PAHD: Physical Activity for Hypertension Database

PAHD (Physical Activity for Hypertension Database) is a structured database of the results of 253 randomized controlled trials of exercise interventions and blood pressure (BP). Each row is one comparison reported in a study: an outcome, two sides (groups and/or time points), the statistics of each side, the reported difference and the p-value. Each row also carries the study design and the arms' exercise protocols and baseline characteristics. The data were extracted from the full texts with large language models (LLMs), checked against the papers, standardized and quality-checked.

The studies come from a PubMed search for hypertension, exercise and blood pressure: randomized controlled trials in adults, in English, published between 2005-01-01 and 2025-12-31, with an accessible abstract and full text. An LLM screened 2,548 records by title and abstract and then by full text against five eligibility criteria, which selected 255 studies. Two of them were removed during database construction, leaving 253 studies.

Files:

• PAHD_flat_table.csv / .parquet: the database, 17,273 rows (one per reported comparison) and 90 readable columns
• PAHD_flat_table_full.csv / .parquet: the same rows with 186 columns: the numeric statistics of each side and of the contrast (means, SDs, CIs, differences), IDs, and the protocol and baseline fields in structured form
• included_studies.csv: the 253 studies, with PMID, PMCID, authors, year, title, journal, DOI and number of rows
• column_dictionary.csv: every column of the three tables, with description, type, unit, allowed values, fill rate and an example
• README.md: how to read a row, the standard vocabularies, the quality status and the flag codes

Standard vocabularies: US OMB SPD 15 (2024) for race and ethnicity; WHO ATC classes for medications; ICD-10 for comorbidities; WHO adult cut-offs for BMI categories; the 2025 AHA/ACC stage for hypertension; and ISO 3166-1 alpha-3 codes for regions.

Quality: every row was checked by deterministic rules against its own numbers, against the quotes the LLMs gave as evidence, and against the paper text. Flagged BP rows were reviewed by hand against the papers. Each row has a qc_status:

• clean: 15,389 rows
• auto_fixed: 390 rows
• reviewed: 392 rows
• flagged: 1,102 rows

Any open flags are listed in qc_flags.

Web portal: https://01a02594-cf5d-61a3-bb24-057117ed5c26.share.connect.posit.cloud/

Benchmarks used to evaluate the pipeline: https://doi.org/10.5281/zenodo.23785622

Please cite: Li C, Obioha C, Cui X, Zhou X, Rosenzweig A, Wang M. A Knowledge Database of Physical Activity Interventions for Hypertension using Large Language Models. Manuscript submitted for publication.