A Decade of AI-Based Synthetic Data in Biomedicine: annotated PubMed corpus (2015-2025), LLM-assisted facet annotations and citation analysis [Data set]

Dataset title:
  A Decade of AI-Based Synthetic Data in Biomedicine: annotated PubMed corpus
  (2015-2025), LLM-assisted facet annotations and citation analysis

Contact person (deposit and data curation):
  Gorka Epelde | Strategy and Impact Department, Biogipuzkoa Health Research
  Institute, San Sebastian, Spain | gorka.epeldeunanue@bio-gipuzkoa.eus |
  ORCID: 0000-0002-5179-415X

Project description:
  This dataset supports a systematic literature review of AI-based synthetic
  data generation in biomedicine over the decade 2015-2025. The study maps the
  methodological landscape, its temporal evolution, and patterns of adoption,
  citation impact and practical translation, based on a corpus of 4,143 PubMed
  records retrieved with a structured Boolean query targeting the intersection
  of synthetic/augmented data, biomedical domains, and AI/ML technologies.

  Beyond the landscape findings, the study contributes a methodology for scaling
  systematic reviews through LLM-assisted annotation validated against human
  expert consensus. Six independent expert annotators established ground truth
  on successive batches of papers, iteratively refining the labelling
  guidelines; multiple LLMs were then used as semi-automated annotators and
  evaluated against that ground truth before being applied to the full corpus.

Data description:
  The deposit contains four groups of tabular files in .xlsx format, plus the
  analysis code:

  (a) The source corpus of 4,143 PubMed records
      -> csv-SyntheticD-set2015-April2025.xlsx
  (b) The consolidated majority-vote annotations across five conceptual facets
      -> majority_vote_annotations.xlsx
  (c) The corpus enriched with OpenAlex citation statistics, both as a whole
      and as one file per facet
      -> papers_with_citation_stats_*.xlsx
  (d) The expert annotation ground truth and the per-model LLM annotations
      -> Human-results.xlsx, GPT4/Ollama/QWEN-results-cleanedAndUnified.xlsx
  (e) The analysis code: one Jupyter notebook for the citation analysis, and
      seventeen Python modules covering majority-vote consolidation,
      inter-annotator agreement, temporal trends and facet distributions
      -> citation_analysis.ipynb, *.py, dependency_map_code.pdf

  All tabular files are linked through the "Paper ID" key. Missing values are
  represented as empty cells. Where a facet label could not be assigned, the
  value "Other" is used in the Medical Category facet; the remaining facets
  have no residual category.

Publication date: 2026-08-03
Creation date:    2025-11-06 (OpenAlex citation retrieval date; see Section 3)
Language:         en