Truace
Subscribe

Large language model responses to structured textual descriptions of salivary gland imaging findings: impact of contextual enrichment on self-reported diagnostic confidence and reporting behavior

Purpose: Large language models (LLMs) are increasingly applied in clinical decision support, yet their behavior when presented with structured textual descriptions of imaging findings, particularly in salivary gland disease, remains poorly defined. The present study was set out to evaluate how the responses of different LLMs change as the textual input is progressively enriched. The models did not receive imaging data; all inputs were text. Methods: This exploratory, simulation-based observational study assessed…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2044· cite & verify
Large language model responses to structured textual descriptions of salivary gland imaging findings: impact of contextual enrichment on self-reported diagnostic confidence and reporting behavior
Doctor Wenlock outside St Thomas's Hospital - geograph.org.uk - 3101668 by PAUL FARMER. CC BY-SA 2.0 · https://creativecommons.org/licenses/by-sa/2.0

In brief

Researchers gave ChatGPT, Gemini and Claude text-only descriptions of imaging findings from 10 salivary gland cases. Each case was presented in three stages: a single-plane description, a multiplanar description, and a multiplanar description with clinical information, yielding 90 observations.

The results suggest that more detailed prompts can make a model sound more certain without showing that its reasoning improved. That matters for anyone considering LLMs in clinical support. The study was a small exploratory pilot with no reference diagnosis, no image interpretation and no support for standalone diagnostic use, so accuracy remains unknown.

Main points

  1. Self-reported confidence rose significantly with richer input (Friedman χ²=18.86; P < 0.001).
  2. The cumulative odds ratio for confidence was 5.43 (95% CI 3.00-9.81) for the fullest input versus the single-plane description.
  3. Severity scores did not differ significantly across the three models (P = 0.31).
  4. Claude reported a mean of 5.67 additional anatomical structures versus 0.00 for ChatGPT.
  5. Gemini gave lower severity scores but comparable self-reported confidence.

The problem

Self-reported diagnostic confidence in three LLMs rose as prompts were enriched, which may reflect expressed certainty rather than better diagnostic reasoning, and no outputs were checked against a reference diagnosis.

The rundown

The study was simulation-based and observational. It used 10 salivary gland cases, each with a standardized textual description of imaging findings. No image files were uploaded, so the models never analyzed actual imaging.

The primary outcomes were the model-generated severity score, self-reported diagnostic confidence, and the number of additional anatomical structures reported. The main analysis used a proportional-odds model with cluster-robust standard errors by case, and non-parametric tests served as sensitivity analyses.

Reporting behavior varied by model. ChatGPT produced stable outputs with no additional structures, while Claude produced much more detailed reports, and a mixed model showed a significant effect of AI system on verbosity (P < 0.001). The authors treat the link between reported parenchymal changes and confidence as a within-output association, not an independent diagnostic relationship.

Sources

  1. Peer-reviewedEuropean Archives of Oto-Rhino-Laryngology2026-10-20

The debate