Presentation (CTI)

OBJECTIVES: Attribute development for discrete choice experiments (DCEs) is traditionally resource-intensive and context-specific, relying on manual qualitative synthesis of literature and expert input. This study evaluated an AI-supported attribute mining approach across three disease areas to assess its replicability, validity, and limitations relative to conventional qualitative methods.
METHODS: An AI-based text mining and clustering pipeline was applied to published qualitative studies, patient-reported outcomes literature, and HTA documents across three disease areas. Candidate attributes and levels were extracted, normalised, and grouped thematically. Outputs were compared against attributes derived through traditional qualitative research for the same diseases. Validation focused on (1) attribute overlap and coverage, (2) conceptual consistency with established DCE frameworks, and (3) disease-specific versus cross-cutting attributes. Expert review was used to assess face validity and identify false positives or missing concepts.
RESULTS: Across the three disease areas, the AI-supported approach replicated a majority of core attributes identified through traditional methods, particularly those related to treatment effectiveness, safety, and burden of administration. Cross-disease commonalities were consistently identified, while disease-specific attributes showed greater variability in precision and required expert refinement. The AI approach generated additional candidate attributes not previously prioritised, some of which were deemed conceptually relevant, while others reflected semantic noise or insufficient contextual grounding. Replication performance was highest where source literature was rich and well-structured.
CONCLUSIONS: AI-supported attribute mining can efficiently replicate core DCE attribute structures across disease areas and serve as a systematic starting point for attribute development. However, expert validation remains essential, particularly for disease-specific concepts and level refinement. These findings support AI as an assistive tool in DCE design rather than a replacement for qualitative research.