Datasets:

AIM-Harvard
/

TrialPanorama-filtered

license: apache-2.0
language:
  - en
tags:
  - medical
size_categories:
  - 10M<n<100M
config_names:
  - studies
  - conditions
  - drugs
  - disposition
  - outcomes
  - results
  - biomarkers
  - endpoints
  - relations
  - drug_moa
configs:
  - config_name: studies
    data_files:
      - split: all
        path: studies*.parquet
  - config_name: conditions
    data_files:
      - split: all
        path: conditions*.parquet
  - config_name: drugs
    data_files:
      - split: all
        path: drugs*.parquet
  - config_name: disposition
    data_files:
      - split: all
        path: disposition*.parquet
  - config_name: outcomes
    data_files:
      - split: all
        path: outcomes*.parquet
  - config_name: adverse_events
    data_files:
      - split: all
        path: adverse_events*.parquet
  - config_name: results
    data_files:
      - split: all
        path: results*.parquet
  - config_name: biomarkers
    data_files:
      - split: all
        path: biomarkers*.parquet
  - config_name: endpoints
    data_files:
      - split: all
        path: endpoints*.parquet
  - config_name: relations
    data_files:
      - split: all
        path: relations.parquet
  - config_name: drug_moa
    data_files:
      - split: all
        path: drug_moa.parquet

TrialPanorama-filtered

Filtered version of TrialPanorama/TrialPanorama-database. This release removes any rows where study_id overlaps with IDs present in the task CSVs (NCT IDs and PubMed IDs).

Source

Original dataset: TrialPanorama/TrialPanorama-database
Task CSVs used for overlap IDs:
- task1_status_transition.csv
- task2_completion_termination.csv
- task3_publication_emergence.csv
- task4_serious_ae.csv

Filtering rules

NCT IDs are collected from CSV columns whose headers contain nct (case-insensitive).
PubMed IDs are collected from CSV columns whose headers contain pmid (case-insensitive).
Cells are tokenized on whitespace, commas, semicolons, and |.
A row is removed when study_id matches an overlapping NCT ID (^NCT\d+$) or an overlapping PubMed ID (^\d+$).
Rows with missing or empty study_id are retained.
Files without a study_id column are copied unchanged.

Overlap counts from task CSVs

NCT IDs: 39,779
PubMed IDs: 2,635

Files included

All parquet files from the source dataset are retained with identical schemas. Only row-level deletions are applied where study_id is present.

Subsplit summaries

adverse_events.parquet

Adverse event terms linked to studies (with MedDRA identifiers and seriousness).

Rows (original): 5,400,796
Rows (filtered): 5,329,726
Rows removed: 71,070
Filter applied: yes
Columns: study_id, study_source, adverse_event_name, adverse_event_description, is_serious, meddra_id, meddra_id_type

adverse_events_part_2.parquet

Continuation of adverse events.

Rows (original): 332,068
Rows (filtered): 328,348
Rows removed: 3,720
Filter applied: yes
Columns: study_id, study_source, adverse_event_name, adverse_event_description, is_serious, meddra_id, meddra_id_type

biomarkers.parquet

Biomarker mentions associated with studies.

Rows (original): 256,034
Rows (filtered): 254,298
Rows removed: 1,736
Filter applied: yes
Columns: study_id, study_source, biomarker_type, biomarker_name, biomarker_genes

biomarkers_part_2.parquet

Continuation of biomarkers.

Rows (original): 84,035
Rows (filtered): 78,416
Rows removed: 5,619
Filter applied: yes
Columns: study_id, study_source, biomarker_type, biomarker_name, biomarker_genes

conditions.parquet

Conditions linked to studies with MeSH identifiers.

Rows (original): 6,069,464
Rows (filtered): 6,005,009
Rows removed: 64,455
Filter applied: yes
Columns: study_source, study_id, condition_name, condition_mesh_id, condition_mesh_type

conditions_part_2.parquet

Continuation of conditions.

Rows (original): 1,982,348
Rows (filtered): 1,773,112
Rows removed: 209,236
Filter applied: yes
Columns: study_source, study_id, condition_name, condition_mesh_id, condition_mesh_type

disposition.parquet

Disposition/intervention groups and subject counts.

Rows (original): 2,512,148
Rows (filtered): 2,490,343
Rows removed: 21,805
Filter applied: yes
Columns: study_id, intervention_type, group_type, intervention_name, intervention_description, study_source, number_of_subjects

disposition_part_2.parquet

Continuation of disposition.

Rows (original): 765,033
Rows (filtered): 689,812
Rows removed: 75,221
Filter applied: yes
Columns: study_id, intervention_type, group_type, intervention_name, intervention_description, study_source, number_of_subjects

drug_moa.parquet

Drug mechanism-of-action and target activity annotations (not study-specific).

Rows (original): 19,378
Rows (filtered): 19,378
Rows removed: 0
Filter applied: no
Columns: drug_name, drug_moa_id, target_name, target_class, accession, gene, swissprot, act_value, act_unit, act_type, act_comment, act_source, relation, moa, moa_source, act_source_url, moa_source_url, action_type, tdl, organism

drugs.parquet

Drug interventions linked to studies with normalization and approvals.

Rows (original): 1,028,836
Rows (filtered): 1,022,795
Rows removed: 6,041
Filter applied: yes
Columns: study_id, study_source, drug_moa_id, drug_name, rx_normalized_name, rx_source, rx_mapping_id, drugbank_name, drugbank_id, drugbank_mesh_id, drug_source, fda_approved, ema_approved, pmda_approved, drug_description

drugs_part_2.parquet

Continuation of drugs.

Rows (original): 287,165
Rows (filtered): 269,798
Rows removed: 17,367
Filter applied: yes
Columns: study_id, study_source, drug_moa_id, drug_name, rx_normalized_name, rx_source, rx_mapping_id, drugbank_name, drugbank_id, drugbank_mesh_id, drug_source, fda_approved, ema_approved, pmda_approved, drug_description

endpoints.parquet

Primary endpoints with domain and subdomain.

Rows (original): 1,071,103
Rows (filtered): 1,059,876
Rows removed: 11,227
Filter applied: yes
Columns: study_id, study_source, primary_endpoint, primary_endpoint_domain, primary_endpoint_subdomain

endpoints_part_2.parquet

Continuation of endpoints.

Rows (original): 368,043
Rows (filtered): 320,051
Rows removed: 47,992
Filter applied: yes
Columns: study_id, study_source, primary_endpoint, primary_endpoint_domain, primary_endpoint_subdomain

outcomes.parquet

Study outcomes and termination reasons.

Rows (original): 365,521
Rows (filtered): 357,876
Rows removed: 7,645
Filter applied: yes
Columns: study_id, study_source, overall_status, outcome_type, why_terminated

outcomes_part_2.parquet

Continuation of outcomes.

Rows (original): 159,027
Rows (filtered): 128,112
Rows removed: 30,915
Filter applied: yes
Columns: study_id, study_source, overall_status, outcome_type, why_terminated

relations.parquet

Entity-to-entity relation edges (not study-specific).

Rows (original): 940,582
Rows (filtered): 940,582
Rows removed: 0
Filter applied: no
Columns: head_id, head_type, tail_id, tail_type, relation_type

results.parquet

Aggregated results with population/interventions/outcomes text.

Rows (original): 3,746,215
Rows (filtered): 3,735,669
Rows removed: 10,546
Filter applied: yes
Columns: study_id, study_source, population, interventions, outcomes, group_type

results_part_2.parquet

Continuation of results.

Rows (original): 726,383
Rows (filtered): 722,595
Rows removed: 3,788
Filter applied: yes
Columns: study_id, study_source, population, interventions, outcomes, group_type

studies.parquet

Study-level metadata (titles, recruitment status, phase, accrual, demographics).

Rows (original): 1,010,037
Rows (filtered): 1,002,045
Rows removed: 7,992
Filter applied: yes
Columns: study_id, study_source, trial_type, title, abstract, sponsor, start_year, recruitment_status, phase, actual_accrual, target_accrual, min_age, max_age, min_age_unit, max_age_unit, sex, healthy_volunteers

studies_part_2.parquet

Continuation of studies.

Rows (original): 322,104
Rows (filtered): 290,738
Rows removed: 31,366
Filter applied: yes
Columns: study_id, study_source, trial_type, title, abstract, sponsor, start_year, recruitment_status, phase, actual_accrual, target_accrual, min_age, max_age, min_age_unit, max_age_unit, sex, healthy_volunteers

Quick start

The easiest way to download the dataset to your local is to use huggingface-cli. The specific command you can use is

huggingface-cli download zifeng-ai/TrialPanorama-database --local-dir LOCAL_DIR --repo-type dataset

where LOCAL_DIR should be replaced with the target directory you want to save your dataset to.

Update history

Aug.4 2025: updated tables with the full set of studies

Dataset website: https://ryanwangzf.github.io/projects/trialpanorama

TrialPanorama

TrialPanorama is a large-scale, structured database and benchmark designed to support AI-driven tasks in clinical trial workflows—including systematic review and trial design. This repo hosts the database of clinical trials. To conduct benchmarking experiments on trial design and systematic review asks, check this dataset instead: https://huggingface.co/datasets/zifeng-ai/TrialPanorama-benchmark

Dataset Overview

Aggregates over 1M clinical trial records from ClinicalTrials and PubMed
Captures standardized elements such as trial setups, interventions, conditions, biomarkers, outcomes, and links to biomedical ontologies (e.g. DrugBank, MedDRA).
Structured into multiple conceptual clusters and tables (e.g. trial-level attributes, protocol design, results, links)

Benchmark Suite

The dataset supports a suite of 8 benchmark tasks across two domains:

Systematic Review:
- Study search
- Study screening
- Evidence summarization
Trial Design:
- Arm design
- Eligibility criteria
- Endpoint selection
- Sample size estimation
- Trial completion assessment

Data Schema (Major Tables)

studies — core metadata (e.g. study_id, intervention type, sponsor type, start year, trial phase, recruitment status)
protocols, interventions, conditions, biomarkers, outcomes, results — each standardized to a unified schema

Use Cases

Developing and evaluating LLMs and ML models for clinical trial-related tasks
Benchmarking AI capabilities in trial search, screening, design, and summarization
Conducting meta-analyses and exploring evidence synthesis across disease areas
Enabling interoperability with biomedical ontologies to support richer clinical trial reasoning

Citation

If you use TrialPanorama, we appreciate your citation:

@article{wang2025trialpanorama,
  title={TrialPanorama: Database and Benchmark for Systematic Review and Design of Clinical Trials},
  author={Wang, Zifeng and Jin, Qiao and Lin, Jiacheng and Gao, Junyi and Pradeepkumar, Jathurshan and Jiang, Pengcheng and Danek, Benjamin and Lu, Zhiyong and Sun, Jimeng},
  journal={arXiv preprint arXiv:2505.16097},
  year={2025}
}