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

Area of Science:

• Artificial Intelligence in Medicine
• Ophthalmology
• Medical Imaging Analysis

Background:

• Medical foundation models, pre-trained on large datasets, offer efficiency in clinical applications.
• The impact of pre-training data on the generalizability and fairness of these models is not well understood.

Purpose of the Study:

• To investigate how pre-training data characteristics influence the generalizability and fairness of medical foundation models.
• To evaluate the performance and fairness of retinal foundation models trained on distinct large-scale cohorts.

Main Methods:

• Trained parallel foundation models on two distinct cohorts (Moorfields Eye Hospital and Shanghai Diabetes Prevention Program) comprising 904,170 fundus photographs each.
• Evaluated model performance and fairness on downstream tasks using public datasets and held-out site-specific data.
• Assessed fairness across age, sex, and ethnicity subgroups.

Main Results:

• Retinal foundation models demonstrated competitive performance even when evaluated on data dissimilar to their pre-training sets, indicating strong generalizability.
• Fairness gaps were observed across different age subgroups.
• Sex and ethnicity subgroups showed minimal impact on model fairness.

Conclusions:

• Foundation models for retinal imaging exhibit good generalizability across diverse datasets.
• Pre-training data demographics significantly shape model fairness, particularly concerning age.
• Domain-specific, fine-grained data curation is crucial for developing equitable and efficient medical foundation models.