Deductive Coding of Psychosocial Autopsy Data Using LLM

This research article investigates the integration of a Large Language Model (LLM) with qualitative research procedures, specifically focusing on deductively coding and summarizing interview data from psychosocial autopsy studies. The study analyzes data from 38 semi-structured interviews with individuals bereaved by suicide, assessing the LLM's performance in three tasks: binary classification of coded segments, independent classification using a sliding window approach, and summarization of coded data. Results indicate substantial agreement between the LLM and qualitative researchers, with accuracy rates of 0.84 for binary classification and 0.67 for the sliding window task. The LLM's summaries were generally rated as adequate or good by independent researchers. The findings suggest that LLMs can enhance the efficiency of qualitative research by reducing time and resource investment while maintaining quality. The study recommends a collaborative model where LLM coding is complemented by human review and interpretation, highlighting the potential for real-time monitoring in public health contexts.