Abstract:

Objective To construct a nursing evidence dataset，develop a generative artificial intelligence-based nursing evidence question-answering model，and evaluate its usability，with the aim of providing real-time decision support for clinical nurses. Methods Domestic and international nursing literature was comprehensively retrieved and classified according to the “5S” evidence pyramid. A training dataset was constructed，comprising more than 6，000 evidence summaries，over 15，000 guidelines and expert consensuses，and more than 500，000 systematic reviews. A domestically developed open-source large language model with strong performance was selected as the base model. Based on the training dataset，over 800，000 instruction fine-tuning sample pairs and more than 700，000 nursing evidence were generated. Supervised fine-tuning and retrieval-augmented generation were employed to train and develop a nursing evidence question-answering model. In September 2025，23 clinical nurses were recruited via convenience sampling to test the model. The usability of the model was evaluated using a self-developed model usability evaluation questionnaire. A lower score indicates better usability. Results All 23 nurses completed the model trial. In the model usability evaluation questionnaire，the scores were as follows：model quality dimension（1.99±0.85） points，information quality dimension（2.59±1.03） points，interface quality dimension（2.28±1.25） points，user satisfaction dimension（2.76±1.00） points，intention to use dimension（2.19±1.32） points，and net benefits dimension（2.03±0.78） points. Conclusion The overall usability of the nursing evidence question-answering model was good，and its use may help clinical nurses obtain evidence-based nursing information more efficiently. Future work could further expand the model’s functions and enrich the output information.

Key words: Evidence-Based Nursing, Generative Artificial Intelligence, Question-Answering Model, Usability Test

摘要：

目的 构建护理证据数据集，训练并开发基于生成式人工智能的护理证据智慧问答模型，并评价模型的可用性，以期为临床护士提供即时决策支持。方法 全面检索国内外护理领域的文献资料，根据“5S”证据金字塔模型进行分类，构建包含6 000余份证据总结、1.5万余份指南及专家共识、50余万份系统评价等资料的训练数据集。选取国内开源性能较好的大语言模型作为基座模型，依托训练数据集构建了80余万条大语言模型指令微调样本对和70余万条内容的护理证据向量知识库，采用监督微调和检索增强生成技术，训练并开发护理证据智慧问答模型。2025年9月，便利选取23名临床护士试用该模型。采用自行编制的模型可用性评价问卷评价模型的可用性，得分越低表明可用性越好。结果 23名护士均完成了模型的试用，模型可用性评价问卷中模型质量维度得分为（1.99±0.85）分、信息质量维度为（2.59±1.03）分、界面质量维度为（2.28±1.25）分、用户满意维度为（2.76±1.00）分、使用意愿维度为（2.19±1.32）分、净收益维度为（2.03±0.78）分。结论 该护理证据智慧问答模型的整体可用性良好，使用该模型可帮助临床护士更高效地获取循证护理证据。未来可进一步拓展该模型的功能、丰富输出信息。

关键词: 循证护理, 生成式人工智能, 问答模型, 可用性评价