Abstract
Artificial intelligence (AI) is gradually transforming healthcare. However, despite its promised benefits, AI in healthcare also raises a number of ethical, legal and social concerns. Compliance by design (CbD) has been proposed as one way of addressing some of these concerns. In the context of healthcare, CbD efforts could focus on building compliance with existing clinical guidelines (CGs), given that they provide the best practices identified according to evidence-based medicine. In this paper we use the example of AI-based clinical decision support systems (CDSS) to theoretically examine whether medical AI tools could be designed to be inherently compliant with CGs, and implication for ethics and trust. We argue that AI-based CDSS systematically complying with CGs when applied to specific patient cases are not desirable, as CGs, despite their usefulness in guiding medical decision-making, are only recommendations on how to diagnose and treat medical conditions. We thus propose a new understanding of CbD for CGs as a sociotechnical program supported by AI that applies to the whole clinical decision-making process rather than just understanding CbD for CGs as a process located only within the AI tool. This implies taking into account emerging knowledge from actual clinical practices to put CGs in perspective, reflexivity from users regarding the information needed for decision-making, as well as a shift in the design culture, from AI as a stand-alone tool to AI as an in-situ service located within particular healthcare settings.