LLM agents draft sarcoma data elements from PubMed abstracts
Original title: Agentic induction of common data elements and information extraction from sarcoma case report abstracts
From 24,588 abstracts, agents kept 53 elements; extracted values matched humans on 94.9% of non-empty pairs.
Researchers tested whether large language model agents can propose common data elements from sarcoma case-report abstracts and then extract them. An agent produced 72,909 candidate elements from 24,588 PubMed abstracts; consolidation left 1,570, and 53 supported by at least 50 abstracts were retained. Extraction completed for every retained element in 79.2% of abstracts, counting non-mentions. Two reviewers approved 83.0% of retained elements for relevance, 67.9% for clarity, and 56.6% for correctness. On 100 sampled abstracts, exact agreement with the LLM value was 94.9% among 1,116 non-empty pairs.
Unconfirmed: Reviewer correctness approval was only 56.6%, so the drafted schema still needs expert curation.
- Candidate elements numbered 72,909 before consolidation to 1,570.
- Retained elements needed support from at least 50 abstracts.
- Bootstrap 95% CI for exact agreement was 93.0–96.4%.
- Errors concentrated in anatomical and temporal fields.
- Organizations
- PubMed
From the original article
1 min read at arxiv.orgCommon data elements (CDEs) support clinical data comparability but are usually defined manually. We investigated whether large language model agents could derive a CDE schema from sarcoma case report abstracts and apply it at scale. An agent proposed 72,909 candidate CDEs from 24,588 PubMed abstracts; agentic consolidation yielded 1,570 CDEs.
By Daniel Reichenpfader · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.