LLM framework mines 5,254 papers to uncover 44 research cohorts

Researchers led by Moritz Sturm and colleagues have developed a question-driven framework that uses large language models (LLMs) to automatically discover study cohorts from scientific literature. The framework generates multiple PubMed queries from configurable vocabularies and templates, retrieves relevant records through the PubMed API, and then uses an LLM to screen titles and abstracts, extracting explicit cohort names with a tailored prompt. A human review step deduplicates the extracted names. Published on arXiv (2608.15909), the approach is designed to complement traditional cohort catalogues and manual literature searches, which often miss relevant cohorts when planning multi-study analyses.

As a use case, the team applied the framework to youth aggression genetics. From 5,400 generated queries, it retrieved 5,254 unique records and identified 188 candidate cohorts. Manual screening using predefined criteria — including participant age and genetic-data availability — retained 44 eligible cohorts. Notably, automated LLM-based name extraction was within the agreement range of human annotators. When the same research question was searched across four established cohort catalogues, their combined results contained only 27 of the 44 eligible cohorts, while 17 were not returned by any catalogue search.

The framework effectively converts research-question-specific vocabulary into screenable cohort inventories via a large, automated literature search. It can be adapted across populations, phenotypes, data modalities, and study designs, making it a flexible tool for researchers. By surfacing cohorts that curated catalogues miss, the framework helps accelerate the discovery of participants for multi-study analyses, particularly in fields like genetics where precise phenotypic matching is critical. The authors have made configurable code, prompts, and example outputs publicly available, allowing other researchers to adapt the framework to their own questions.