National University of Singapore Launches First Global Medical AI Foundation Model

By Irene Yeh 

December 4, 2025 | Since artificial intelligence (AI) made its way into the medical field, foundation models—machine learning systems that are trained on broad data and adaptable to a wide range of downstream tasks—have become a focus. These foundation models are used to perform various medical tasks with minimal amounts of data. Of course, this means that these models require robust training, which poses a challenge for researchers. These models are usually trained on datasets that are not standardized on a global level, resulting in limitations of the models’ effectiveness.  

To overcome this obstacle, researchers from the Yong Loo Lin School of Medicine, National University of Singapore (NUS Medicine) founded the Global RETFound Consortium, which is led by Dr. Yih Chung Tham, assistant professor at NUS Medicine and an NUS presidential young professor; Pearse A. Keane, professor at Moorfields Eye Hospital and University College London, Institute of Ophthalmology; and Carol Y. Cheung, professor at the Chinese University of Hong Kong, CUHK. 

Developed from the success of RETFound, the first ophthalmic foundation AI model, the Global RETFound Consortium’s initiative aims to build the world’s first truly global and equitable medical foundation model. A position statement of the Consortium was published earlier this fall in Nature Medicine (DOI: 10.1038/s41591-025-03859-5). 

RETFound and Global RETFound 

The RETFound foundation model was initially developed using 1.6 million color fundus photographs (CFPs) and trained to detect eye and systemic diseases. After successfully demonstrating the ability to detect diseases, the team realized that most medical AI models are trained on limited geographic regional data, which inspired the establishment of the Global RETFound Consortium. The Global RETFound model will be a substantial expansion of the RETFound model on over 100 million CFPs from more than 100 study groups across 65 countries. This results in a much more geographically and ethnically diverse, fair, and global representation of data. 

Foundation models learn from very large datasets, typically composed of millions of images. “In simple terms, it is very similar to equipping the foundation model with a deep and extensive knowledge,” explains Tham. Thus, the model can be adapted broadly in new disease detection tasks without requiring large amounts of annotated examples of diseases. This minimizes the costly and tedious process of curating high-quality, human-annotated medical datasets. 

“By learning general patterns and representation from vast data sources, foundation models provide a strong starting point for researchers and clinicians to develop new, task-specific disease detection models more quickly and efficiently,” continues Tham. 

During an exploratory study, the team investigated the optimal real-to-synthetic image ratio for the model. They realized that there are massive computational demands in generating synthetic images because it may take close to a month to generate millions of synthetic images. “When you consider scaling this to our numerous partner organizations, the challenge becomes clear,” says Tham.    

Further development of the foundation model will require a great deal of work. The team aims to identify the optimal balance between efficiency and performance, as well as to ensure a scalable computing infrastructure that can support computational needs efficiently. 

Tackling AI Bias 

One of the biggest concerns that emerges from AI models is bias. AI bias occurs when imbalances in training data cause models to produce skewed results. Data will always have varying degrees of bias, and when an AI model learns from datasets, the development of bias toward certain groups, ideas, and more is always a possibility.  

To prevent bias, Global RETFound prioritizes securing training data from as many countries as possible, including from Africa, the Middle East, Asia, and South America. There is also emphasis on communities that are underrepresented in AI development. “The model will be trained using proportionally balanced data across different demographic groups to minimize bias and ensure equitable performance across all populations,” says Tham.  

Right now, the team is trying to secure CFPs from more than 100 study groups across Asia, Europe, North and South America, and Africa, representing a wide spectrum of ethnic, demographic and geographic diversity, as well as variations in retinal imaging devices.   

According to Tham, different countries have distinct AI regulations. Before any data or model weight sharing, formal agreements will be signed to define data usage rights, compliance obligations, and ownership. Only after signing will collaborative training or model utilization proceed. The Global RetFound model will be available solely for non-commercial research and secondary development, ensuring ethical and lawful international collaboration. 

The Future of a Global-Scale Foundation Model 

Once the first version of the model is trained and validated, the team plans to fine-tune the model in each participating region, including Europe, Asia, and North America, to local retinal datasets. Then, it will move onto external validation and clinical trials, which will be conducted across hospitals in different countries and evaluated for performance across ethnicities, imaging devices, and disease subtypes. 

There will also be different versions for different countries “ideally so, under a structured ecosystem.” As Tham explains, there needs to be a global core model, a shared backbone pre-trained on diverse international data that represents common ophthalmic knowledge. There also needs to be regional adaptors—country-specific modules that are fine-tuned on local datasets to capture differences in disease prevalence, image acquisition styles, and local deployment models. Depending on regulations and hardware availability, there could be high-complexity versions for academic and tertiary hospitals and low-resource versions for community clinics or mobile screening.