Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and generation, leading to growing interest in their use in healthcare. While much of the existing research focuses on diagnostic reasoning or benchmark performance, real-world clinical practice often requires LLMs to serve in more supportive and pragmatic roles. This study investigates the application of LLMs to pre-consultation symptom-taking—a routine yet time-consuming task in clinical workflows. We compare four widely used models: ChatGPT, DeepSeek, ERNIE Bot, and Qwen, in their ability to conduct multi-turn symptom-gathering dialogues and generate structured summaries using simulated pediatric cases from the IMCS21 dataset. This comparative approach enabled us to evaluate overall model performance and identify the communication styles and traits clinicians found most effective. To assess the clinical relevance and usability of these outputs, we conducted a questionnaire study with 61 physicians from diverse specialties. Participants evaluated the LLM-generated dialogues and summaries across key dimensions including relevance, clarity, empathy, and completeness. Results show that ChatGPT consistently outperformed other models, especially in generating coherent, empathetic interactions and clinically useful summaries. Physicians identified logical structure, natural communication, and content completeness as essential for clinical acceptance. While trust in LLMs remains cautious, there is strong interest in adopting such tools under physician oversight. To the best of our knowledge, this is the first study to examine the use of LLMs for pre-consultation symptom-taking and to assess physician perceptions of their practical utility. The findings offer timely guidance for designing LLM-based clinical support systems that prioritize clarity, coherence, and human-centered interaction.