Emotional Framing Effects on Medical Triage in General-Purpose and Medical-Adapted LLMs

Large language models (LLMs) are increasingly used for health advice, yet their triage decisions are usually evaluated on vignettes written in neutral language, whereas real patients often describe their symptoms with worry or reassurance. This study examines whether emotional framing changes the triage decisions of a general-purpose model, Gemma 3 27B, and its medical counterpart, MedGemma 27B, which shares the same architecture and size. A total of 448 class-balanced vignettes from a public triage dataset were neutralized and then presented under four conditions: neutral, mild anxiety, high anxiety, and self-dismissing framing. Each of the 1,792 prompts was run three times per model, and the primary measure was the shift in each model’s triage level relative to its own neutral prediction. Framing changed the predictions of both models, although the effects were small and affected at most 7.4% of cases. Self-dismissing framing lowered urgency in Gemma 3 but not in MedGemma. Contrary to expectations, anxious framing never raised urgency and instead lowered it in both models, likely because both already over-triaged most cases at baseline. Emergency cases remained almost entirely stable. MedGemma was more robust to self-dismissal but responded to high anxiety at least as strongly as Gemma 3. These findings suggest that medical domain adaptation changes which emotional cues a model responds to rather than making it robust to framing in general. Accepted for presentation at ICICyTA 2026 (Paper #1571376873).

Recommended citation: Taslim, S. V., Amethyst, M. G., Tarigan, G. A., Nugroho, K. S., & Pardamean, B. (2026). "Emotional Framing Effects on Medical Triage in General-Purpose and Medical-Adapted LLMs." 2026 International Conference on Intelligent Cybernetics Technology & Applications (ICICyTA).