How large language model-powered conversational agents influence decision making in domestic medical triage contexts

Abstract

Read online

IntroductionEffective delivery of healthcare depends on timely and accurate triage decisions, directing patients to appropriate care pathways and reducing unnecessary visits. Artificial Intelligence (AI) solutions, particularly those based on Large Language Models (LLMs), may enable non-experts to make better triage decisions at home, thus easing the healthcare system's load. We investigate how LLM-powered conversational agents influence non-experts in making triage decisions, further studying different persona profiles embedded via prompting.MethodsWe designed a randomized experiment where participants first assessed patient symptom vignettes independently, then consulted one of the two agent profiles—rational or empathic—for advice, and finally revised their triage ratings. We used linear models to quantify the effect of the agent profile and confidence on the weight of advice. We examined changes in confidence and accuracy of triage decisions, along with participants' perceptions of the agents.ResultsIn a study with 49 layperson participants, we found that persona profiles can be differentiated in LLM-powered conversational agents. However, these profiles did not significantly affect the weight of advice. Notably, less confident participants were more influenced by LLM advice, leading to larger adjustments to initial decisions. AI guidance improved alignment with correct triage levels and boosted confidence in participants' decisions.DiscussionWhile LLM advice improves triage recommendations accuracy, confidence plays an important role in its adoption. Our findings raise design considerations for human-AI interfaces, highlighting two key aspects: encouraging appropriate alignment with LLMs' advice and ensuring that people are not easily swayed in situations of uncertainty.

Keywords

• human-AI interaction
• decision-making
• empirical studies
• LLMs
• triage