The eligibility layer

ClinicalTrials.gov publishes eligibility criteria as free text — one prose block per study. That is readable by a person and not by a program. OncoMatch parses that text into typed fields: which biomarker a trial requires or excludes, at what threshold, by which assay, for which cohort; how many prior lines of therapy are allowed; which stages qualify.

The website is one consumer of that layer. This page describes the layer itself.

Coverage

Trial count is queried live: published, recruiting, and mapped to a cancer-type panel — the set the matcher can actually return. The other three figures were measured 2026-09-09 across 7,282 published recruiting trials.

What is structured

Biomarker criteria carry the gene, the alteration and its category (mutation, expression, fusion, amplification), a canonical form of the alteration, the assay where the protocol names one, the cohort the criterion applies to, and a logic group so that alternatives (“EGFR or ALK”) stay distinguishable from conjunctions.

Each criterion is required, excluded, or allowed. Those are not interchangeable: matching scores required and excluded criteria only. allowed is a deliberately neutral bucket — it holds permissive mentions, stratification factors, and per-arm genotypes in umbrella trials, which are not the same thing and should not be scored as if they were.

Prior therapy is structured separately: minimum and maximum prior lines, treatment-naive requirements, and specific agents or classes that disqualify.

Provenance

Most structured facts record the sentence they came from. A claim that a trial requires EGFR T790M is checkable against the protocol text that produced it, rather than taken on trust. Coverage is high and, as below, not complete.

Measured 2026-09-09. Coverage is not complete: where a source sentence is absent the field is shown without one rather than with a reconstructed quote.

Freshness

Recruitment status is re-checked daily against ClinicalTrials.gov, and status changes, enrollment revisions and completion-date movements are appended to a history table rather than overwritten — so the record is a time series, not a snapshot. New trials and materially changed eligibility text are re-extracted weekly.

Accuracy — what is and is not established

Extraction is performed by an ensemble of three language models whose outputs are reconciled field by field; a trial that cannot reach agreement is withheld rather than published. That is a process control. It is not an accuracy measurement, and it should not be read as one.

The measurement is a pre-registered benchmark: a locked, stratified sample of 150 recruiting trials drawn from a frozen frame, a protocol fixed in advance of any result, human abstraction with a blind subset held back to check the drafts, and a scoring definition naming exactly which fields count and which are ignored.

That benchmark has not been run yet, so there is no accuracy figure to publish. When there is one it will be reported against the frozen protocol, including the parts that score badly. Until then, treat these fields as a structured reading of the protocol text — useful for finding trials worth asking about, and not a substitute for the protocol itself or for the study team.

Machine access

The layer is queryable over the Model Context Protocol at https://www.oncomatch.org/api/mcp — trial search by cancer type and biomarker, biomarker suggestions, and full structured detail for a given NCT ID. No authentication, no key, and queries are not stored.

Underlying registry data is public and belongs to ClinicalTrials.gov; every trial links back to its record there. What is ours is the structuring.

Limits worth knowing

• Extraction reads eligibility text, title, conditions and official title — not arm or intervention tables. Criteria that only resolve by reference to a named arm may be incomplete.
• Trials whose cohorts carry different entry requirements are represented in a flat schema. Where that flattening loses a distinction, the trial is still listed.
• A trial that has not been mapped to a cancer-type panel has a page but does not appear in matching. That gap is tracked, not silent.
• Phase 1 trials are under-represented relative to the registry as a whole.

Questions about the data, or want to build on it? info@oncomatch.org