient recourse. Remediation required before deployment: a formal, accessible, time-bound appeals process with independent clinical review.
Physician override rate is not monitored or reported. A model whose determinations are overridden by treating physicians at high rates is producing incorrect outputs by definition. Remediation: continuous monitoring of physician override rates with escalation when rates exceed a defined threshold.
Methodology is not disclosed to affected parties. Patients and physicians cannot understand why the model reached its determination. Remediation: minimum methodology transparency sufficient for physicians to evaluate the basis of the decision.
Bias testing has not been conducted across patient demographic groups. A model applying a statistical average to a heterogeneous population will systematically produce worse outcomes for patients who deviate from the average — including the sickest patients, who are precisely those who most need post-acute care. Remediation: mandatory bias audit before deployment in clinical decision contexts.
What It Cost
The human cost is documented in the lawsuit: two deaths are on the record. The number of patients affected by systematic early discharge across all Medicare Advantage plans using nH Predict is not yet fully established — the lawsuit is ongoing.
The financial exposure is significant. UnitedHealth Group is the largest Medicare Advantage provider in the United States. A federal court finding of systematic AI-driven wrongful denial of medically necessary care, with deaths documented, represents material legal liability.
The regulatory and political cost is accelerating. Congressional scrutiny of AI in healthcare insurance has increased directly following this case. STAT News's investigative series — which documented the 90% error rate before the lawsuit was filed — remains one of the most cited pieces of AI accountability journalism in recent memory.
The Lesson
Human Agency is not the most technically complex pillar in the Atlas Framework. It is the most important one.
This case answers the question Atlas asks at the close of every assessment: would you trust this AI with your own money? Your child's education? Your medical diagnosis?
The answer nH Predict gave, at scale, to hundreds of thousands of elderly Medicare Advantage patients, was: it doesn't matter what you think. The algorithm has decided.
A framework that is designed to detect this failure before deployment rather than after lives have been lost is not a luxury. It is a minimum standard of care for any AI system that touches healthcare decisions.
References
- Estate of Gene B. Lokken et al. v. UnitedHealth Group, Inc. et al., Case No. 0:23-cv-03514, U.S. District Court for the District of Minnesota, filed November 2023.
- STAT News investigative series: AI algorithm denial rates in Medicare Advantage plans.
- CBS News: "UnitedHealth uses faulty AI to deny elderly patients medically necessary coverage, lawsuit claims."