PRISM

Appearance

Utilization Phenotypes ​

Status: concept — the central claim about what there is to learn; the synthetic prototype demonstrated the machinery on one manufactured instance of one.

A utilization phenotype is the pattern a developing condition leaves in how a patient uses the healthcare system — the signature written in visits, referrals, prescriptions, and escalations rather than in lab values or imaging. It is the thing PRISM actually learns to recognize.

A second kind of phenotype ​

Medicine describes conditions through clinical phenotypes: physical findings, laboratory abnormalities, imaging results. But a condition that is making someone ill also changes their behavior in the healthcare system, and the system's behavior toward them. A medication stops working and gets escalated. A vague symptom earns an office visit, then another. A specialist is consulted for one piece of the picture, a different specialist for another. An emergency room absorbs a crisis at 2 a.m. that no scheduled appointment caught.

Each of those events produces a billing record, because everything that expects payment produces a billing record. The sequence of those records — what was done, where, by what kind of provider, justified by what diagnosis, in what order and at what tempo — is the utilization phenotype. It exists whether or not any individual clinician ever recognizes it, because it is generated automatically by the ordinary operation of the claims system, uniformly across every provider and care setting.

Two properties make it unusually good material for pattern learning. It is objective: no interpretation, no documentation-quality variance, no missing notes — the record is what was billed. And it is complete across providers: the pattern that no single office can see, because each provider holds only a slice, is intact in the one place the slices reunite.

Earlier than the diagnosis ​

Utilization phenotypes typically begin before anyone knows what they are looking at. The healthcare system responds to symptoms long before it names causes: treatments are tried, tests are ordered, referrals are made — all of it billed — while the underlying condition is still unrecognized. By the time a diagnosis code finally appears in a record, the phenotype may have been legible for months or years.

That precedence is the entire opportunity. A system that can recognize the phenotype does not have to wait for the diagnosis, because the phenotype is the part that comes first. This is what "early" means throughout these docs: not predicting the future, but recognizing a pattern that is already present in the record before its explanation is.

What PRISM does with it ​

PRISM never defines phenotypes explicitly — no rules, no feature lists, no clinician-authored criteria. Its models are trained to do exactly one thing: continue a patient's timeline with the next plausible rows. A condition's utilization phenotype is, from the model's perspective, nothing more than a regularity in how certain timelines unfold — and a model trained on enough timelines continues a phenotype-bearing history differently than a background one. When the natural continuation of a pattern is a diagnostic test, the model emits the test; that is the whole mechanism of a screening suggestion.

This is also why nobody has to know what a given condition's phenotype is. The three-pattern construction selects training examples by outcome anchors (a TEST followed by early or late treatment), not by naming the precursor pattern; whatever regularity precedes the anchor is what the models absorb. In production, the precursor patterns can stay unnamed forever — although consistently learned ones would themselves be research material, hypotheses about disease progression surfaced from population-scale data.

The manufactured instance ​

The 2026 synthetic prototype built one utilization phenotype by hand — Primary Veladrin Excess, a fictional condition with a designed lead-up of workups, labs, scattered specialist visits, and vague symptom encounters — precisely so the ground truth would be knowable. Because the phenotype was injected under strict isolation, the prototype could prove the machinery recognizes exactly the pattern and nothing else: the ensemble fired when precursor rows were present, stayed silent when they were stripped, and needed as little as one recognizable event to fire.

The honest boundary of the claim: the prototype proves the machinery works given a learnable phenotype. Whether any particular real condition leaves one — rich enough, consistent enough, present in claims at all — is an empirical question the method cannot answer in advance. That question is exactly what the condition-selection criteria and the real-data phase exist to test, condition by condition.

See also ​