Retrieval over 2026 Clinical Guidelines Recovers Post-Cutoff Knowledge in MedGemma-27B: an Open, Verifiable Pilot
Preprint, not peer reviewed. A lightweight RAG pipeline built on a fully open and verifiable foundation: 773 CC BY guideline and consensus abstracts published in 2026 (Europe PMC), BM25 retrieval, a dev-tuned confidence gate and MedGemma-27B run locally. On 131 held-out questions accuracy rose from 0.649 to 0.908 (paired z = 6.07). Open-book evaluation: questions generated by the same model from indexed abstracts, no clinical review. Research use only, not for diagnosis. Not affiliated with or endorsed by Google; MedGemma is subject to the HAI-DEF Terms of Use.Questa versione include anche una versione divulgativa in italiano per un pubblico non tecnico (file linee-guida-2026-rag-it_divulgativa_it.pdf).
Authors
- Giovanni Del Papa
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-10-26
- DOI
- https://doi.org/10.5281/zenodo.23664674
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- preprint