PRISM

Appearance

What is PRISM? ​

Status: concept, demonstrated — the method was validated end-to-end in the 2026 synthetic prototype.

PRISM (Predictive Recommendations for Improved Screening in Medicine) turns the insurance-claims histories a payer already holds into early-screening suggestions — flags to a physician that a specific, established diagnostic test may be worth ordering now. This page is the whole idea in one pass; every claim links to the article that owns it.

The problem ​

For a whole class of conditions, an established, non-invasive, inexpensive diagnostic test exists — and most of the people who would benefit never receive it until years of escalating symptoms, ineffective treatments, and complications have accumulated. The knowledge exists, the test exists, the treatment exists; what is missing is a systematic way of noticing which patient should get which test, and when. The screening gap describes why this failure is persistent: each provider sees one slice of a patient's journey, and nobody along the way is responsible for looking at the whole. One party does hold the whole record. Every visit, prescription, lab, and admission produces a claim, so the insurer's vantage point captures the complete journey across providers, years, and care settings — a record assembled for billing that happens to be exactly the longitudinal view early screening needs.

The mechanism ​

PRISM reduces each patient's history to a six-column markdown table of standardized billing codes — PERSON, WHEN, WHERE, WHO, WHAT, WHY — one row per medical event. Those six columns are the only thing any PRISM model ever sees, in training or inference; no name, address, or free text exists anywhere in the pipeline, a property with real regulatory consequences (anonymous by architecture).

A model's entire job is to continue the table with the next plausible rows. The models are deliberately over-specialized — trained until continuing this table is all they can do. A screening suggestion "fires" when a model's forced continuation contains the diagnostic TEST code: asked what plausibly comes next for this patient, the model answers "this test." The pattern a developing condition leaves in billing data — its utilization phenotype — is what makes that answer learnable, and three-pattern learning is how timelines are turned into training material that teaches it.

No single model is ever trusted. The training population is split into disjoint pools, one independent model is trained per pool, and every patient is screened by all of them. Because the models share no data, weights, or adapters, agreement between them is evidence. The output is a consensus count — "N of 5 surfaced this test" — a flag for a physician, never a probability and never a diagnosis.

What has been demonstrated ​

The method was validated end-to-end in the 2026 synthetic proof of concept. A fictional condition — Primary Veladrin Excess, a fabricated phenotype whose codes exist in no real codeset — was injected into randomly chosen carriers among 25,000 synthetic patients across five pools, so that any firing could only be attributable to the injected signal. Five models were trained and every phenotype example was run against all five: 16,250 runs, zero errors. For the four models that had never seen a given patient, firing rates were 100.0% on carriers whose signal led to timely testing, 99.8% on carriers tested too late, and 90.0% on prompts cut ~20 rows earlier — ending before the test ever appeared, the operational meaning of "earlier." Control carriers with no lead-up pattern drew fires only 1.5% of the time per model, collapsing to 0.07% under unanimous 5-of-5 consensus; stripping the injected rows collapsed firing to near zero. The full grid, ablations, and every explained miss are in results — along with what the prototype deliberately does not prove: the data was synthetic and clean, so these rates are an upper bound, not a forecast, and whether real conditions leave analogous precursor patterns remains the open question.

Safeguards ​

Two properties are architectural, not policy. PRISM is constructive-only: the system counts positive suggestions and nothing else, and no pathway exists by which it could emit "don't test" or feed a denial of care. And PRISM is a recall instrument, never a predictor: models are forced to continue every timeline rather than being allowed to stop, silence is never reported as "no need," and suggestions reach only the patient's primary-care physician as ordinary clinical decision support the physician is free to ignore.

Business model ​

PRISM is paid only when a suggestion works. Results-based compensation means revenue exists only when a suggested test is performed, tracked, and followed by documented early treatment — with the patient's cost of the test covered, and no fee-for-service or per-member fees. The system runs on claims data the insurer already has, on PRISM-managed hardware inside the insurer's walls, with zero integration into provider workflows. The company is a Public Benefit Corporation: patient outcomes are a charter-level obligation that survives any change of ownership.

What comes next ​

The prototype proved the method works given a learnable precursor pattern; the next phase tests whether real conditions leave such patterns in real claims data. That phase ingests real claims, runs the same two training rounds per pool, and evaluates retrospectively — every historical patient whose late diagnosis is already on record becomes a test of whether the ensemble would have flagged them sooner. The production vision — on the order of 100 pools, continuously retrained, screening a whole population in the background — is design intent, marked as such throughout this doc set.

See also ​