Operationalizing AI ethics in medicine-a co-creation workshop study.

backgroundA majority of AI ethics frameworks focus on high-level principles but lack actionable guidance. Effectively implementing AI in projects requires the operationalization of AI ethics, translating principles into requirements. This paper proposes a novel method for operationalizing AI ethics through co-creation workshops.

methodsThe study adopted a qualitative, participatory research approach. Stakeholders from the VALIDATE project, a Horizon Europe action developing an AI-based clinical decision support system for stroke patient stratification, were divided into five diverse teams. The workshops aimed to: (i) determine the low-level ethical requirements stakeholders considered essential for the project, and (ii) identify requirements not included in the EU Trustworthy AI Guidelines. The methodology included storytelling, content analysis to identify ethical issues and dilemmas, eliciting quantifiable requirements using a standardized planning language (Planguage), and a procedural feedback process.

resultsThe workshops identified explainability, privacy, model robustness, model validity, epistemic authority, fairness, and transparency as key ethical issues. Participants drafted low-level requirements related to privacy, explainability, transparency, and validity. Six issues (time sensitivity, validity, prevention of harm to patients, patient-inclusive care, quality of life, and lawsuit prevention) could not be mapped to the EU Guidelines. Participants did not draft requirements in relation to the latter issues. Challenges included the diverse interpretations of concepts, such as validity. Participants generally had favorable impressions of the workshops, although they found formulating requirements in Planguage format more challenging than storytelling and topic prioritization.

conclusionsThe workshops elicited concrete, quantifiable, and actionable requirements, which were useful in developing a project-specific ethical framework. The proposed methodology is resource-efficient and requires fewer AI ethics experts than existing methods while remaining compatible with established guidelines. Procedural feedback results suggest that participants and facilitators would benefit from additional training in the use of Planguage. Potential challenges included the impact of power dynamics among participants on discussions, blind spots due to overlooked issues, and the absence of stroke patients among the participants.