Main Session

Sep 28

PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology

2652 - Do Large Language Models Read or Remember? Analyzing LLM Performance In Biomedical Text Processing with Progressive Content Removal and Counterfactual Results

10:45am - 12:00pm ET

Poster Hall - Exhibit Hall A

Screen: 22

POSTER

Presenter(s)

Daniel Zwahlen, MD, MBA - Department of Radiation Oncology, Winterthur, Zurich

P. Windisch^1,2, C. Koechli², F. Dennstaedt², D. M. Aebersold², D. R. Zwahlen³, R. Foerster^1,2, and C. Schroeder^1,2; ¹Department of Radiation Oncology, Cantonal Hospital Winterthur, Winterthur, Switzerland, ²Department of Radiation Oncology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland, ³Department of Radiation Oncology, Kantonsspital Winterthur, Winterthur, Switzerland

Purpose/Objective(s):

Large language models (LLMs) can classify biomedical documents accurately, but strong performance does not prove they are using the supplied text rather than identifier-triggered parametric knowledge. We tested whether oncology trial-success classification reflects “reading” of abstract evidence or “remembering” of known trials.

Materials/Methods:

We used a corpus of 250 two-arm oncology randomized controlled trials from seven major journals (2005 - 2023) and asked the flagship models of three commercial vendors (OpenAI, Google, and Anthropic) to output a single label indicating whether the primary endpoint was met. For each trial we created five deterministic inputs: title+abstract (baseline), title-only, DOI-only, counterfactual title+abstract with the primary endpoint outcome minimally flipped, and the same counterfactual title+abstract paired with the original DOI to induce an identifier-text conflict.

Results:

With full title+abstract, models achieved near-ceiling performance (accuracy and F1 Score 0.96 - 0.97) and high format adherence (97.2 - 100%). Performance degraded stepwise with content removal (title-only accuracy and F1 Score 0.79 - 0.88, DOI-only 0.63 - 0.67), consistent with above-chance identifier-driven signal. Under counterfactual results, models followed the edited evidence (accuracy and F1 Score 0.96 - 0.99 against inverted labels). Adding the real DOI minimally affected GPT (accuracy and F1 Score ˜ 0.99) but modestly reduced Gemini (accuracy and F1 Score ˜ 0.97) and Claude (accuracy and F1 Score ˜ 0.95), mainly via lower sensitivity.

Conclusion: LLMs robustly track explicit endpoint statements in abstracts, yet identifiers can support above-chance predictions and occasionally compete with textual evidence. Progressive ablations plus counterfactual conflicts provide a practical, reproducible audit for grounding in biomedical LLM evaluations.