Multimodule Human-Artificial Intelligence Collaboration Pipeline for Large Language Model-Assisted Thematic Analysis...

Qualitative thematic analysis is widely used in health research to examine patient experiences and inform the refinement of digital health interventions, but it is time- and labor-intensive. Large language models (LLMs) may help accelerate this process, yet their performance may depend not only on the model itself but also on how the analytic workflow is structured. Current evidence remains limited on how different LLMs perform across multistage thematic analysis workflows and across multiple health-related qualitative datasets.

This study aimed to evaluate a modular human-artificial intelligence (AI) collaboration pipeline for LLM-assisted thematic analysis and compare how model choice and workflow strategy influence alignment between AI-generated and human-generated themes across 3 qualitative health studies.

The framework was applied to analyze deidentified semistructured interview transcripts from 3 completed qualitative health studies involving patients with interstitial lung disease, postural orthostatic tachycardia syndrome, and chronic obstructive pulmonary disease. Three LLMs were compared: Gemini (Gemini 3 Pro), ChatGPT (GPT-5.2-thinking), and Opus (version 4.6). The workflow separated analysis into code extraction, code combination, and theme generation, and 5 strategies were tested. AI-generated themes were embedded using sentence-t5-xxl and compared with human-generated themes using cosine similarity after alignment with Hungarian and Greedy matching. Runtime and output-format consistency were also examined.

Output volume differed substantially by model. Gemini generated the fewest codes and themes, while ChatGPT showed a similar but higher output ceiling. Opus produced the largest and most variable codebooks and theme sets. Across the 3 studies, Opus showed the strongest and most consistent alignment with human-generated themes, with the best cosine similarity scores observed in postural orthostatic tachycardia syndrome-direct coding (mean 0.893, SD 0.041), chronic obstructive pulmonary disease-direct grouping (mean 0.891, SD 0.027), and interstitial lung disease-L3 (mean 0.889, SD 0.032). ChatGPT was competitive in selected settings, whereas Gemini generally produced slightly lower similarity scores but had the shortest runtime. ChatGPT and Opus also showed better formatting consistency and workflow usability than Gemini.

A modular human-AI pipeline can support thematic analysis across multiple digital health interview studies, but performance depends strongly on both model choice and workflow design. Opus produced the most consistently human-aligned themes, while Gemini and ChatGPT showed different trade-offs in speed, fidelity, and usability. These findings support the use of LLMs as structured, human-supervised analytic assistants rather than replacements for qualitative researchers.