Researchers working on a first medical AI model with globally representative data

A bold new experiment is underway in the world of medical AI. A coalition of over 100 research groups from more than 65 countries has come together to build the first truly global AI foundation model for medicine. Their tool of choice? The retina — specifically, more than 100 million color fundus images collected from nearly every corner of the globe.

As Dr. Yih Chung Tham from NUS Medicine puts it, "Current foundational models are trained on data that is geographically and demographically 'narrow', which limits their effectiveness and can perpetuate existing health inequalities." In other words, if you want AI that serves everyone, you need data from everyone.

How does it work?

The initiative — dubbed Global RETFound — is using a two-track approach to data sharing that's as clever as it is practical:

• Local fine-tuning: Each participating institution can train the AI model on its own data, sharing back only the learned "weights," not patient data.
• Secure central sharing: For institutions without local GPUs or technical expertise, there's a path to upload de-identified images to a secure central hub.

This flexible framework ensures no one is left out — whether they're a top-tier research center in London or a rural clinic in South America. The project builds on RETFound, the first retinal foundation model, trained on 1.6 million images at Moorfields Eye Hospital and UCL. However, this time, the dataset is an order of magnitude larger, encompassing images from Africa, Asia, the Americas, Europe, and beyond.

Pearse Keane, professor at UCL, explains, "This dual approach allows participation from research groups regardless of their resource levels. By combining real and synthetic data generation techniques, we can build a diverse, globally representative dataset without compromising security."

Why does it matter?

Diversity in medical data isn't just a "nice-to-have" — it's a lifesaver. AI models trained on narrow datasets can overlook critical variations across populations, resulting in misdiagnoses or biased results. By incorporating images from every continent (save Antarctica), the Global RETFound model could:

• Improve detection of conditions like diabetic retinopathy, glaucoma, and macular degeneration in populations that have historically been underrepresented.
• Serve as a springboard for research into systemic diseases like cardiovascular conditions.
• Help close the gap in healthcare inequality by making cutting-edge AI tools accessible to low-resource settings.

And here's the kicker: the model will be released under a Creative Commons license, making it freely available for non-commercial research. That's a major win for transparency and global collaboration.

The context

This project didn't spring from nowhere. RETFound, its predecessor, made waves when it was published in Nature in 2023. That model showed just how powerful retinal imaging could be for detecting both eye-specific and systemic diseases. But the team knew they were working with a limited dataset — mostly from the UK.

Global RETFound is the next leap forward, both in ambition and scale. It's not just a model — it's a statement that medical AI should serve the world, not just the wealthy and well-studied. It's a move toward an AI future where innovation doesn't stop at national borders.