Foundation Models in Healthcare: Opportunities, Biases and Regulatory Prospects in Europe

This article concerns the rise of a new paradigm in AI - “foundation models,” which are pre-trained on broad data at scale and subsequently adapted to particular downstream tasks. In particular, it explores the issue from the perspective of healthcare and biomedicine, focusing on the benefits of foundation models, as well as their propensity to encode bias, which threatens to exacerbate discriminatory practices already experienced by patients in Europe. Section 1 offers a brief introduction concerning the use of AI in healthcare and biomedicine and the problem of divergencies in access to and quality of healthcare across Europe. Section 2 familiarises the reader with the technical qualities of foundation models and recent developments in the field. Section 3 explains how the new health data strategy proposed by the EU could foster the development of foundation models in healthcare. Section 4 elaborates on their benefits in healthcare and biomedicine, while Sect. 5 explores the risk of bias exhibited by foundation models. Section 6 comments on the uncertain status of foundation models under the proposed Artificial Intelligence Act and offers brief recommendations concerning future regulation. Section 7 concludes.