LLM agents draft sarcoma data schema from 24,588 abstracts, with expert curation still required
Researchers tested whether large language model agents could build a common data element schema from sarcoma case report abstracts and apply it across a large corpus, reporting mixed but promising review results.
In plain words
Common data elements are standard fields that make clinical records comparable across studies; they are usually defined by hand. In this arXiv paper, AI agents proposed candidate fields from sarcoma study summaries, then consolidated them into a smaller schema used to pull structured answers from the same abstracts. Human reviewers and a sampling check found the results useful but imperfect, so expert review remains necessary.
What was announced
The agents proposed 72,909 candidate common data elements from 24,588 PubMed abstracts, and an agentic consolidation step reduced these to 1,570. The authors retained 53 elements supported by at least 50 abstracts and ran evidence-linked extraction across the corpus. Processing succeeded for all retained elements in 79.2% of abstracts, counting responses that indicated a field was not mentioned.
Two human reviewers approved 83.0% of the retained elements for relevance, 67.9% for clarity, and 56.6% for correctness. In 100 sampled abstracts, exact agreement with the model's extracted value was 94.9% among 1,116 pairs with a non-empty answer, with a bootstrap 95% confidence interval of 93.0-96.4% at the abstract level. Errors concentrated in anatomical and temporal fields. The authors conclude the approach supports agent-assisted schema drafting followed by expert curation before large-scale extraction.
What it does not show
- The source is an arXiv abstract, so methods detail, peer review status, and full results are not available in the supplied evidence.
- Reviewer approval for correctness was 56.6%, and extraction agreement was measured on only 100 sampled abstracts, so performance across the full corpus is not established.
- The study covers sarcoma case report abstracts; applicability to other clinical domains is not shown.
For security and IT teams
- Scope
- The reported pipeline covers sarcoma case report abstracts from PubMed and does not address other data types or clinical areas.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.