Researchers build LLM-based database of exercise interventions for hypertension

[medRxiv] A preprint study by Li et al. describes PAHD (Physical Activity for Hypertension Database), a structured database of exercise interventions for hypertension built with large language models. According to the study, one LLM screened 2,548 PubMed records by title and abstract and then by full text, selecting 255 studies; LLMs extracted each study's groups, exercise protocols, patient characteristics and results, two LLMs cross-checked study structure, and every value carries a verbatim quote. Code standardized and joined the extractions into 17,312 records from 254 studies, one per reported comparison. Against adjudicated benchmarks, screening reached 86.3% recall and 80.6% precision, and extraction recovered 90.1% of blood pressure values with 90.5% precision, according to the study. A web portal matches patients' hypertension stage and comorbidities to studies. The preprint has not been peer-reviewed.