Automated Thematic Analysis for Clinical Qualitative Data: Iterative Codebook Refinement with Full Provenance

A multi-institutional research team has published a new framework that automates thematic analysis (TA) for clinical qualitative data using large language models (LLMs). The system addresses critical limitations in existing automated TA approaches, which often produce codebooks with poor generalizability and lack analytic auditability. By combining iterative codebook refinement with comprehensive provenance tracking, the framework ensures both quality and reproducibility. When evaluated across five diverse corpora—including clinical interviews, social media data, and public transcripts—it achieved the highest composite quality score on four datasets compared to six baseline methods.

The iterative refinement process yielded statistically significant improvements with large effect sizes on four of the five datasets. These gains were primarily driven by enhanced code reusability and distributional consistency while maintaining descriptive quality. Most notably, on two pediatric cardiology clinical corpora, the AI-generated themes demonstrated strong alignment with themes previously annotated by human experts. The framework represents a substantial advancement for health informatics, offering researchers a scalable, reproducible method to extract meaningful patterns from patient narratives and interview data while maintaining the audit trail necessary for rigorous scientific review.