Generative AI for Qualitative Analysis in a Maternal Health Study: Coding In-depth Interviews using Large Language Models (LLMs)

Study Objectives: In-depth interviews are one of the most widely used approaches for qualitative studies in public health. The coding of transcripts is a critical step for information extraction and preliminary analysis. However, manual coding is often labor-intensive and time-consuming. The emergence of generative artificial intelligence (GenAI), supported by Large Language Models (LLMs), presents new opportunities to understand human languages, which may significantly facilitate the coding process. This study aims to build a computational coding framework that uses GenAI to automatically detect and extract themes from in-depth interview transcripts. Methods: We conducted an experiment using transcripts of in-depth interviews with maternity care providers in South Carolina. We leveraged ChatGPT to perform two tasks automatically: (1) deductive coding, which involves applying a predefined set of codes to dialogues; and (2) inductive coding, which can generate codes from dialogues without any preconceptions or assumptions. We fine-tuned ChatGPT to understand the content of the interview transcripts, enabling it to detect and summarize codes. We then evaluated the performance of the proposed approach by comparing the codes generated by ChatGPT with those generated manually by human coders, involving human-in-the-loop evaluation. Results: The results demonstrated the potential of GenAI in detecting and summarizing codes from in-depth interview transcripts. ChatGPT could be utilized for both deductive and inductive coding processes. The overall accuracy of GenAI is higher than 80% and the codes it generated showed high positive associations with those generated manually. More impressively, GenAI reduced the time required for coding by 81%, demonstrating its efficiency compared to traditional methods. Discussion: GenAI models like ChatGPT show high generalizability, scalability and efficiency in handling large datasets, and are proficient in multi-level semantic structure identification. They demonstrate promising results in qualitative coding, making it a valuable tool for supporting people in public health research. However, challenges such as inaccuracy, systematic biases, and privacy concerns must be addressed when using them in practice. GenAI-based coding results should be handled with caution and reviewed by human coders to ensure accuracy and reliability.