Building the world’s first truly global medical foundation model

Abstract

Since 2022, the field of medical artificial intelligence (AI) has begun a shift toward foundation models, machine learning systems that are trained on broad data at scale and are adaptable to a wide range of downstream tasks^1,2. Medical foundation models are rapidly evolving, driven by the synergy of expanding medical data repositories, advances in neural network architecture (especially transformers), self-supervised learning approaches and computing power. Medical foundation models are capable of performing, or can be adapted to perform, a range of medical tasks with a minimal amount of annotated data. To date, several promising breakthroughs with foundation models have been demonstrated across diverse medical domains, including pathology, radiology and ophthalmology^3,4,5.

Nevertheless, training robust medical foundation models requires large, diverse and clinically useful representative data⁶. Assembling such datasets remains a major challenge for the research community because of strict data-sharing regulations intended to protect patient privacy and ensure ethical compliance. For these reasons, most existing foundational models are trained on datasets that are geographically and demographically ‘narrow’ (that is, not globally representative), limiting their generalizability and effectiveness, particularly in under-represented regions and populations^6,7,8.