PRISM · Predictive Recommendations for Improved Screening in Medicine

The whole patient journey is already written down.
PRISM reads it.

For a whole class of conditions, an inexpensive diagnostic test exists — and most people who would benefit never get it until years of symptoms have accumulated. PRISM turns the anonymized claims history an insurer already holds into a flag to the patient's physician: this specific test may be worth ordering now.

A carrier from the 2026 prototype population — patient 21752, pool 5

PERSONWHENWHEREWHOWHATWHY
SIGNAL34 female2025-01-06POS-81: Independent LaboratoryNUCC-291U00000X: Clinical Medical LaboratoryCPT-83541: 24-hour urine mineral-excretion panelICD-E87.A3: Recurrent electrolyte mineral derangement;
34 female2025-01-06POS-11: OfficeNUCC-208D00000X: General PracticeCPT-99213: Office or other outpatient visit…ICD-M54.50: Low back pain, unspecified;
34 female2025-01-10POS-11: OfficeNUCC-111N00000X: ChiropractorCPT-98941: Chiropractic manipulative treatment (CMT); spinal, 3-4 regionsICD-M54.50: Low back pain, unspecified;
SIGNAL34 female2025-01-12POS-11: OfficeNUCC-207V00000X: Obstetrics & GynecologyCPT-83543: Gynecologic mineral-axis evaluationICD-N94.A1: Female pelvic mineral-axis symptom;
SIGNAL34 female2025-02-15POS-01: PharmacyNUCC-333600000X: PharmacyATC-C03AZ08-211584: hydralozide 25 MG Oral Tablet [#30 DS:30]ICD-R03.A1: Resistant blood pressure elevation on multiple agents;
SIGNAL34 female2025-05-30POS-11: OfficeNUCC-207RC0000X: Internal Medicine, Cardiovascular DiseaseCPT-93785: Extended ambulatory blood-pressure monitoringICD-R03.A1: Resistant blood pressure elevation on multiple agents;
34 female2025-06-03POS-01: PharmacyNUCC-333600000X: PharmacyATC-M01AE02-206165: Naproxen 500 MG Oral Tablet [#30 DS:30]ICD-M54.50: Low back pain, unspecified;
▸ the ensemble continues the table
TEST34 female2025-06-06POS-81: Independent LaboratoryNUCC-291U00000X: Clinical Medical LaboratoryCPT-82197: Serum veladrin / regulatory-mineral ratio, quantitativeICD-E27.A4: Suspected mineralocorticoid-axis dysregulation;

Six columns are all any PRISM model ever sees — no names, no notes, no identifiers. The SIGNAL/TEST marks are annotations for you, the reader; no model ever sees them.

Every model screens every patient — independently

model 1 model 2 model 3 model 4 model 5

5 of 5 independent models, each continuing this table on its own, surfaced the screening test. That agreement — not any single model — is the flag a physician sees.

What the ensemble keyed on: the four highlighted rows — a urine panel, a gynecologic evaluation, a pharmacy fill, a blood-pressure study — scattered across five months, four different providers, each unremarkable on its own. No one along the way sees them side by side. The claims record does.

This is synthetic demonstration data: the condition, its codes, and this patient are fabricated by design, so that the 2026 proof of concept could isolate exactly what the models learned. In the full run, all 250 carriers with this timely-tested arc drew a unanimous 5-of-5 flag.

The idea in three moves

concept

No diagnosis engine. No risk score.
A table, continued.

1 · Reduce

Every patient becomes a six-column table of standardized billing codes — PERSON, WHEN, WHERE, WHO, WHAT, WHY — one row per medical event. Nothing else exists in the pipeline: no names, no addresses, no free text.

2 · Continue

A model's entire job is to continue the table with the next plausible rows. It is deliberately over-specialized — trained until this is all it can do. A suggestion "fires" when its continuation contains a specific diagnostic test.

3 · Agree

The training population is split into disjoint pools; one independent model is trained per pool; every patient is screened by all of them. Because the models share nothing, agreement is evidence. The output is a consensus count — a flag, never a probability.

The 2026 proof of concept

demonstrated

Tested the way you'd want it tested:
on a condition that cannot exist.

A fictional condition was injected into randomly chosen carriers among 25,000 synthetic patients, its codes present in no real codeset — so any firing could only be attributable to the injected pattern. Five models were trained; every carrier was screened by all five.

Full grid
16,250
runs — every example × every model, zero errors
Caught in time
100%
firing on timely-tested carriers, by models that never saw them
Surfaced earlier
90%
firing on prompts cut ~20 rows earlier — ending before the test ever appeared
False flags
0.07%
on no-signal controls when all 5 models must agree

Honest limits

These rates were measured on deliberately clean synthetic data — an upper bound that shows the method works, not a forecast of real-world performance. Whether real conditions leave analogous precursor patterns in claims data is the open question the next phase exists to test. Read the results and their limits →

Non-negotiables

concept

Built so it cannot become the thing you're worried about.

Constructive-only

The system counts positive suggestions and nothing else. No pathway exists by which it could emit "don't test," rank patients for denial, or feed a coverage decision.

A recall instrument, never a predictor

Models are forced to continue every timeline — never allowed to "decide" by stopping. A fire means "consider this test." Silence is never stored, reported, or usable as "no need."

Anonymous by architecture

The six columns contain standardized codes, an age, and a sex. Names, addresses, and free text don't reach the pipeline at all — identifying information isn't removed from the data; it was never in it.

Paid only on outcomes

Revenue exists only when a suggested test is performed and followed by documented early treatment. PRISM earns by catching what was missed — never by gatekeeping. And the company is a Public Benefit Corporation: patient outcomes are a charter-level obligation.

All safeguards, in depth →

How this site talks

Every claim wears its status.

PRISM mixes locked experimental results with production intent, and the two must never blur. Everything on this site is tagged with how much weight to put on it — the same discipline used in the project's own documentation.

concept demonstrated vision history

concept — a durable design principle · demonstrated — validated end-to-end in the 2026 synthetic prototype · vision — production intent, not yet built · history — past work, kept for provenance.

Read the full documentation →