Exploring AI Explainability: A Multidimensional Approach to XAI in Healthcare. Understanding heterogeneous perspectives of hospital stakeholder groups.

Sammanfattning

Artificial intelligence (AI) is significantly changing the way companies work. Implementing the most suitable AI systems is a critical and complex task that demands careful consideration of various criteria. Key among these considerations is the concept of AI explainability (XAI), which facilitates users' understanding of AI-based decision-making. The extant literature largely treats XAI as a technical-oriented concept; only a few studies acknowledge that users may have different perspectives and needs. However, how users' perspectives differ and why has not been theorized in literature. For whom is XAI important? What are the different users' perceived purposes of XAI? When would the users like XAI to be provided? All these questions need further investigation. The healthcare sector, characterized by its slow digital adoption, heavily regulated environment, and high-risk status under the EU AI Act, is a fruitful research context for further exploring these XAI considerations. Acknowledging the importance of contextual factors of a specific organization, this research further narrows the scope by exploring how XAI resonates among different stakeholder groups at Karolinska University Hospital, Stockholm. This qualitative study explores the heterogeneous needs and perspectives on XAI of six key hospital stakeholder groups: clinicians, management, patients, nurses, research & development, and regulatory. Data was collected through semi-structured interviews with 27 interviewees, each representing one of the identified hospital stakeholder groups. The empirical results indicate significant differences between stakeholder groups regarding when, how, and why to receive XAI. Moreover, it became evident that the attributes of AI trust, AI literacy, medical literacy, and accountability for AI-based decision-making impact a group's perspective on XAI. A conceptual framework was created based on literature, populated, and extended with insights from the interviewees from all stakeholder groups. The study contributes to the existing literature by providing a thorough understanding of the human-centered view on XAI and emphasizing the concept's multidimensionality. It reveals the differences between hospital stakeholder groups regarding their perspectives on XAI and four attributes that have proven to affect these perspectives. By examining how the different manifestations of these attributes impact stakeholder groups' perspectives on XAI, the authors provide valuable insights to practitioners within the field.