Domain Atlas / Clinical decision support & deterioration alerting
Cigna PxDx
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 1 assumed · 9 published baseline.
ProPublica and The Capitol Forum (Patrick Rucker, Maya Miller, David Armstrong), computing from internal Cigna records and interviews with former employees, reported on March 25, 2023 that Cigna's PxDx review flags claims where the billed procedure code does not pair with the diagnosis code on a payer-authored list, and that company medical directors then sign the flagged denials in batches without opening patient files: over 300,000 payment requests denied through this method across a two-month period in 2022, at an average of 1.2 seconds of physician attention per case, with individual medical directors signing between roughly 60,000 and 121,000 denials in one to two months. A former Cigna doctor told the reporters, 'We literally click and submit. It takes all of 10 seconds to do 50 at a time.' The same records, per the investigation, show the match list being extended on cost grounds: adding autonomic-nervous-system testing in 2014 carried an internal projection of roughly 2.4 million dollars a year in savings, and the executive credited with developing the process said it had 'undoubtedly saved billions of dollars.' Cigna publicly disputes the article's characterization of the process - a spokesperson called a complaint built on it 'based on an article riddled with factual errors and misinformation' - and has published no substitute figures; the underlying documents are now in discovery. The investigation received the April 2023 Sidney Award. These figures are the investigation's computation and are not adjudicated fact.[3]
What happened
Every other deployment in this domain decides something about care. This one decides about money, after the fact. PxDx — Cigna's own name for it, short for "procedure to diagnosis" — is post-service claim review. A physician treats a patient, bills for the treatment, and the claim arrives at a screen that compares the procedure code billed against a payer-authored list of diagnosis codes deemed acceptable for that procedure. Where they pair, the claim is approved and paid; Cigna's published account is that 94 percent of the claims subject to this review are approved automatically, that the list covers roughly 50 common, relatively low-cost tests and procedures, and that denials issued through it are less than 1 percent of total claim volume. Where they do not pair, the claim goes into a queue. Both sides of the dispute agree on what the screen is not: Cigna calls it "simple sorting technology that has been used for more than a decade" and states it "does not involve algorithms, artificial intelligence, or machine learning," and the investigative account describes a list lookup. Nothing here is a learned model. That is the point of the case rather than a caveat on it, because the contested object is not a model's accuracy — it is the layer above the match, and the list beneath it.
On March 25, 2023, ProPublica and The Capitol Forum (Patrick Rucker, Maya Miller, David Armstrong) published what internal company records and former employees said that layer consisted of. Medical directors, they reported, signed denials in batches from the queue of flagged mismatches without opening patient files; a former Cigna doctor described the work as "We literally click and submit. It takes all of 10 seconds to do 50 at a time." Computing from company records, the reporters put the volume at over 300,000 payment requests denied through this method across a two-month period in 2022, at an average of 1.2 seconds of physician attention per case, with individual directors signing between roughly 60,000 and 121,000 denials in one to two months. The same records, per the investigation, show how the list grew: adding autonomic-nervous-system testing in 2014 carried an internal projection of roughly $2.4 million a year in savings, and the executive credited with developing the process, Dr. Alan Muney, told the reporters it had "undoubtedly saved billions of dollars." The investigation also reported the design assumption underneath the correction channel — that Cigna internally estimated only about 5 percent of people would appeal a denial — on a set of claims selected for being low-dollar. The emblematic arc is one physician's own: a roughly $350 vitamin-D blood test denied in autumn 2021 as not medically necessary, an internal appeal decided by a different Cigna doctor and lost, and a reversal at external independent review roughly seven months later. Cigna disputed the reporting publicly and has continued to; a spokesperson later called a class-action complaint built on it "based on an article riddled with factual errors and misinformation." It has published no substitute figures. The work won the April 2023 Sidney Award.
Four outside channels then acted, and only one of them has closed. Congress wrote first: on May 16, 2023 the House Energy and Commerce Committee's Republican chairs demanded PXDX process documents, legality memoranda, the list of plans subject to the review, per-medical-director denial records, and 2022 counts of claims reviewed, denied, appealed and overturned, by May 30 — and the same day the reporters described scrutiny from the U.S. Department of Labor and from the California, Washington and Delaware regulators. No public committee findings, hearing record or released production has been located since; the inquiry is a demand on the record, not an outcome. The courts wrote second and narrower than the headline suggests. Kisting-Leung v. Cigna Corp. was filed in the Eastern District of California on July 24, 2023; on March 31, 2025 Judge Dale A. Drozd dismissed the Employee Retirement Income Security Act (ERISA) denial-of-benefits claim with leave to amend and plaintiffs elected not to replead it, so the surviving claims are the ERISA fiduciary-duty claim and a California unfair-competition claim resting on Health & Safety Code § 1367.01(e), which permits a medical-necessity denial only from a licensed physician or a licensed health care professional competent to evaluate the specific clinical issues. Three of the six original plaintiffs, including the named lead, were dismissed for lack of standing after Cigna's own declaration evidence showed there were no PxDx denial letters on their claims — which establishes, from the defense side, that not every denial of this insurer's runs through this review. Assuming the most deferential standard for the sake of argument, the court held at the pleading stage that reading the plan's medical-director requirement "as allowing an algorithm to make the decision so long as a medical director pushes the button" would conflict with the plan's plain language. That is an interpretation of allegations assumed true, not a finding about what happened. Cigna answered in May 2025; one further plaintiff was voluntarily dismissed in August 2026, leaving two; roughly 2.1 million pages have been produced; depositions are coordinated with the parallel Connecticut action so Cigna's witnesses sit once for both; fact discovery closes in autumn 2026 and class-certification briefing opens October 29, 2026. No class exists yet.
The regulator wrote the only closed finding. On October 8, 2025 the California Department of Managed Health Care fined Cigna HealthCare of California, Inc. $500,000 for improperly denying providers' claims as not medically necessary, finding that the plan "reviewed and denied claims without physicians conducting clinical reviews of the claims prior to issuing denials" and that it had used a different review process than the policy it filed with the Department. Cigna agreed to pay, to re-review denials issued under the non-compliant process going back two years, and to revise and refile its policy. Those are findings agreed to by the regulated entity and they may be stated as fact — with one boundary that matters: the Department's release names neither PxDx nor the investigation. The conduct found matches the documented pattern and the action is described here as addressing the plan's claims-review practice, consistent with that pattern, never as a PxDx fine; and the respondent is the California-regulated plan entity, not the national group. No report of the re-review's completion or its results has been located.
The sociotechnical reading
The governed object here is a list and a signature, and the machine between them is trivial by design. That inversion is worth dwelling on, because the domain's other deployments invite a question this one refuses: how accurate is the model? A deterministic code match has no accuracy in that sense. Its error is a flagged mismatch that a clinical review of the individual claim would not have denied under the governing coverage terms — which makes the error entirely a property of who wrote the list and of what a flag is permitted to become downstream. The map draws exactly those two surfaces. The match list is the widest inbound pathway on the board because it IS the decision boundary, and nothing in the record audits its entries against coverage terms. What the signing physician reads from the claim record at the moment of determination is drawn at the floor. Those two widths in the same diagram are the investigation's finding and the regulator's finding rendered as structure rather than asserted in prose: the boundary is wide and the read at the point of decision is not.
The record side carries the case's strangest loop. Additions to the match list were justified, per internal documents the investigation quotes, by projected savings from denying the added category — and those savings are computed from the denials the list produces. The boundary grows on its own output. No sibling in this domain has that shape, and it is drawn as a pathway from the determination record back into the list, carrying the privacy flag it earns: what a claim was submitted for was payment, and what determination data does here is fund the next extension. A second store makes the oversight failure legible without narrating it. The review policy the plan filed with the state regulator is the only artefact through which an outside body could read this pipeline, and the Department's 2025 finding is that the process in use was not the process on file. So the diagram draws an oversight channel reading a different record than the one running — which is not a metaphor, it is what the finding says.
The correction channels are real, narrow, and honest about which is which. A second company doctor reads an appealed claim with the file open; that is more than the original determination is documented to have had, and in the one arc the record follows end to end it upheld the denial. An independent reviewer outside the plan reversed the same claim about seven months later. The board therefore carries the outside correction's per-item strength and the width of the pathway feeding it as two separate numbers, never averaged: a channel can reach the right answer on nearly everything that gets to it while almost nothing gets to it. The thing that gates the pathway is not a failure of the channel — it is a design expectation, reported from company records, that about 5 percent of people would appeal, on claims chosen for being small. The rarest object in the whole payer cluster sits at the end: a retroactive undo. The regulator's settlement reaches determinations nobody appealed, ordering two years of denials re-read against the claims they were made on. It is drawn present but low — one state's regulated entity, agreed once, no standing reconciliation, and no published result.
What this file will not do is worth stating as plainly as what it does. The quantitative core is one investigative computation from internal records, disputed by the operator without counter-figures, and now being tested in discovery. Every throughput figure is attributed wherever it appears. The operator's counter-frame is carried rather than rebutted, because two parts of it are structural facts the map is built on: no care is gated by anything here, and the screen is a list lookup rather than a learned system. The 5 percent figure is an expectation, not a measured appeal rate, and it is drawn as pathway width, never as a rate. The Medicare Advantage appeal and overturn statistics that appear in the congressional correspondence measure a different programme and were used there as an analogy; no value here rests on them. Served patients and providers are not in the dynamics: no coverage decision, clinical outcome or financial outcome for any person is computed from anything on the diagram, and the record contains no denial, appeal or overturn measurement for this review disaggregated by any characteristic of a served person, so none is manufactured. The deprivation the record documents is payment for care already delivered, and the burden of appealing it. Both of those live here, in the case file, and not on the board.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.