Dataset

clinical-trials-eligibility-graphs-rerank

Public Records & Filings Download 1 source

Parsed eligibility criteria from 13,229 ClinicalTrials.gov studies representing the candidate pool gathered by two first-stage retrievers during TREC Clinical Trials 2021–2023, formatted as typed entity-relation graphs for reranking models.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

This dataset contains parsed eligibility criteria from 13,229 ClinicalTrials.gov trials, represented as entity and relation graphs produced by a 20B-parameter biomedical extraction model. It is published as a rerank candidate set used in TREC Clinical Trials 2021–23 research.

What this dataset contains

The dataset is a single train split with 13,229 rows, each corresponding to a ClinicalTrials.gov trial identified by nct_id. Inclusion and exclusion criteria have been extracted into entity and relation graphs stored as JSON strings in the inclusion_entities, inclusion_relations, exclusion_entities, and exclusion_relations columns. Coverage is limited to trials surfaced in the top-100 retrieval results for TREC Clinical Trials topics across 2021–23, and one trial (NCT03595566) is absent due to an extraction failure.

Structure and fields

Each row carries the nct_id large_string identifier alongside four graph columns stored as large_string JSON: inclusion_entities, inclusion_relations, exclusion_entities, and exclusion_relations. Four int64 counters, n_inc_entities, n_inc_relations, n_exc_entities, and n_exc_relations, summarize graph size and can be used to filter trials where one criterion side is empty. The schema is heterogeneous on purpose, with optional code_expr fields and variable relation arity, so graphs should be parsed on read rather than flattened.

How analysts use it

The dataset is intended for reranking and eligibility-classification research on TREC Clinical Trials, allowing a CrossGAT eligibility model to be trained without invoking an LLM. Users parse the JSON graph columns, filter on n_inc_entities or n_exc_entities where both sides are needed, and combine the graphs with the underlying TREC collection under its data usage agreement. Extracted content is model output and contains extraction errors, so it should be treated as research input rather than curated annotation.

Documentation & references 1