ChatGPT應用於質性研究: 以台灣個案管理中心癌症篩檢服務模式與社區新冠疫情為例

Abstract

質性研究為了解事件脈絡建立理論架構發展後續科學評估與介入之基礎。過往質性研究以訪談文稿運用沉浸式文本閱讀萃取訪談事件關鍵字與編碼建立受訪者對於訪談主題及相關陳述，並於彙整主題後形成理論基礎。人工智慧大型語言模型發展起始於自然語言處理技術(Natural Language Processing，NLP)，由初始之字詞庫對照(Bag of Word)與n-gram 語詞分節摘取與字詞預測技術發展至詞頻權重(Term frequency-Inverse Document frequency，TF-IDF)矩陣化技術提高運算效率，並應用文字語詞語矩陣(Word2Vec及Term2Vec)技術結合遞迴神經網絡(Recurrent Neural Network，RNN)整合多種深度學習技術預訓練(Pre-train)與微調(Fine tune)發展成為目前應用如ChatGPT之大型語言模型(Large language model，LLM)。此過程亦與質性研究由訪談中萃取關鍵字詞形成脈絡思維並行，因此，使用LLM模式進行訪談應當非常有效率，而且可以節省時間及人力。 本研究運用大型語言模型ChatGPT進行質性研究包含訪談摘要生成以及關鍵字萃取、主題結構形成。並且以ChatGPT所萃取之關鍵字與文本主題進行包含情感分析、因素分析，與因素路徑分析。情感分析包含主題分佈分析，評估受訪者對於訪談主題之情感傾向語對於主題結構。所萃取之關鍵字則以TF-IDF形成詞頻權重矩陣進行因素分析，量化主要因素包含之關鍵字並形成因素路徑。 所使用質性研究評估之訪談資料為台灣癌症篩檢個案管理中心訪談以及某社區里長及工作人員新冠疫情防疫應變訪談。台灣癌症篩檢個案管理中心訪談為運用衛生福利部國民健康署 2014 年「癌症篩檢個案管理中心輔導計畫」3位訪談者於全台19處個案管理中心對57位管理中心成員以衛生計畫與評價循環模式設計半結構式問卷建立訪談逐字稿資料集。社區里長及工作人員新冠疫情防疫應變訪談則為2024年由社區免疫服務對4個社區里長以及新冠疫情期間里工作人員進行質性訪談形成之逐字稿。 研究以ChatGPT對所建立之訪談文稿資料集運用一致之提示詞包含訪談文稿摘要生成提示，以及由訪談文稿摘錄之主題，各主題摘取關鍵字編碼個數並彙整成資料表形成訪談逐字稿摘要以及萃取關鍵字、開放編碼與主題編碼數位資料庫。運用所摘取之關鍵字進行因素路徑分析。 癌症個案管理中心訪談文稿萃取關鍵字情感分析(Sentiment analysis)於19處縣市衛生局皆以正向情感高於負向情感。主題形成個案管理經歷與能力培育、個案管理中心規劃推行、工作規劃、跨部門資源整合服務合作、癌症篩檢與個案管理成效目標，以及民眾認知溝通與教育六大類面向，其中以個案管理經歷與能力培育以及個案管理中心規劃推行為訪談中最頻繁提及之主題類別。依據受訪談管理中心所屬區域(北、中、南、東)主題亦以此兩大類別主題出現頻次最高，部分縣市(縣市C、縣市J，以及縣市R)則以個案管理工作規畫為主要主題。跨部門資源整合服務合作則較頻繁出現於縣市G與縣市S訪談中。因素路徑分析顯示癌症篩檢個案管理訪談可形成人員與工作能力培養(因素1)、資源整合(因素2)、衛生局政策與目標(因素3)、困難個案管理實務(因素4)、陽性個案追蹤(因素5)，以及民眾溝通(因素6)。北、中、南、東區癌症篩檢個案管理中新因素路徑以全國為主，並呈現各區特性因素。 社區里長與工作人員新冠疫情訪談萃取關鍵字進行情感分析於該社區以及四里皆以負向情感高於正向情感。訪談中提及之主題包含心理支持與調適、防疫措施推行、里長職責與社區互動、防疫資源整合、疫苗疑慮，以及政治與防疫影響六項主要主題類別。心理調適與支持及防疫措施推行相關主題在四里皆為訪談出現頻次第一或第二位。因素路徑分析顯示社區里長與工作人員新冠疫情訪談形成里長職責與民眾溝通(因素1)、抗體檢測協調(因素2)、心理支持與壓力(因素3)、防疫措施與封城(因素4)、政治與疫情(因素5)、防疫物資與社區衝擊(因素6)。 本研究運用大型語言模型ChatGPT對質性訪談逐字稿進行摘要以及文本分析，萃取關鍵字以及主題編碼，並建構影響因素網絡。本論文所提供ChatGPT應用於質性研究之脈絡評估架構可以做為未來質性研究分析LLM模式使用之參考。

Qualitative research aims to construct frameworks and networks of relevant factors for better understanding events through observations like interviews. This approach helps plan and administer evaluation and intervention. Conventionally, it relies on interview transcripts and well-trained personnel to extract keywords, coding, and themes, following a rigorous process to build perspectives from interview statements, supporting theory formation. The development of large language models (LLMs) began with natural language processing (NLP) techniques, evolving from bag-of-words and n-grams to TF-IDF, and then to word matrices like Word2Vec for better computational efficiency. Recurrent neural networks (RNNs) further advanced NLP by improving sentence prediction. The integration of deep learning techniques for pre-training and fine-tuning led to the evolution of LLMs like ChatGPT. This progression parallels qualitative research, wherein keyword extraction and theme formation from unstructured transcripts support theory development. The qualitative research data were derived from interviews with staff members of the Taiwan Cancer Screening Center and interviews with community leaders and staff members from a community in Taiwan. The Taiwan Cancer Screening Case Management Centers Counseling project involved three interviewers who conducted semi-structured interviews for 57 members from 19 centers across Taiwan, resulting in a verbatim transcript dataset based on a semi-structured questionnaire. The COVID-19 pandemic response interviews were conducted in 2024 by the community COVID-19 services project, forming verbatim transcripts from qualitative interviews with leaders and staff from four communities in Taiwan. This study used ChatGPT for qualitative research tasks including interview summarization, keyword extraction, and theme coding. The extracted keywords were utilized for sentiment analysis, factor analysis, and pathway analysis. Sentiment analysis evaluated emotional tendencies in the interviews, while factor analysis quantified the keywords and connected them through pathway analysis. Sentiment analysis of keywords from the cancer screening case management center interviews shows higher positive sentiments across all 19 health bureaus. Key themes include case management experience and capacity building, project planning and administration, work arrangement, interdisciplinary resource integration, goals and effectiveness of cancer screening, and public awareness. The themes most frequently mentioned include case management experience and capacity building. Project planning and administration were also commonly noted in County C, County J, and County R. Interdisciplinary resource integration was notably mentioned in County G and County S. Pathway analysis identified six factors: personnel and work capacity building, resource integration, health bureau policies and goals, difficult case management, positive case management and tracking, and public communication. These patterns were consistent across North, Central, South, and East regions, with the keywords and factors supporting each theme reflecting regional characteristics. Sentiment analysis of the community leaders and staff for COVID-19 interviews reveals higher negative sentiments across all four communities. Key themes include psychological support and coping strategies, epidemic prevention measures, community leader obligations, public communication, resource integration, vaccine misinformation, and political impacts on prevention strategies. The most frequently mentioned themes are psychological adjustment and epidemic prevention measures. Pathway analysis identified six factors: community leader responsibilities and communication, antibody testing coordination, psychological support, epidemic measures and lockdowns, politics, and the impact of preventive supplies. This study applied ChatGPT to summarize qualitative interview transcripts and perform text analysis, extracting keywords and coding themes, constructing an influencing factor network. This study also provides a framework for assessing the context of qualitative research using ChatGPT.