Product Field Work All builds
BuildLegible

Overturned

When a California HMO denies care and the patient appeals to the state, an independent doctor decides, and the written reasoning is published. A model read all 42,746 decisions since 2001 in about eight minutes for $2.55. Insurers now lose most of these appeals, and the cases show what separates the patients who win from the ones who don’t.

The problem

Insurers deny tens of millions of claims a year, and almost nobody pushes back. People don’t know they can appeal, don’t know whether they would win, and don’t know what to send. The best evidence about what wins sits in public, in the written findings of independent reviewers, as free text no patient, doctor or regulator has time to read.

85Min-network claims denied by HealthCare.gov insurers in 2024
<1%of those denials were appealed by consumers
34%of Marketplace enrollees knew they had a right to an external appeal

Source: KFF, Claims Denials and Appeals in ACA Marketplace Plans in 2024. Marketplace data only; employer plans don’t report.

The cost is care people give up on, bills they pay themselves, and denials that keep coming because challenging them is rare. A denial nobody appeals is a denial that works.

How I measure the problem
  1. How often independent reviewers side with the patient, by year and by what was disputed, from the state’s own outcome field
  2. What the winning cases contain that the losing ones don’t, read from every reviewer’s findings
  3. Whether the model’s reading holds up, checked against the state’s own case-type field and a keyword test
Why I chose it

It’s the flip side of rising healthcare costs that ordinary people can actually act on. The data is public, de-identified and already judged, so every insight comes with an answer key. And the finding is practical: it can tell a real person whether to fight and what to bring.

The ideaAsk a small decision model six typed questions about every reviewer’s findings, join its answers to the state’s outcome, and turn the patterns into odds and a checklist a patient can use.

Sources: California Department of Managed Health Care, Independent Medical Review Determinations (42,746 decisions, 2001 to 2026); TypeSafe jev-1.13.0 answers, run September 27, 2026.

Every appeal since 2001

Each dot is one decision. Green means the reviewer overturned the insurer and the patient got the care. Regroup the dots to see which kinds of disputes patients win. Tap any dot to read the case.

Patient wonInsurer upheld

Tap any dot to read that case.

Should I appeal?

Describe your denial. You’ll see how similar appeals turned out, what decided them, a checklist built from the winning cases, and how to file.

What the insurer said
Have you already tried other treatments that failed, caused problems, or aren’t safe for you?
–

0%100%

What to bring

    What decided cases like yours

    When the patient won

    When the insurer was upheld

    Real decisions like yours

    How to file

    1. Appeal to your plan first, in writing.Ask for the denial reason and the criteria they used. In California you must go through the plan’s grievance process, which it has 30 days to resolve, before the state will review.
    2. Build the file from the checklist.A letter from your doctor naming the guideline, and records of every treatment you’ve tried and what happened.
    3. Ask for an independent review if the plan says no.In California, apply to the DMHC Help Center within six months of the plan’s decision. It’s free, and urgent cases can be expedited. Call 1-888-466-2219 or apply at dmhc.ca.gov.
    4. Outside California, you likely have the same right.Most private plans must offer an independent external review after an internal appeal. Start at HealthCare.gov’s external review page or your state insurance department.

    These are historical results from California independent medical reviews, not a prediction for your case and not medical or legal advice. They include only people who appealed all the way to the state.

    The yardstick changed, and patients started winning

    The share of appeals reviewers decided for the patient, by year. Pick a dispute type to compare it with all appeals.

    Denied, then reversed almost every time

    Treatments ranked by how often reviewers overturned the insurer. When a denial loses nine times in ten, the denial itself is the pattern worth examining.

    What a complete answer costs

    Say a patient advocacy group, a health system’s denials team or a regulator wants these six answers for every decision. Here is the cost by hand against the cost with the model and a person reviewing what it’s unsure about.

    –saved per full pass
    –analyst hours by hand
    –analyst hours with the model

      In practice nobody reads 42,746 decisions by hand. They count the state’s outcome field, which says who won but never why. The hand cost is what the “why” would take.

      From public data

      Median pay, BLS, 2024.

      Private industry, BLS ECEC.

      Assumptions you can change

      Decisions average about 1,500 characters. No public benchmark exists for coding time; these are starting points, not findings.

      What it means for the people

      Today the people who fight denials work from memory and anecdote, one case at a time. Here is how the work shifts, and who it helps.

      Today
      Reading and coding every decision
      With the model
      Spot checks
      Unsure cases
      Freed for helping people

      Patients and familiesSee their odds before giving up, and walk into the appeal with the evidence that wins cases like theirs instead of a blank form.Uses: the odds, the checklist, real decisions
      Doctors’ officesWrite letters of medical necessity aimed at what reviewers weigh: the named guideline and the treatments already tried. Fewer letters, fewer losses on paperwork.Uses: what decided cases like yours
      Advocates and legal aidTriage a waiting list by likely outcome, and spend scarce hours on the cases that are winnable but currently missing evidence.Uses: the odds by dispute and treatment
      Regulators and employersSpot denials that lose on appeal nine times in ten and ask why they are still being issued. Those are candidates for prior-authorization reform.Uses: the reversal ranking and the trend
      What disappears, and what doesn’t

      Reading thousands of decisions to find a pattern disappears. Deciding to appeal, writing the letter, and the medical judgment itself stay with people. The model reads what reviewers wrote, not the medical record, so it can say what tends to win, never whether a specific patient should get a specific treatment.

      How it works

      One public fileEvery California independent medical review since 2001, with the state’s outcome and the reviewer’s written findings.
      Six typed questions, one requestWhat was disputed, what decided it, prior treatments, guideline cited, paid first, and how serious going without would be.
      Code does the countingEvery rate, grouping and checklist is computed in the page from cached answers, so nothing calls the model again.
      Every number opens a caseTap a dot to read the decision behind it, with the model’s answers beside it.

      Evidence

      –of cases the model called “proven or experimental” were ones the state also filed as Experimental/Investigational; it found of them
      –agreement between the model’s “got the care first” answer and a plain keyword check for reimbursement
      $2.55for the whole archive: 42,746 requests, about 60 million input tokens, about eight minutes of wall time, no failed calls
      What I haven’t proven yet

      This covers California plans regulated by the Department of Managed Health Care, mostly HMOs. It leaves out PPOs regulated by the Department of Insurance, self-funded employer plans and Medicare. It includes only people who appealed all the way to the state, who are not a random sample of everyone denied. Findings are the reviewer’s summary, not the medical record, and they don’t name insurers. Associations here aren’t causes: patients who document failed treatments may differ in other ways. I haven’t built an independent hand-labeled set; accuracy rests on the two field checks above and spot reads. The model’s severity answer reflects how reviewers wrote each case, so I use it only in case cards. Results are pinned to jev-1.13.0.

      Choices

      Read the reasoning, not just the outcomeThe state’s fields say who won. Only the reviewer’s words say why, and the why is what a patient can act on.
      A small decision model over a chat modelTyped answers with probabilities, six questions in one request, input tokens only at $0.042 per million. No paragraphs to parse, and every unsure answer is flagged.
      Odds with a sample size, never a promiseThe tool shows how many cases it drew on and a range, and says when it had to broaden the match because the data was thin.