JMIR Publications

ABSTRACT

Background:

Artificial Intelligence-driven clinical decision support systems (AI-CDSS) are increasingly introduced to enhance clinical decision-making, yet their adoption in healthcare remains limited. Research tends to explain this limited uptake in primarily technical or operational terms, overlooking how social and cultural characteristics shape how AI is interpreted and used.

Objective:

This study aims to examine how clinicians and AI experts make sense of AI-CDSS in clinical practice by introducing feminist and constructivist Science and Technology Studies (STS) concepts to the analysis of qualitative interview data. We explore how sociotechnical conditions influence perceptions of understanding, value, accountability, and use of AI-CDSS beyond informational or technical requirements.

Methods:

We conducted a reanalysis of semi-structured interviews conducted in 2024 with clinicians (N=8) and AI experts (N=8) in the Netherlands. The original study focused on their informational needs for AI-CDSS. The current paper examines how participants constructed knowledge, identified relevant information, and determined when AI was considered meaningful and safe to use. We align themes with concepts from STS, which we use as interpretive lenses and organizing structures to understand the complex relations observed.

Results:

We connect five empirical themes to five STS concepts. Situated knowledge shows that informational needs of clinicians are not universal but relate to their roles and specific clinical contexts, with understanding oriented less toward technical transparency than toward actionable guidance, relevance, and integration into decision-making. Standpoint theory illuminates differences between stakeholder priorities: clinicians emphasized information on how to use and act on AI outputs within existing workflows, whereas AI experts focused more on providing technical characteristics and performance metrics. Tinkering captures how clinicians experimented with AI systems in practice to develop situated, lay forms of expertise. Situated value highlights that perceptions of an AI-CDSS’s worth are relational, with clinicians valuing novel, context-sensitive insights and developers prioritizing performance metrics. Finally, response-ability reveals how trust and accountability were largely delegated to institutional structures, while clinicians reported limited legal, organizational, and structural support to respond meaningfully to uncertainty or errors when using AI-CDSS.

Conclusions:

Our findings highlight that adoption of AI in healthcare cannot be reduced to technical transparency or communication strategies, but requires attention to situated practices, values, and responsibilities of clinicians and AI experts. Processes and considerations of how clinicians make sense of AI-CDSS are not universal, but shaped by social, cultural, and institutional contexts. This highlights limitations of purely technical or Human Computer Interaction (HCI)-focused approaches and underscores the need for context-sensitive, multi-stakeholder strategies to support meaningful and responsible AI adoption in clinical practice.