AI models extract sex-based blood pressure data from medical literature

A team of researchers has demonstrated the feasibility of deploying large language models and few-shot learning to systematically extract blood pressure data from medical publications, with particular attention to how readings differ across biological sex.

The problem with current standards#

Modern blood pressure diagnostics rely on guidelines established decades ago that remain in global use today, even though they typically do not account for individual demographic variations tied to sex or age. While these uniform standards help clinicians identify at-risk patients, they lack the granularity needed for fully accurate diagnosis because they ignore meaningful differences across populations.

Methodology#

The team built a pipeline combining natural language processing with LLM-based extraction. They:

• Developed a Solr-based search engine to retrieve blood pressure-related articles from PubMed, filtering for publications that discussed biological sex differences
• Manually reviewed 213 articles, isolating 90 that reported blood pressure statistics broken down by sex
• Tested three automated extraction approaches: a few-shot learning method, as well as zero-shot techniques using LLaMA3 and GPT-3.5
• Extracted means and standard deviations of blood pressure readings, paired with sex classification

Analysis and visualization#

Once the data was automatically extracted, the researchers generated heatmaps and contour plots to visualize how blood pressure values vary across biological sex categories. This visualization approach allowed patterns in the literature to emerge in a form that could guide future standard-setting.

Why it matters#

Automating information extraction from the scientific literature could accelerate the process of refining clinical guidelines to reflect real demographic diversity—potentially improving diagnostic accuracy and reducing the risk of missed or misclassified hypertension cases across different populations.