Patient tone influences AI-generated clinical advice in E-medicine

A recent study has shown how large language models (LLMs) treat patient tone as clinical input, leading to adjustments in triage, follow-up, prescribing, and sick-leave decisions despite identical symptoms.

“These tone-sensitive shifts may introduce hidden biases, affect resource use, and enable misuse in E-medicine workflows,” the researchers said.

A total of 1,000 clinician-validated primary-care vignettes (500 clinical, 500 sick-leave) were created in this study, and each was presented in eight communication styles. Five agentic LLMs then generated structured outputs for triage urgency, sick-leave decisions, and other outputs.

The researchers used chi-square tests (Cramer’s V) and t-tests (Cohen’s d), with FDR correction, to examine the differences from the neutral control. As external validation, they also processed 40 real patient e-messages from a large health network using the same pipeline.

Patient tone generated clear and reproducible shifts across 120,000 agent runs, with urgent, threatening, and demanding framings increasing same-day or urgent care from 14 percent to 63 percent (V up to 0.69; p<0.001). Medication advice shifted modestly toward prescription options (Rx from 5 percent to 9 percent; p<0.001), while emotional tone boosted empathy-based responses from 62 percent to 86 percent (p<0.001).

In sick-leave tasks, a threatening tone resulted in reduced approvals (58 percent to 50 percent) and granted days (2.60 to 2.36; d=‒0.12), but an emotional tone tended to increase both slightly. The same directional effects were observed in the real-world validation, confirming that tone influenced model outputs even in authentic messages.

“E-medicine use has surged, and health systems are exploring LLMs for message triage,” the researchers said. “However, it is still unknown whether patient tone alone alters AI-generated clinical or administrative decisions.”