ProtocolEC

A protocol-derived benchmark for complete clinical-trial eligibility-criteria (EC) generation.

ProtocolEC pairs each of 4,302 completed Phase III trials (22 therapeutic areas) with (i) its ClinicalTrials.gov registry metadata and registry EC, and (ii) a more complete EC set extracted from the trial's protocol PDF. Protocol EC contain roughly twice the criteria and words of the registry EC. Splits are 80/10/10, stratified by therapeutic area.

Split	Trials
train	3,443
validation	429
test	430

Task

Given a trial's features (title, summary, phase, conditions, interventions, demographics), generate the complete, protocol-level eligibility criteria (complete eligibility criteria). The registry EC (Eligibility Criteria (registry)) is a compressed public summary and serves as a baseline / comparison, not the target.

from datasets import load_dataset
ds = load_dataset("<your-org>/protocolec")
ex = ds["test"][0]
ex["Eligibility Criteria (registry)"]   # compressed registry summary (baseline)
ex["complete eligibility criteria"]     # protocol-extracted GOLD target

Columns (35)

• Identifier: NCT ID
• Target (gold): complete eligibility criteria — protocol-extracted complete EC
• Registry EC (baseline): Eligibility Criteria (registry)
• Trial features (input): Brief Title, Official Title, Brief Summary, Detailed Description, Phases, Conditions, Interventions, Intervention Model, Primary Purpose, Masking, Arms, Study Type, Keywords, SNOMED CT, primary_therapeutic_area, Enrollment Count, Overall Status, Organization, Acronym, Primary Outcomes, Secondary Outcomes, Start Date, Completion Date, Primary Completion Date
• Demographics: Sex, Minimum Age, Maximum Age, Standard Ages, Healthy Volunteers
• Provenance: Protocol Doc Count, Protocol Filenames, Protocol URLs

Protocol PDFs

The source protocol PDFs are not redistributed here. Each row carries Protocol URLs (ClinicalTrials.gov CDN links), from which the PDFs can be downloaded directly if needed.

Construction (summary)

Phase III trials with an attached protocol PDF were retrieved from the ClinicalTrials.gov v2 API; the EC section of each PDF was localized with a deterministic three-stage locator; EC were extracted verbatim with an LLM extractor; multiple protocol documents per trial were merged; each trial was assigned a primary therapeutic area; and a blinded three-judge audit validated coverage/enrichment against the registry EC. See the paper for full details.

License & ethics

Released for non-commercial research use (CC BY-NC 4.0). Source records are U.S. public-domain ClinicalTrials.gov content. The data contains no patient-level records or identifiable health information: eligibility criteria describe general study-population requirements, not individuals.

Citation

@article{protocolec,
  title  = {Beyond the Registry: A Protocol-Derived Dataset and Embedding-Guided Method for Complete Clinical Trial Eligibility Criteria Generation},
  author = {<authors>},
  year   = {2026}
}