[Participatory processes in the development of AI systems in a clinical context: A qualitative secondary analysis]

Participatory processes in the development of AI systems in a clinical context: A qualitative secondary analysis Abstract: Background: Despite technological advances, the implementation of clinical AI systems frequently fails. Participatory approaches are regarded as promising solutions, but empirical evidence on their implementation and effectiveness is lacking. Aim: This study analyses the knowledge potential of participatory methods in the development of clinical AI systems. It examines how these contribute to making profession-specific experiential knowledge, documentation-related data practices visible and support negotiation processes. Methods: A qualitative secondary analysis of participatory activities in the KIDELIR project (development of an AI-based decision support system for delirium prediction) was conducted. Structured, qualitative content analysis was used to analyse photo documentation, observation protocols from workshops (case vignettes, health information mapping, group discussions, participatory system mapping, theory of change modelling), and workshop and interview transcripts. Results: Different formats revealed distinct knowledge forms: case vignettes uncovered intuitive assessment logics, mapping methods exposed structural information flows, and interviews highlighted profession-specific requirements. Not only were functional expectations collected, but also implicit routines, data breaks, and communication problems became visible. Conclusions: Participatory formats serve to gather requirements and reflect on context-specific forms of knowledge. They make profession-specific knowledge usable for development decisions without requiring direct technical translatability.