Collecting patient-reported outcome measures (PROMs) is essential for clinical care and research, yet traditional form-based approaches are often tedious for patients and burdensome for clinicians. We developed a generative AI conversational agent(CA) using GPT-5 to collect back pain data according to the NIH Task Force's Recommended Minimal Dataset. Unlike prior CAs that ask questions one-by-one, our CA engages users in topic-based conversations, allowing multiple data items to be captured in a single exchange. Through iterative development and pilot testing with clinicians and a consumer panel, we identified key design principles for health data collection CAs. These principles extend established clinical decision support design guidelines to conversational interfaces, addressing: flexibility of interaction style, personality calibration, data quality assurance through confidence visualization, patient safety constraints, and interoperability requirements. We present our prompt design methodology and discuss challenges encountered, including managing conversation length, handling ambiguous responses, and adapting to LLM version changes. Our design principles provide a practical framework for developers creating conversational agents for patient questionnaire completion. The CA is available at https://chatgpt.com/g/g-68f4869548f48191af0544f110ee91c6-backpain-data-collection-assistant (requires ChatGPT registration and subscription for unlimited use).

翻译：收集患者报告结局指标对于临床诊疗与研究至关重要，然而传统的表单式采集方式常使患者感到冗长乏味，并给临床工作者带来沉重负担。我们基于GPT-5开发了一款生成式人工智能对话代理，用于依据美国国立卫生研究院特别工作组推荐的最小数据集采集背痛数据。与既往逐项提问的对话代理不同，本系统通过主题式对话与用户互动，支持在单轮交流中捕获多项数据条目。通过临床医生和消费者小组参与的迭代开发与试点测试，我们提炼出健康数据采集对话代理的关键设计原则。这些原则将成熟的临床决策支持设计准则延伸至对话式界面，涵盖：交互方式的灵活性、人格化校准、通过置信度可视化的数据质量保障、患者安全约束以及互操作性要求。本文展示了提示设计方法论，并探讨了开发过程中遇到的挑战，包括对话长度控制、模糊应答处理以及大语言模型版本适配等问题。本设计原则为开发患者问卷自动填写对话代理提供了实用框架。该对话代理可通过https://chatgpt.com/g/g-68f4869548f48191af0544f110ee91c6-backpain-data-collection-assistant访问（需ChatGPT注册及订阅以获得无限制使用权限）。