Academic Journal
Reasoning or reciting? A temporal contamination audit of large language models in clinical medicine.
| Τίτλος: | Reasoning or reciting? A temporal contamination audit of large language models in clinical medicine. |
|---|---|
| Συγγραφείς: | Sheppert AP; Internal Medicine, Legacy Health System, Portland, OR 98686, United States.; Internal Medicine, The Vancouver Clinic, Vancouver, WA 98686, United States.; Artificial Intelligence, Capitol Technology University, Laurel, MD 20708, United States., Adams B; Vascular Surgery, Intermountain Healthcare, Salt Lake City, UT 84604, United States.; Bioinformatics, University of Utah, Salt Lake City, UT 84112, United States., Sheppert AD; Internal Medicine, Legacy Health System, Portland, OR 98686, United States.; Internal Medicine, The Vancouver Clinic, Vancouver, WA 98686, United States., Riley S; Internal Medicine, Legacy Health System, Portland, OR 98686, United States.; Internal Medicine, The Vancouver Clinic, Vancouver, WA 98686, United States. |
| Πηγή: | Journal of the American Medical Informatics Association : JAMIA [J Am Med Inform Assoc] 2026 Sep 01; Vol. 33 (9), pp. 1617-1626. |
| Τύπος έκδοσης: | Journal Article |
| Γλώσσα: | English |
| Στοιχεία περιοδικού: | Publisher: Oxford University Press Country of Publication: England NLM ID: 9430800 Publication Model: Print Cited Medium: Internet ISSN: 1527-974X (Electronic) Linking ISSN: 10675027 NLM ISO Abbreviation: J Am Med Inform Assoc Subsets: MEDLINE |
| Imprint Name(s): | Publication: 2015- : Oxford : Oxford University Press Original Publication: Philadelphia, PA : Hanley & Belfus, c1993- |
| Ιατρικοί όροι (MeSH): | Large Language Models* , Clinical Reasoning* , Artificial Intelligence* , Clinical Medicine*, Humans ; Semantics |
| Περίληψη: | Objective: Evaluate whether large language models reason or simply regurgitate training data in clinical diagnosis. Materials and Methods: We audited 2000 clinical case reports from PubMed Central: 1000 from 2021 to 2022 (within training data) and 1000 from 2025 (after training cutoffs). Five frontier LLMs generated diagnoses evaluated by an independent AI judge validated against physician consensus (n = 10 000 evaluations). Results: Diagnostic accuracy was virtually identical across temporal cohorts (66.8% contaminated vs 66.9% clean), directly contradicting the memorization hypothesis. Lexical similarity was uniformly low (mean ROUGE-L 0.057), and semantic similarity measured by BERTScore showed no memorization signal (F1 0.8182 contaminated vs 0.8195 clean, Δ = +0.0013), confirming that models generate novel reasoning rather than regurgitating training data. Discussion: This large-scale audit, using both lexical and semantic similarity metrics, provides compelling evidence that LLMs engage in genuine clinical reasoning rather than regurgitating memorized training data. Conclusion: Models demonstrated equivalent accuracy on cases they could not have seen during training, suggesting they have internalized generalizable medical knowledge rather than memorizing specific cases. (© The Author(s) 2026. Published by Oxford University Press on behalf of the American Medical Informatics Association. All rights reserved. For commercial re-use, please contact reprints@oup.com for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact journals.permissions@oup.com.) |
| Contributed Indexing: | Keywords: artificial intelligence; clinical reasoning; data contamination; large language models; medical diagnosis |
| Entry Date(s): | Date Created: 20260612 Date Completed: 20260916 Latest Revision: 20260918 |
| Update Code: | 20260918 |
| PubMed Central ID: | PMC13580742 |
| DOI: | 10.1093/jamia/ocag069 |
| PMID: | 42281378 |
| Βάση Δεδομένων: | MEDLINE |
| ISSN: | 1527-974X |
|---|---|
| DOI: | 10.1093/jamia/ocag069 |