Understanding pre-training data effects in retinal foundation models using two large fundus cohorts

Abstract

Abstract Medical foundation models, pre-trained on large-scale unlabelled data, show strong performance and data efficiency when adapted to various clinically relevant applications. However, how pre-training data shape the generalisability and fairness of these models remains unexplored. Here we address this using two cohorts from Moorfields Eye Hospital (UK) and the Shanghai Diabetes Prevention Program (China), each containing 904,170 fundus photographs for model pre-training. Using identical pipelines, we train parallel foundation models using individual cohort and evaluate them on downstream tasks with publicly available datasets and held-out data from each site. The parallel models show competitive performance to data that differ substantially from their pre-training data. Nevertheless, we observe fairness gaps over age subgroups, whereas sex and ethnicity show minimal impact. These results demonstrate the good generalisability of retinal foundation models and indicate that pre-training demographic attributes shape fairness differently, highlighting the importance of domain-specific, fine-grained data curation for efficient foundation model development.