nH Predict Utilization Review

What this models

This example runs on the nH-Predict-class payer utilization review network: 12 components and 25 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 1 assumed · 9 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  Topology. Twelve components, all documented, none decorative - and the payer-side shape is the point: every other org in this domain is a bedside alert read by a treating clinician, while this network's governed object is the order of operations inside a coverage determination. Two model nodes because the record documents two distinct channels (a vendor-operated similar-patient stay estimator completed during the acute stay, and a plan-side machine-assisted case-file review approved by an internal committee in April 2021). An enforcement node because a recorded adverse determination automatically issues as a denial or termination notice - the downstream action the regulator's 2024 guidance constrains. Three absences are equally derived: no worklist (the documented pressure is per-review time compression, not a queue or backlog), no guardrail (no automated output screen is documented - the documented pre-determination control is human verification of primary evidence, carried on the operator pathways), and no external boundary (the vendor platform is a contracted part of the governed pipeline, drawn as nodes and stores, and the record documents no egress beyond it). The diagram is drawn at the coarsest granularity that keeps every documented mechanism distinguishable: where one documented act travels two ways - the coordinator completing a tool that frames what they enter, or a tool contribution reaching the decider inside the record the rule tells that class to read - it is drawn once, on the pathway the record documents as consequential, and the other direction is narrated in that pathway's description.

• baseline

  Heavy workload against limited capacity. Workload: a 46.2-million-determination program-wide request stream in 2022, this plan's post-acute denial rate tripling from 10.9% to 22.7% over 2020-2022, a review workflow reorganized to be faster (six to ten minutes removed from the average review), and an intake the vendor instructed its own call handlers not to help complete. Capacity above the very-low floor: the human channel is a real, separately-mandated clinical function - first-level nurse review plus a medical-director duty the federal rule names. Capacity held below full: the same record documents the review time budget being cut as the check was introduced, a regulator restating the individualized re-assessment duty in February 2024, and the one correction with a measured effect sitting outside the plan and reaching 9.9% of denials.

• baseline

  Baselines mirror the PAN org's evidence-derived pathway widths as strength rungs (the full mapping is recorded in the derivation notes). The two structural contrasts are deliberate and documented. First, the cohort pathway into the estimator runs at full strength while the individual record's pathway runs faint, because the regulator's stated objection to tools of this class is that they determine coverage from a larger data set instead of the individual patient's medical history, physician recommendations and clinical notes - the width asymmetry is that sentence, drawn. Second, the appeal channel's per-item effect runs at a substantial level (83.2% of 2022 appeals overturned the initial decision, held below that figure because appeals are self-selected) while its inbound reach runs faint (9.9% of denials appealed): per item the channel corrects most of what reaches it, and almost nothing reaches it. Averaging that pair into one middling number is exactly what this network exists to refuse.

• baseline

  The self-fulfilling store. The record-to-record pathway from the case record into the similar-patient cohort is this network's signature loop: a determination shortens or lengthens an actual stay, the stay enters the vendor's comparison cohort, and the next estimate is measured against it - so a determination writes itself back into the reference for the next determination. It runs at a substantial level rather than full because the comparison cohort is compiled across providers and payers rather than from this plan's determinations alone. No bedside-alert sibling in this domain has a reference store its own decisions manufacture.

• baseline

  The committee pair. The internal utilization-management committee carries a reinforcing pathway (its approval sets the tool and the review time budget the desk works inside, at a substantial level) that is stronger than its check (drawn faint). That ordering is not an accusation, it is the record: the body is real, meets, minutes its reasoning and can withhold approval - and its minutes show that, presented with faster handle times together with an increase in adverse determination rate, it voted to tentatively approve the model at the following meeting. The Senate report's third recommendation - that the regulator expand utilization-management-committee rules to prevent this class of tool from unduly influencing human reviewers - is aimed at exactly this surface, and it is why the committee's check is present, faint, rather than absent.

• baseline

  The enforcement node and its check. Determination notice issuance is drawn as an enforcement node because the record documents a downstream action that follows the record as a matter of course: 3.4 million denials issued across the program in 2022. The re-assessment check back onto the record is present, faint, not empty, because the standing rule (medical necessity determinations on the enrollee's medical history, physician recommendations and clinical notes, with medical-director involvement where appropriate) existed throughout, and the February 2024 guidance restates the sequence - an algorithm may assist, and a stay estimate alone cannot be the basis to terminate post-acute care. It is not higher because that guidance was written for this exact operation after the period in which the plan's post-acute denial rate roughly doubled while its review workflow was being made faster.

• baseline

  What this network refuses to carry, and the refusals are load-bearing. (1) The figure most often attached to this deployment - that over 90 percent of claim denials are reversed on appeal - is a pleaded allegation recited by the Lokken court under the motion-to-dismiss standard, followed in the same paragraph by the court's note that the plan denies any use of the tool; it appears nowhere as a value here, and the overturn figure used instead is the KFF analysis of the plans' own federal reporting. (2) The parallel payer's use of the same tool is established only at the pleading level in Barrows and contributes to no rung. (3) The December 2022 workgroup that explored using appeal data to identify cases likely to be appealed is a documented workgroup with two meetings, not a deployed system, and no third model is drawn. (4) The 2023 vendor target of keeping rehab stays within one percent of projected days is carried as a documented internal target, not as a measured behaviour of any coordinator.

• baseline

  What error means here, and what orders it. On the estimator channel an error is an output that mis-states what an individual enrollee's circumstances require, measured against the coverage criteria that legally govern the determination - never an accuracy figure about a person. The value is ordered, not set, by the one ground-truth audit in the record: the 2022 government evaluation found 13 percent of sampled 2019 prior authorization denials (fifteen large organizations pooled, one June week) met Medicare coverage rules, with post-acute facility stays among the report's own examples. That audit predates this vendor's management of the benefit and pools fifteen organizations, so it orders the error term for the decision class and measures nothing about this deployment.

• baseline

  Evidence status. The deployment-specific facts on this diagram come from documents the deploying institutions produced under compulsion - more than 280,000 subpoenaed pages behind the Senate report - plus a government audit of the decision class, a regulator's own guidance, court recitals, and one investigative series read only to its free preview. No independent technical evaluation of this deployment exists in the public record, which is why the model-side check pathway is drawn empty on that documented absence. Naming currency: the vendor brand was retired in Q1 2024 into a successor organization, so this network names the tool and vendor as the subject of the 2020-2023 record, not as a current operating brand.

• assumed

  Served enrollees are not in the dynamics. No coverage decision, length of stay, medical-necessity judgement, health outcome or financial outcome for any person is computed from anything drawn here, and no score over a person is authored anywhere in this network. The denial, appeal and overturn rates on this diagram are properties of a request stream measured by the government and by an analysis of the plans' own federal reporting - recorded, never derived. The record's documented disparities are between service categories and between insurers, not between subpopulations of served people, and this network does not convert a service-line rate into a subpopulation rate.

What this example does not show

• Whether the estimator was used to MAKE coverage determinations is contested, not found: the plan denies any use of the tool in litigation, the vendor publicly described it as a guide, and both federal class actions - survived in narrowed form, contract and implied-covenant counts in both plus unjust enrichment and common-law fraud in the Humana action - are in discovery with no dispositive merits ruling. What this network draws as structure is the uncontested record: a vendor presentation showing coordinators completing the tool during the acute stay, and the plan-side machine-assisted approvals with their measured time and adverse-determination effects.
• The figure most often attached to this deployment - over 90 percent of denials reversed on appeal - is a pleaded allegation recited under the motion-to-dismiss standard and contradicted in the same paragraph of the same order; it appears nowhere here. The overturn figure used is KFF's 83.2 percent, which is program-wide across Medicare Advantage in 2022, not plan- or service-line-specific, and is conditioned on the self-selected 9.9 percent of denials that were appealed.
• The 13 and 18 percent figures are a government audit of one June 2019 week of denials across fifteen pooled organizations, predating this vendor's management of the benefit. They order an error term for the decision class; they measure nothing about this deployment, and no published deployment error rate or independent evaluation of this deployment exists.
• The record's disparities are between SERVICE CATEGORIES and between INSURERS - post-acute denial rates against overall rates - not between subpopulations of served people. No subpopulation-disaggregated denial, appeal or overturn measurement exists for this plan or benefit, and this network does not manufacture one.
• Served enrollees are not modeled. Coverage decisions, lengths of stay, health and financial outcomes are boundary quantities recorded in the case file; the Lab models institutional propagation through the operator network and computes no enrollee outcome from anything on this diagram. The vendor brand named in the record was retired in Q1 2024; this scenario names it as the subject of the 2020-2023 documentary record, not as a current operating brand.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• 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.

  empirical
  • Government U.S. Senate Permanent Subcommittee on Investigations (Committee on Homeland Security and Governmental Affairs), Majority Staff Report, Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care, 17 October 2024 https://www.hsgac.senate.gov/wp-content/uploads/2024.10.17-PSI-Majority-Staff-Report-on-Medicare-Advantage.pdf
  • Investigative Casey Ross and Bob Herman, UnitedHealth pushed employees to follow an algorithm to cut off Medicare patients' rehab care, STAT, 14 November 2023 (part of the Denied by AI series; free preview read, full text paywalled) https://www.statnews.com/2023/11/14/unitedhealth-algorithm-medicare-advantage-investigation/

• Across Medicare Advantage in 2022, of 46.2 million prior authorization determinations, 3.4 million (7.4 percent) were denied in whole or in part; 9.9 percent of denials were appealed; and 83.2 percent of appeals resulted in the initial decision being overturned (KFF analysis of the plans' own federal reporting). The figures are program-wide, not plan- or service-line-specific, and the overturn rate is conditioned on the self-selected minority of denials that were appealed, so it overstates what the same review would correct if applied to every denial.

  empirical
  • Reference KFF, Medicare Advantage Plans Denied a Larger Share of Prior Authorization Requests in 2022 Than in Prior Years (8 August 2024), analysis of federal Medicare Advantage prior authorization and appeals reporting https://www.kff.org/medicare/medicare-advantage-plans-denied-a-larger-share-of-prior-authorization-requests-in-2022-than-in-prior-years/

• Whether nH Predict was used to make coverage determinations is contested and unadjudicated. The vendor's public statement in STAT's March 2023 series opener was that the tool 'is not used to make coverage determinations' and 'is used as a guide'; the Lokken order (D. Minn., Feb 13, 2025) recites - as pleaded allegations taken as true on a motion to dismiss - 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'; the Barrows order (W.D. Ky., Aug 14, 2025) recites Humana's use of the tool as pleadings only. Both class actions survived motions to dismiss in narrowed form - contract and implied-covenant counts in both, plus unjust enrichment and common-law fraud in the Humana action and are in active discovery as of August 2026, with no dispositive merits ruling. CMS's contract-year-2024 rule and its February 6, 2024 FAQ constrain the order of operations: an algorithm may assist a coverage determination, the plan remains responsible, an algorithm determining coverage from a larger data set instead of the individual patient's medical history, physician recommendations, or clinical notes would not comply with 42 CFR 422.101(c), and a predicted length of stay alone cannot be the basis to terminate post-acute care services - only re-assessing the individual patient's condition can. The naviHealth brand was retired in Q1 2024 into Optum 'Home & Community Care'; the pipeline continues under that name.

  empirical
  • Government Estate of Gene B. Lokken v. UnitedHealth Group, Inc., UnitedHealthcare, Inc., and naviHealth, Inc., Civil No. 23-3514 (JRT/DJF), U.S. District Court for the District of Minnesota, Memorandum Opinion and Order Granting in Part and Denying in Part Defendants' Motion to Dismiss, Doc. 91, 13 February 2025 (Tunheim, J.) https://storage.courtlistener.com/recap/gov.uscourts.mnd.211721/gov.uscourts.mnd.211721.91.0.pdf
  • Government Centers for Medicare & Medicaid Services, Health Plan Management System memo to all Medicare Advantage Organizations and Medicare-Medicaid Plans, Frequently Asked Questions related to Coverage Criteria and Utilization Management Requirements in CMS Final Rule (CMS-4201-F), 6 February 2024 https://www.aha.org/system/files/media/file/2024/02/faqs-related-to-coverage-criteria-and-utilization-management-requirements-in-cms-final-rule-cms-4201-f.pdf
  • Government 42 CFR 422.101(c), Medical necessity determinations and special coverage provisions (Centers for Medicare & Medicaid Services, as amended by the contract-year-2024 Medicare Advantage final rule CMS-4201-F, applicable to coverage beginning 1 January 2024) https://www.ecfr.gov/current/title-42/section-422.101
  • Investigative CBS News, UnitedHealth uses faulty AI to deny elderly patients medically necessary coverage, lawsuit claims https://www.cbsnews.com/news/unitedhealth-lawsuit-ai-deny-claims-medicare-advantage-health-insurance-denials/
  • Investigative Casey Ross and Bob Herman, Denied by AI: How Medicare Advantage plans use algorithms to cut off care for seniors in need, STAT, 13 March 2023 (series opener; free preview read, full text paywalled) https://www.statnews.com/2023/03/13/medicare-advantage-plans-denial-artificial-intelligence/
  • Government Barrows, et al. v. Humana, Inc., No. 3:23-cv-654-RGJ, U.S. District Court for the Western District of Kentucky, Memorandum Opinion and Order Granting in Part and Denying in Part Defendant's Motion to Dismiss, Doc. 82, 14 August 2025 (Jennings, J.) https://litigationtracker.law.georgetown.edu/wp-content/uploads/2023/12/Barrows-et-al_2025.08.15_MEMORANDUM-OPINION-ORDER.pdf
  • Investigative Bob Herman and Casey Ross, UnitedHealth discontinues a controversial brand amid scrutiny of algorithmic care denials, STAT, 23 October 2023 https://www.statnews.com/2023/10/23/unitedhealth-optum-navihealth-rebranding-algorithm/
  • Reference Georgetown Law, O'Neill Institute Health Care Litigation Tracker, Estate of Gene B. Lokken et al. v. UnitedHealth Group Inc. et al. (tracker entry, procedural posture read August 2026) https://litigationtracker.law.georgetown.edu/litigation/estate-of-gene-b-lokken-the-et-al-v-unitedhealth-group-inc-et-al/
  • Reference Georgetown Law, O'Neill Institute Health Care Litigation Tracker, Barrows et al. v. Humana Inc. (tracker entry, procedural posture read August 2026) https://litigationtracker.law.georgetown.edu/litigation/barrows-et-al-v-humana-inc/

• The closest ground-truth audit of this decision class is HHS OIG evaluation OEI-09-18-00260 (April 2022): reviewing a stratified random sample of 250 prior authorization denials and 250 payment denials issued by 15 of the largest Medicare Advantage organizations during one week of June 2019, health-care coding experts and physician reviewers found 13 percent of the prior authorization denials and 18 percent of the payment denials met Medicare coverage rules, with identified causes including internal clinical criteria applied beyond Medicare rules, insufficient-documentation findings the reviewers judged unfounded, manual processing errors, and system programming failures, and with stays in post-acute facilities among the report's own examples. The evaluation predates naviHealth's management of this benefit and pools fifteen organizations, so it orders - and does not measure - an error term for this deployment.

  empirical
  • Government evaluation U.S. Department of Health and Human Services, Office of Inspector General (2022, April 27). Some Medicare Advantage Organization Denials of Prior Authorization Requests Raise Concerns About Beneficiary Access to Medically Necessary Care (OEI-09-18-00260) https://oig.hhs.gov/oei/reports/OEI-09-18-00260.asp

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.

All of them in context on the Clinical decision support & deterioration alerting domain page.

Levers available here and the patterns behind them

• Upgrade model — Improve the model
• Keep prompts neutral — Framing and mirroring reduction
• Mark AI-written records — Provenance labeling
• Gate record entries — Human-in-the-loop write gating
• Pause AI on alarms — Deployment circuit-breaker
• Escalate checks — State-feedback vigilance
• Assign a challenger — Structured dissent
• Understand the system — Understand the system
• Check copied records — Reconcile copied records
• Review on schedule — Oversight cadence & retrospectives
• Check with a second model — Cross-model verification
• Gate vendor updates — Vendor quality gate
• Store less data — Data minimization

Documented case histories

• nH Predict Utilization Review
• TREWS sepsis early-warning system
• Advance Alert Monitor (AAM) deterioration model
• Sepsis Watch deep-learning detection system
• Proprietary EHR sepsis model (external validation)
• Cost-Proxy Care Stratification
• CA-CDS Child Abuse Alerting
• IDx-DR Autonomous Screening
• Viz.ai LVO Stroke Triage
• IBM Watson for Oncology
• OPTN eGFR Waiting-Time Correction
• Practice Fusion Pain CDS
• UBH Level of Care Guidelines (Wit v. UBH)
• EviCore by Evernorth: the review threshold
• Cigna PxDx