NUS Ophthalmology AI Model Advances Disease Prediction with Time- and Person-Sensitive Foundation Breakthrough

Revolutionizing Disease Prediction: NUS Contribution to RETFound Plus AI Model

Researchers from the National University of Singapore (NUS) have played a pivotal role in the development of RETFound Plus, a groundbreaking time- and person-sensitive foundation model designed for predicting and stratifying risks of both ocular and systemic diseases using retinal fundus photographs. Published on March 14, 2026, in npj Digital Medicine, this advancement builds on the original RETFound model and introduces temporal modeling to capture disease progression over multiple patient visits, marking a significant leap in ophthalmic artificial intelligence (AI).

The model was trained on an expansive dataset comprising 1,304,292 color fundus photographs (CFPs) from 304,345 participants, enabling it to learn progression-aware representations that account for individual variability and time-dependent changes. Key contributors from NUS include Ke Zou from the Centre for Innovation and Precision Eye Health and the Department of Ophthalmology at Yong Loo Lin School of Medicine, alongside collaborators from Singapore Eye Research Institute (SERI) and Singapore National Eye Centre.

Understanding Foundation Models in Ophthalmology

Foundation models in AI, first popularized in natural language processing, are large-scale pre-trained systems that can be fine-tuned for diverse downstream tasks. In ophthalmology, RETFound—introduced in a 2023 Nature paper—set the benchmark by learning generalizable representations from unlabeled retinal images for detecting conditions like diabetic retinopathy (DR) and glaucoma. RETFound Plus extends this by incorporating longitudinal data, making predictions 'time-sensitive' through temporal modeling and 'person-sensitive' via personalized risk stratification across multi-ethnic cohorts.

This evolution addresses limitations of cross-sectional models, which excel at static diagnosis but falter in forecasting incidence and progression. For instance, the model predicts 5-year risks for stroke, myocardial infarction, diabetes, hypertension, DR, and glaucoma with improved concordance index (c-index) gains of 4-10% for systemic diseases and 3-7% for ocular ones compared to its predecessor.

The Technical Innovation Behind RETFound Plus

RETFound Plus employs self-supervised learning on sequential CFPs, modeling patient trajectories across visits to discern subtle progression patterns. Self-supervised pre-training on vast unlabeled data reduces reliance on costly annotations, while temporal components—like transformer-based sequence modeling—integrate time-to-event endpoints for survival analysis.

Performance was validated on external datasets from the UK, US, Singapore, Hong Kong, and Denmark, demonstrating robustness across ethnicities. In Singapore validation, it showed enhanced calibration for hypertension and DR, critical given local demographics. Risk stratification improved hazard ratios by 1.2-2.1-fold for systemic risks, enabling precise patient prioritization.

Singapore's Eye Health Landscape and AI Relevance

Singapore faces a rising burden of eye diseases, exacerbated by an aging population and high diabetes prevalence—around 11.6% in adults, per recent health ministry data, with projections to 15% by 2030. DR affects up to 30% of diabetics, while glaucoma cases, currently ~57,800 in those over 60, are expected to surge 43% to 85,800 by 2040. Notably, over one-third of seniors have undiagnosed age-related conditions like cataracts (40.8% undiagnosed) and glaucoma (48.1%), heightening blindness risk.

Younger adults are increasingly affected; severe glaucoma surgeries for 40-49-year-olds nearly tripled by 2025, linked to untreated myopia—affecting 80% of youth. AI models like RETFound Plus promise early intervention, aligning with national screening programs where AI has boosted DR detection efficiency.

Photo by Hyundai Motor Group on Unsplash

NUS Yong Loo Lin School: A Hub for Ophthalmic AI

The Yong Loo Lin School of Medicine at NUS, through its Ophthalmology Department and SERI, has pioneered AI applications. Achievements include national AI deployment for DR screening, processing millions of images with >90% accuracy, and collaborations like the Global RETFound Consortium involving 100+ groups and 100 million images for equitable AI.

Faculty like Yih Chung Tham and Tien Yin Wong lead oculomics research, using retinal images as 'windows to health' for systemic predictions. Ke Zou's work on RETFound Plus exemplifies this, enhancing model generalizability for Asian populations underrepresented in Western datasets.

Other milestones: AI for glaucoma progression, myopia prediction (DeepMyopia), and cognitive decline via retinal ageing markers, validated to forecast dementia 5 years early.

Performance Metrics and Clinical Validation

• Systemic Diseases: Stroke (+10% c-index), myocardial infarction (+8%), diabetes (+6%), hypertension (+4%).
• Ocular Diseases: DR (+7%), glaucoma (+3-5%).
• Risk Stratification: Decile-wise hazard ratios show steeper gradients, aiding triage.

External testing on Singapore data confirmed gains, with better calibration (Brier scores improved 5-15%). Compared to clinician benchmarks, it rivals experts in risk assessment while scaling to population levels.

Access the full study here for detailed benchmarks.

Implications for Precision Medicine in Singapore

In Singapore's universal healthcare, RETFound Plus could integrate into polyclinics for proactive screening. By predicting risks from routine fundus photos—non-invasive and cost-effective—it enables targeted interventions, potentially averting 20-30% of DR progressions and reducing systemic events via early flags.

Stakeholders praise its equity: multi-ethnic training mitigates bias, vital for Singapore's diverse populace (74% Chinese, 13% Malay, 9% Indian). Clinicians note reduced workload; one SERI expert highlighted, "AI shifts focus from detection to prevention."

Challenges and Ethical Considerations

Despite promise, challenges persist: data privacy in longitudinal records, generalizability beyond fundus imaging, and regulatory hurdles for SaMD (software as medical device). Singapore's Health Sciences Authority fast-tracks such innovations, but explainability remains key—RETFound Plus uses attention maps for transparency.

Ethical deployment requires diverse validation; the model's global testing addresses this. Future audits for bias in underrepresented subgroups are essential.

Photo by Chunjiang on Unsplash

Future Directions and Global Collaborations

NUS leads extensions like Global RETFound (100M images), multimodal integration (OCT, language-vision models), and real-world trials. Partnerships with UCL, Moorfields, and CUHK accelerate deployment.

In Singapore, integration with Smart Nation initiatives could yield national oculomics platforms by 2030, forecasting healthcare savings of SGD 100M+ annually from prevented blindness.

Actionable Insights for Researchers and Clinicians

• For Researchers: Leverage open-source RETFound weights; fine-tune on local cohorts for myopia-glaucoma links.
• For Clinicians: Adopt in screening; interpret risk scores alongside biomarkers.
• Training: Upskill via NUS AI ophthalmology modules.
• Patients: Routine fundus checks for holistic health insights.

This model heralds an era of predictive oculomics, positioning NUS at the forefront of Singapore's biomedical innovation.