人工智能在狼疮妊娠定性研究中的应用:大语言模型与提示策略的潜力
本研究评估不同大语言模型(LLM)翻译和分析系统性红斑狼疮(SLE)女性妊娠经历定性访谈的能力,并比较LLM与人类研究者识别的主题。对26份葡萄牙语访谈使用三种提示策略和四种LLM生成主题,结果显示LLM生成的主题与人类研究者高度一致,few-shot提示表现最佳,Claude Sonnet 4.5最稳定。
要点
- LLM生成的主题与人类研究者识别的主题高度一致,主题统一性和语义相似性高。
- Few-shot提示始终优于zero-shot和思维链提示,主题提取最全面连贯。
- LLM辅助定性分析可补充人工解读,支持非英语数据和多元患者视角。
核心结论
- LLM可作为定性研究的支持工具,保持主题连贯性并捕捉SLE女性妊娠经历的复杂性。
- Few-shot提示在本研究中实现了最全面和连贯的主题提取,而zero-shot提示主题结果较低。
研究设计与样本
- 分析26份由Rodrigues等人最初以葡萄牙语进行的半结构化访谈,对象为SLE女性,探讨围产期护理经历。
- 使用四种先进LLM(GPT-4o、Claude Sonnet 4.5、Gemini 2.5 Flash、DeepSeek R)和三种提示策略(zero-shot、few-shot、思维链)生成关键主题。
- 采用LLM-as-a-judge框架评估主题覆盖度、主题统一性和语义相似性。
主要发现
- 跨模型和提示策略,LLM生成的主题相似,且与人类研究者识别的主题密切对齐。
- LLM-as-a-judge评估显示高主题统一性和语义相似性。
- Claude Sonnet 4.5表现最一致。
- Few-shot提示始终优于zero-shot和思维链方法。
临床怎么用
- 在定性研究中可考虑使用few-shot提示策略以获取更全面连贯的主题提取。
- LLM辅助分析可作为人类解读的补充,并有助于纳入非英语语言数据和多元患者视角。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: Our objective was to evaluate the ability of different large language models (LLMs) to translate and analyze qualitative interviews describing pregnancy experiences among women with systemic lupus erythematosus (SLE). An additional objective was to compare themes identified by LLMs with those identified by human researchers. METHODS: We analyzed 26 semi-structured interviews originally conducted in Portuguese by Rodrigues et al. among women with SLE exploring their perinatal care experiences. Three prompting strategies (zero-shot, few-shot, and chain-of-thought) were used on the transcripts across four state-of-the-art LLMs (GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash, and DeepSeek R) to generate key themes. Model outputs were evaluated using an LLM-as-a-judge framework assessing thematic coverage, thematic unity, and semantic similarity. RESULTS: Across models and prompting strategies, LLMs generated similar themes that closely aligned with those identified by human researchers. LLM-as-a-judge evaluations demonstrated high thematic unity and semantic similarity. Claude Sonnet 4.5 showed the most consistent performance. Few-shot prompting consistently outperformed zero-shot and chain-of-thought approaches. CONCLUSION: LLMs show promise as supportive tools for analyzing qualitative research by preserving thematic coherence and capturing the complexity of pregnancy experiences in women with SLE. Although zero-shot prompting is widely used, it produced lower thematic results in our study, whereas few-shot prompting achieved the most comprehensive and coherent theme extraction. LLM-assisted qualitative analysis may complement human interpretation and has the potential to support the inclusion of non-English language data and diverse patient perspectives in lupus research.
引用
- Boers JRM, Terry G, Rodrigues L, et al. Using Artificial Intelligence in Qualitative Lupus Pregnancy Research: The Potential of Large Language Models and Prompting Strategies. Arthritis Care Res (Hoboken). 2026. PMID: 42901873. doi:10.1002/acr.80193
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。