Domain Atlas / Clinical decision support & deterioration alerting
nH Predict Utilization Review
Internal records subpoenaed by the U.S. Senate Permanent Subcommittee on Investigations show that early-2021 testing of an auto-authorization model inside UnitedHealthcare produced faster handle times together with an increase in adverse determination rate - attributed to finding contraindicated evidence missed in original review - and the internal committee voted to tentatively approve the model at the following meeting; the April 2021 approval of 'Machine Assisted Prior Authorization' was paired with testing that removed six to ten minutes from the average review while the reviewing doctor or nurse still had to verify that the primary evidence is acceptable. Over the same period the insurer's post-acute prior authorization denial rate went from 10.9 percent (2020) to 16.3 percent (2021) to 22.7 percent (2022), and its 2019 skilled-nursing-facility denial rate was nine times lower than its 2022 rate. A January 2022 vendor presentation shows a naviHealth care coordinator completing nH Predict to determine optimal post-acute placement while the patient is still hospitalized, and an April 2022 vendor instruction told call handlers not to guide providers on the questions used to collect the information determinations are made from.[2]
What happened
Every other case in this domain is a bedside alert: a model warns a clinician about the patient in front of them. This one sits on the other side of the claim. When a Medicare Advantage enrollee is ready to leave an acute hospital for a skilled nursing facility, an inpatient rehabilitation facility or a long-term acute care hospital, the plan decides whether that post-acute care is covered, and for how long. From 2021, UnitedHealthcare's Medicare Advantage post-acute benefit was managed by naviHealth — acquired by Optum, part of the same corporate parent, in May 2020 — and a January 2022 naviHealth presentation obtained by the Senate shows how the pipeline opens: a "naviHealth Care Coordinator completes nH Predict" to "determine optimal [post-acute care] placement" while the patient is still in the hospital. nH Predict compares the specific patient with similar prior patients and returns an estimated amount of post-acute care and a projected discharge date, drawn from a vendor-owned comparison cohort of stays that actually happened — at lengths that earlier coverage determinations from this same pipeline had bounded. Requests reach the review through the plan's online provider portal and move to the vendor across three named platforms — the nH Access portal, the PAAN system and the ICUE system; in April 2022 naviHealth instructed its call handlers, "Do NOT guide providers or give providers answers to the questions" used to collect the information determinations are made from.
The Senate Permanent Subcommittee on Investigations opened its inquiry in May 2023 and published "Refusal of Recovery" in October 2024, built on more than 280,000 subpoenaed pages from insurers together covering nearly 60% of Medicare Advantage enrollment. What the internal record shows is the pairing this case file exists to carry. In early 2021 UnitedHealthcare tested an "HCE Auto Authorization Model"; committee minutes record initial testing producing "faster handle times" together with "an increase in adverse determination rate," attributed to finding contraindicated evidence missed in original review — and the committee voted to tentatively approve the model at the following meeting. In April 2021 the same committee approved "Machine Assisted Prior Authorization," was told the reviewing doctor or nurse "still had to verify that the primary evidence is acceptable," and was told in the same record that testing had removed six to ten minutes from the average review. Over that period the plan's post-acute prior authorization denial rate went from 10.9% in 2020 to 16.3% in 2021 to 22.7% in 2022; its 2019 skilled-nursing-facility denial rate was nine times lower than its 2022 rate. STAT's "Denied by AI" investigative series (from March 13, 2023) named nH Predict publicly and reported a 2023 naviHealth target of keeping rehab stays within 1% of the days the algorithm projected — an internal target, not a measured behavior. naviHealth's spokesperson answered in the series opener that "The naviHealth predict tool is not used to make coverage determinations. The tool is used as a guide." The brand did not survive the scrutiny: STAT reported in October 2023 that from Q1 2024 Optum's post-acute-care companies would be called Home & Community Care, and the naviHealth name was retired.
Three outside channels then wrote the rest of the record, each acting on a different surface. The regulator wrote order-of-operations rules rather than banning the tool: under CMS-4201-F and the February 6, 2024 Health Plan Management System (HPMS) FAQ memo, an algorithm may "assist" a coverage determination, the plan remains responsible for compliance, an algorithm that determines coverage "based on a larger data set instead of the individual patient's medical history, the physician's recommendations, or clinical notes" would not comply with 42 CFR 422.101(c) — Title 42, Code of Federal Regulations (CFR) — and a predicted length of stay "alone cannot be used as the basis to terminate post-acute care services" — only re-assessing the individual patient's condition can support the termination notice. The courts narrowed and sustained: in Estate of Gene B. Lokken v. UnitedHealth Group (D. Minn., No. 0:23-cv-03514), Judge John R. Tunheim's February 13, 2025 order held all claims subject to Medicare administrative exhaustion but waived it as futile, found most claims preempted, allowed Count 1 (breach of contract) and Count 2 (implied covenant) to proceed, and dismissed Counts 3-7 with prejudice; in Barrows v. Humana (W.D. Ky., No. 3:23-cv-654-RGJ), Judge Rebecca Grady Jennings' August 14, 2025 order let Counts 1 (breach of contract), 2 (implied covenant), 3 (unjust enrichment) and 8 (common-law fraud) proceed and dismissed Counts 4-7 with prejudice. Both cases are in active discovery as of August 2026 — Lokken with a defense motion to amend set for hearing September 1, 2026, Barrows with class certification due October 15, 2026 — and there is no dispositive merits ruling, no settlement, and no finding that nH Predict was or was not used to make determinations. The Lokken order recites, as pleaded allegations only, an allegation about the share of claim denials reversed on appeal, and records in the same paragraph that "UHC denies any use of nH Predict." And the measurement channel published its numbers: KFF's analysis of the plans' own federal reporting found that of 46.2 million Medicare Advantage prior authorization determinations in 2022, 7.4% were denied in whole or in part, 9.9% of denials were appealed, and 83.2% of appeals overturned the initial decision. The closest ground-truth audit of the decision class remains the 2022 evaluation by the Department of Health and Human Services (HHS) Office of Inspector General (OIG): 13% of sampled 2019 prior authorization denials, across fifteen pooled organizations, met Medicare coverage rules — a study that predates naviHealth's management of this benefit and measures the class, not this deployment.
The sociotechnical reading
The governed object here is not a model's accuracy — no deployment error rate has ever been published — but the order of operations inside a coverage determination: what the reviewer sees first, how long the reviewer has, and what must be checked before an adverse determination leaves the building. The subpoenaed committee minutes are the rare direct evidence of the trade this domain usually only infers: review time and the adverse-determination rate moved together, the body with authority to say no saw both numbers, and it approved. The map draws that honestly rather than accusatorially — the utilization-management committee is present, informed and minuted, its approval edge stronger than its check — because the Senate's own third recommendation aims at exactly that surface: expand committee regulation so predictive tools do not unduly influence human reviewers. The prediction channel's structural fact is stranger than bias: the coordinator completes the tool rather than reviewing its output, so the estimate is the frame the assessment is entered into, and the comparison cohort it draws on is fed by this pipeline's own past determinations — a store that manufactures its own ground truth, drawn on the diagram as the record-to-cohort loop. The regulator's response fits the topology: CMS did not ban the estimator, it sequenced it — assist is permitted, and termination requires re-assessing the individual patient first — which is why the Lab models the CMS instrument as authority over the check on the notice, never as a promised reduction in error.
The correction that demonstrably works is the one with the narrowest reach, and this case supplies the catalogue's cleanest version of that pair: 83.2% of appeals overturned the initial decision in 2022, and 9.9% of denials were appealed. Per item the channel corrects most of what reaches it; almost nothing reaches it. The diagram carries the pair as a strong check edge fed by the thinnest inbound pathway on the board, because averaging the two facts into one number is precisely the misreading the record warns against — and the appeal figures are program-wide, conditioned on self-selection, which is why they order the edges rather than set them. What the record refuses to say matters as much: whether nH Predict made coverage determinations is contested in two live class actions and asserted by no court; the pleaded overturn share is a pleading recited under the motion-to-dismiss standard and contradicted in the same paragraph; the December 2022 workgroup that explored using appeal data to identify appeal-likely denials was a workgroup with two meetings, not a deployed system, and is deliberately not drawn. The record's disparities are between service categories and between insurers — post-acute denial rates at multiples of overall rates — never between subpopulations of served people, so no subpopulation harm is modeled or manufactured. The boundary holds as always: enrollees, their stays, their health and their bills are what this pipeline decides about, and nothing on the diagram computes any of it. What the Lab can show is the machinery those decisions travel through — a vendor between the plan and the review, a time budget cut as a verification duty was restated, an aggregate that could not express where denials concentrated, and an appeal channel doing most of the system's measured correcting from outside the building, one-tenth of the time.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.