Human–AI Collaboration Pipeline for Thematic Analysis: Example of VR-Supported Pulmonary Rehabilitation

Abstract

Thematic analysis of qualitative interviews is a time-consuming iterative process aimed at summarizing content of multiple interviews. Large Language Models (LLM) have been shown to be effective in summarizing textual content. We used results of a previously completed thematic analysis of qualitative interviews of patients who provided feedback on VR-supported pulmonary rehabilitation for comparison with the results of LLM-assisted thematic analysis. Pulmonary rehabilitation (PR) improves outcomes in COPD but remains underused; virtual-reality (VR)–enabled, home-based models may expand access and engagement. We evaluated a modular human–AI pipeline for thematic analysis of qualitative interviews with nine patients introduced to a VR PR app, assessing how closely LLM–generated themes align with human analysis. A three-agent workflow produced layered codebooks and themes; Sentence-T5-XXL embeddings and cosine similarity quantified theme proximity. Mean cosine similarity between matched human and LLM themes was consistently high. Overall, the engineered pipeline approximates human thematic structure while offering tunable granularity, with a recurring tendency for LLMs to over-abstract pragmatic details.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 1 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.