Evaluation of Prompt Design and Internal Reasoning in Chatbot-Based Medical History Taking: Simulation Study

Abstract Background A persistent discrepancy exists between patient-reported information and physician documentation. While conversational agents have been developed to collect medical histories prior to consultations, existing evaluations have largely focused on diagnostic accuracy or user satisfaction rather than on the completeness and clinical relevance of the information collected. There remains a need to assess the extent to which clinically relevant information is captured through chatbot-based interviews, and to understand how model configurations and instructional strategies influence this coverage. Objective This study aimed to evaluate the extent to which a chatbot can obtain clinically useful patient history information, and to examine how prompt detail and internal reasoning influence information coverage during chatbot-based medical interviews. Methods We developed a medical history-taking chatbot using the Qwen3-14B-Instruct model and evaluated 4 configurations in a 2×2 factorial design: detailed and thinking mode, detailed and nonthinking mode, minimal and thinking mode, and minimal and nonthinking mode. These configurations were compared against a rule-based system baseline (choice mode) using 66 standardized primary care clinical cases, with simulated patients interacting with the chatbot according to predefined case scripts. Information coverage (%) was assessed using a checklist inspired by Objective Structured Clinical Examination (OSCE) frameworks. Three physicians independently evaluated transcript coverage, with interrater agreement assessed using full agreement rates and Fleiss κ. For the 4 LLM configurations, coverage was analyzed using 2-way repeated-measures ANOVA to examine the effects of prompt detail, internal reasoning, and their interaction. All 5 configurations, including the rule-based baseline, were additionally compared using 1-way repeated-measures ANOVA with post hoc 2-tailed paired t tests. Results Interrater agreement was substantial (Fleiss κ=0.75). Across all 66 simulated cases, information coverage differed significantly among configurations ( P