A knowledge database of physical activity interventions for hypertension using large language models

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Physical exercise lowers blood pressure, but guidelines provide broad recommendations that rarely account for age, comorbidities, medications and hypertension severity. A structured, queryable resource of exercise interventions for hypertension is lacking, while manual curation is labor-intensive and difficult to scale. We present PAHD (Physical Activity for Hypertension Database), built with large language models (LLMs). An LLM screened 2,548 PubMed records through title-and-abstract and full-text assessment, selecting 255 studies. LLMs extracted arms, exercise protocols, participant characteristics, and outcomes, with two LLMs cross-checking study structure and each value linked to a verbatim quote. Programmatic integration generated 17,273 standardized records from 253 studies, one per reported comparison. Against adjudicated benchmarks, screening achieved 86.3% recall and 80.6% precision, and extraction 90.1% recall and 90.5% precision for blood pressure values. A clinician-facing portal retrieves studies matched to patients’ hypertension stage and comorbidities. PAHD is a reusable, traceable evidence resource for personalized exercise interventions in hypertension.

Author summary

Regular exercise lowers blood pressure, but people with high blood pressure differ in age, other illnesses and medications, and current guidelines give mostly general advice. Hundreds of clinical trials have tested specific exercise programs in specific groups of patients, but their findings are scattered across papers and laborious to collect by hand. We used large language models to find these trials and read them. The models selected 255 trials from 2,548 candidate records and recorded who participated, what exercise they did and how their blood pressure changed, linking every value to a quotation from the paper so that it can be checked. Compared with careful human reviews, the models captured most information accurately. The resulting database holds 17,273 records from 253 trials, and a website lets clinicians find the exercise programs that were tested in patients with characteristics similar to their own patients. We hope this resource helps researchers and clinicians use the existing trial evidence on exercise for hypertension.

Article activity feed