2001jdev/clinical-trials-eligibility-graphs-rerank

Clinical Trial Eligibility Graphs (rerank candidate set)

Parsed eligibility criteria for 13,229 ClinicalTrials.gov trials — the complete candidate set retrieved across TREC Clinical Trials 2021–23 by two first-stage retrievers. Each trial's inclusion and exclusion criteria are represented as entities and typed relations.

This exists so the CrossGAT eligibility model can be run without an LLM. Generating these graphs took ~28 GPU-hours on a 20B-parameter model; this dataset is that output, so reproduction costs nothing.

Schema

column	type	description
nct_id	string	ClinicalTrials.gov identifier
inclusion_entities	string (JSON)	entities parsed from inclusion criteria
inclusion_relations	string (JSON)	typed relations between them
exclusion_entities	string (JSON)	entities parsed from exclusion criteria
exclusion_relations	string (JSON)	typed relations between them
n_inc_entities, n_inc_relations	int	counts, for filtering without parsing
n_exc_entities, n_exc_relations	int	counts

Graph columns are JSON strings rather than nested structs: entity and relation schemas are heterogeneous (optional code_expr, variable relation arity), and flattening them into a fixed Arrow schema would lose fields. Parse on read.

import json
from datasets import load_dataset

ds = load_dataset("2001jdev/clinical-trials-eligibility-graphs-rerank", split="train")
row = ds[0]
inc_entities = json.loads(row["inclusion_entities"])
# {"id": "S0_T1", "type": "Condition", "text": "rheumatoid arthritis", "code_expr": []}
inc_relations = json.loads(row["inclusion_relations"])
# {"id": "S0_R1", "type": "HAS_CONTEXT", "arg1_id": "S0_T2", "arg2_id": "S0_T1"}

Coverage and known gaps

• 13,229 trials, all with a non-empty inclusion graph.
• 859 trials (6.5%) have an empty exclusion graph. Most genuinely state no exclusion criteria; some are parse failures. The model handles an empty side — it is trained on data containing them — but filter on n_exc_entities if you need both sides populated.
• Median 25 inclusion entities and 23 exclusion entities per trial.
• This is a candidate set, not a corpus. It covers what two specific retrievers surfaced in the top-100 for TREC CT topics, not ClinicalTrials.gov at large. A trial neither retriever ever retrieved is absent.

Provenance

Generated with openai/gpt-oss-20b via a CHIA-style extraction prompt, over the ClinicalTrials.gov snapshot used by TREC Clinical Trials 2021–23. Cached by NCT id during benchmark runs and consolidated here; 182 MB of JSON, 17.7 MB as zstd-compressed Parquet. A random 200-trial sample was verified to round-trip byte-identically.

One trial in the original candidate set, NCT03595566, produced no graph on any attempt and is absent.

Intended use

Reranking and eligibility-classification research on TREC CT. Users must comply with the TREC data usage agreements for the underlying collection. Extracted criteria are model output, not curated annotations — they contain extraction errors and must not be used for clinical decisions.