Algorithmic Fairness unfolded: an embedded interdisciplinary collaboration within a medical imaging AI lab

Algorithmic fairness (AF) is increasingly recognized as a crucial dimension of artificial intelligence (AI), particularly in healthcare, where it represents both a pressing ethical concern and a practical challenge. Despite a growing body of literature aimed at conceptualizing and operationalizing AF, in practice it is often overlooked or treated as a technical issue of optimization, detached from its ethical and socio-technical dimensions. This paper reports on collaborative efforts on AF within a medical imaging AI lab for lung cancer screening (LCS). Drawing on a two-year collaboration between engineers and ethicists at a Dutch UMC, we analyze a six-month AF assessment project of AI-based LCS risk models and an ethics roundtable workshop. Drawing on Ethics Parallel Research, we reconstruct and critically examine processes shaping AF assessment, as well as challenges encountered in interdisciplinary collaborations. Our analysis frames AF as a ‘wicked problem’ characterized by uncertainty, competing values, and the absence of definitive solutions. It reveals persistent tensions between comprehensive theoretical accounts of AF and the narrower approaches often adopted in engineering practice, shaped by factors such as data and resource limitations, time pressures, competing incentives, unclear responsibilities, and scarce opportunities for transdisciplinarity. Furthermore, we argue that AF frameworks often assume a top-down ‘translation’ model where ethical principles are specified in advance and subsequently implemented, overlooking how fairness is interpreted, negotiated, and reshaped within concrete development contexts. We distinguish between content-oriented and process-oriented challenges and reflect on attempts to bridge AF theory and practice through transdisciplinarity. Our findings suggest that AF is multidimensional and clinically relevant, yet often out of scope in developers’ practice, and that our research systems are not structure to make it a primary design objective. We conclude by arguing for more context-sensitive and transdisciplinary approaches to AF in medical AI.

Co-authors: Mira Vegter, Colin Jacobs, and Marianne Boenink.