Artificial intelligence is changing how doctors read retinal images. But one question remains underexplored: can eye AI work reliably for people living at high altitude?
A new study by Dr. Wu Yuan’s team at The Chinese University of Hong Kong has been accepted by Science Bulletin . The study, titled “A mixture of ophthalmic foundation models enables mitigation of altitude-associated domain shift,” examines a hidden challenge for ophthalmic AI: altitude-associated domain shift.
The team systematically constructed a cross-altitude, multi-disease fundus image dataset using color retinal photographs from Chinese medical centers. The dataset covered plain, high-altitude and very-high-altitude clinical settings, from about 40 ft to more than 12,000 ft above sea level. The study focused on four image categories: normal fundus, glaucoma, high myopia and age-related macular degeneration.
The researchers tested four widely used foundation models: RETFound, VisionFM, FLAIR and CLIP. The results showed a clear warning sign. When these models were tested across high-altitude cohorts, their performance dropped. For some models, the cross-altitude AUROC gap exceeded 14%.
To reduce this gap, the team developed MIXFound. Instead of training a new giant model, MIXFound combines the strengths of several existing foundation models. It keeps their pretrained image encoders frozen, learns how well each model performs for each disease class, and then mixes their predictions using performance-informed weights.
In simple terms, MIXFound lets different AI models contribute more where they are strongest.
Across seven evaluation cohorts, MIXFound delivered more stable results than individual foundation models. It achieved statistically significant improvements over four comparator models in 27 of 28 cohort-level comparisons after correction for multiple testing. It also outperformed simpler fusion strategies, showing that its benefit was not just the result of averaging models together.
The study also found altitude-associated differences in retinal blood vessel features in healthy individuals. These findings may help explain why retinal images from high-altitude populations can look different to AI systems. The authors caution, however, that altitude should be understood as a marker of a compound real-world shift, not as a single causal factor.
Overall, the study highlights a practical lesson for medical AI: strong performance in familiar datasets does not guarantee reliability everywhere. MIXFound offers a lightweight way to improve robustness when ophthalmic AI is used in geographically diverse and underrepresented populations.
The framework is intended as a decision-support layer, such as for second reading or triage support, rather than as a standalone replacement for clinical diagnosis.
Science Bulletin
Imaging analysis