本研究旨在因應臨床護理溝通訓練中反覆練習機會不足、情境變化有限，以及新進護理師與護理學生因臨床經驗不足而易產生焦慮等問題，建構一套生成式人工智慧驅動之多角色AI溝通訓練系統〈醫言難盡〉。本系統以情境式學習與醫護教育遊戲概念為基礎，設計具關卡任務、角色互動、即時回饋與重複挑戰特性的訓練活動，使參與者能在較安全且可反覆練習的環境中，模擬真實臨床病房中的溝通流程。 〈醫言難盡〉以生成式人工智慧即時生成病人情境，並建構病人、家屬、護理師前輩與醫師四種AI角色。參與者在每輪訓練中需先閱讀病人情境與病歷資訊，再與病人對話以蒐集主訴與疼痛評估相關資訊，過程中亦可能面對家屬插話、情緒反應與資訊干擾；當參與者遇到判斷困難時，可向護理師前輩尋求提示與協助，最後則需向醫師進行病況回報。系統透過醫師考核機制判斷參與者是否完成護理標準作業流程（SOP），包含詢問疼痛評估重點、整合病人資訊與完成醫師回報等任務，並依據通關結果與醫師回饋作為後續學習參考。 本研究採質性分析為主、量化問卷為輔之研究設計。研究活動以多輪次訓練方式進行，參與者於不同關卡中反覆進行情境對話與醫師考核，系統則記錄其對話內容、行為操作、通關結果與關卡進展。本研究進一步分析參與者之通關表現、SOP行為、額外詢問行為、家屬插話應對方式與不同背景參與者之表現差異。此外，本研究亦透過活動前、後焦慮量表分析參與者焦慮變化，並以系統滿意度問卷蒐集其對整體訓練體驗、操作易用性、AI互動、情緒調適與臨床應用期待之回饋。 研究結果發現，參與者之通關表現與SOP行為掌握程度呈現較一致的變化趨勢。隨訓練輪次增加，整體SOP行為占比與通關率皆逐步提升，顯示反覆情境練習有助於參與者熟悉臨床溝通流程。不同背景參與者中，新進護理師與學生組在通關表現與SOP行為掌握上較為明顯；相較之下，額外詢問與家屬應對策略雖能反映個人互動習慣，但與通關表現未呈現穩定一致之關聯，顯示影響任務完成的核心仍以參與者是否掌握並執行SOP流程為主。 此外，研究結果亦顯示，參與者在使用〈醫言難盡〉後，焦慮程度整體呈下降趨勢，且對系統在焦慮調適、臨床溝通練習、AI角色互動與推薦使用意願等面向皆給予正向評價。整體而言，本研究顯示，生成式人工智慧驅動之多角色護理溝通訓練系統可作為護理臨床溝通教育之輔助工具，透過情境生成、角色互動、關卡任務與醫師考核機制，協助參與者熟悉臨床溝通流程、累積情境應對經驗，並降低其面對臨床情境時的不確定感與焦慮。。;This study aimed to address several challenges in clinical nursing communication training， including insufficient opportunities for repeated practice， limited variation in training scenarios， and the anxiety commonly experienced by novice nurses and nursing students due to limited clinical experience. To address these issues， this study developed ， a generative artificial intelligence–driven multi-role AI communication training system. Grounded in situated learning and healthcare education game concepts， the system was designed as a training activity featuring level-based tasks， role interaction， immediate feedback， and repeated challenges. It allows participants to simulate clinical communication processes in a safer environment where repeated practice is possible. uses generative artificial intelligence to generate patient scenarios in real time and constructs four AI roles: patient， family member， senior nurse， and physician. In each round of training， participants first read the patient scenario and medical record information， and then interact with the patient to collect information related to the chief complaint and pain assessment. During the process， they may also encounter family interruptions， emotional responses， and information interference. When participants face difficulties in clinical judgment， they may seek prompts and assistance from the senior nurse. Finally， they are required to report the patient’s condition to the physician. The system uses a physician assessment mechanism to determine whether participants have completed the nursing standard operating procedure (SOP)， including asking key pain assessment questions， integrating patient information， and completing the physician report. The task-passing results and physician feedback then serve as references for subsequent learning. This study adopted a primarily qualitative research design supplemented by quantitative questionnaires. The research activity was conducted through multiple rounds of training， in which participants repeatedly engaged in scenario-based dialogue and physician assessment across different levels. The system recorded their dialogue content， behavioral operations， task-passing results， and level progression. This study further analyzed participants’ task-passing performance， SOP behaviors， additional questioning behaviors， family interruption response strategies， and performance differences among participants with different backgrounds. In addition， pre- and post-activity anxiety scales were used to examine changes in participants’ anxiety， and a system satisfaction questionnaire was administered to collect feedback regarding the overall training experience， ease of use， AI interaction， emotional adjustment， and expectations for clinical application. The results showed that participants’ task-passing performance and mastery of SOP behaviors demonstrated relatively consistent trends. As the number of training rounds increased， both the overall proportion of SOP behaviors and the task-passing rate gradually improved， indicating that repeated scenario-based practice helped participants become familiar with clinical communication processes. Among participants with different backgrounds， the novice nurse and nursing student group showed more noticeable improvement in task-passing performance and SOP behavior mastery. In contrast， although additional questioning and family response strategies reflected individual interaction habits， they did not show a stable and consistent relationship with task-passing performance. This suggests that the core factor influencing task completion was whether participants were able to master and execute the SOP process. Furthermore， the results indicated an overall downward trend in participants’ anxiety after using . Participants also gave positive evaluations of the system in terms of anxiety adjustment， clinical communication practice， AI role interaction， and willingness to recommend its use. Overall， this study suggests that a generative AI–driven multi-role nursing communication training system can serve as a supplementary tool for clinical nursing communication education. Through scenario generation， role interaction， level-based tasks， and a physician assessment mechanism， the system can help participants become familiar with clinical communication processes， accumulate scenario-based response experience， and reduce their uncertainty and anxiety when facing clinical situations.
