Presentation (CTI)

OBJECTIVES: Qualitative interview data collected during clinical trials are often analysed for a predefined purpose. Secondary research questions and patient-centered themes may arise following results. We evaluated whether a large language model (LLM) pipeline could enable rapid re-analysis of existing trial interview transcripts to characterise specific health-related quality-of-life (HRQoL) themes the original analysis did not target, and to recover patient-level data enabling within-patient comparisons originally unaccounted for.
METHODS: The process started with per-patient structured extraction by an LLM custom agent, contextualised to the trial design, followed by quantitative aggregation in R. Double-transcript patients enabled within-patient contrasts between placebo and active-drug periods, the central advance over the period-stratified original analysis. Outputs were one pipe-delimited row per patient for tabulated aggregation. One prompt specified explicit field numbers and content, conditional formatting, and active self-checks. Extraction ran within a secure, access-controlled enterprise environment, each patient processed in one isolated agent session to preserve accuracy and avoid cross-patient contamination.
RESULTS: Structured HRQoL coding was obtained rapidly relative to manual coding. LLM extraction was sensitive to transcript heterogeneity, with field misalignment the dominant challenge, mitigated but not eliminated by self-checks. Misalignment followed empty data fields: quality control confirmed the explicit prompt eliminated false or hallucinated data, yielding empty fields rather than fabrication. Single-transcript outputs showed less misalignment, likely due to shorter expected output. Quantitative aggregation in R succeeded, the main challenge being parsing and coding of textual data.
CONCLUSIONS: A structured LLM pipeline can facilitate rapid within-patient secondary HRQoL analyses of existing trial interview data within a predefined concept. Reliability depends less on model capability than on explicit prompt engineering and transcript-aware handling. This pipeline can quickly uncover new HRQoL themes and their interplay, informing future study design.