Multimodule LLM thematic analysis of digital-health interviews: comparative workflow study

Research applicability

Offers a concrete comparison design for researcher-supervised qualitative analysis, preserving code lineage while testing the effects of consolidation and model choice.

Getting started

Editorial, untested guidance: inspect the published module prompts and appendix, then pilot on consented deidentified transcripts in an approved environment. Retain source-code links, human coding review and explicit evaluation criteria; do not treat similarity alone as validity.

Maintenance

Article, not a maintained software-service claim.

License

Source reuse rights must be checked separately; no open licence admission requirement.

Workflow & human oversight

Researcher-supervised process; automation level intentionally unspecified. No autonomous research or clinical deployment asserted.

Workflow level is not an evidence score; higher is not better evidence.

Approach

A comparative study separates LLM-assisted health-interview analysis into code extraction, code combination and theme generation in independent conversations. It compares ChatGPT GPT-5.2-thinking, Opus 4.6 and Gemini 3 Pro across three datasets and five workflow strategies, using embedding-based alignment with human-generated themes.

Limitations

Partial publisher paper-excerpt review: extraction backend cut the body at about 50,000 characters; figures, supplements, discussion tail and underlying transcripts were not fully reviewed. Authors ran each model-workflow condition once through consumer web interfaces. Semantic similarity and valid code identifiers do not establish qualitative interpretive correctness. No independent execution or reproduction; no universal model-ranking claim. The general interview-method section describes one-hour in-person visits, while the Table 4 footnote describes COPD remote interviews after a 12-month programme. This cohort-context difference is unresolved; no uniform interview-setting claim is made. The cited prior framework has unresolved bibliographic identity and is not assumed to alias another resource.

Reproducibility

No code, data analysis or model outcomes reproduced in this review.

Adoption requirements

Approved access to interview data and LLM platforms, qualitative expertise and privacy/consent controls; manual organization between conversations. Authors used paid personal subscriptions, but that is a reported study configuration, not a recommendation to upload sensitive data to consumer accounts.