Zheyuan Wang, Yukun Zhou, Yilan Wu, Jocelyn Hui Lin Goh, Ke Zou, Zhouyu Guan, Yibing Chen, Gabriel Dawei Yang, Ping Zhang, Changchang Yin, An Ran Ran, Miao Li Chee, Can Can Xue, Zhi Da Soh, Samantha Min Er Yew, Danqi Fang, Xujia Liu, Benjamin Sommer Thinggaard, Jakob Grauslund, Haoxuan Li, Yixiao Jin, Jia Shu, Tingyao Li, Nan Jiang, Tingli Chen, Huating Li, Xiangning Wang, Qiang WU, Charumathi Sabanayagam, Siegfried K. Wagner, Carol Y. Cheung, Ching-Yu Cheng, Bin Sheng, Tien Yin Wong, Pearse A. Keane, Yih-Chung Tham
Foundation models (FMs) enable generalizable medical AI, but existing retinal FMs perform best on cross-sectional classification and detection and are less effective for predicting disease incidence and progression. We present RETFound Plus, a CFP-based FM trained with temporal modeling on 1,304,292 fundus photographs from 304,345 participants across multiple visits to learn progression-aware representations. Compared with RETFound, RETFound Plus improved calibration and 5-year risk prediction across systemic and ocular diseases, with larger gains for systemic outcomes (stroke, myocardial infarction, diabetes and hypertension; +4-10% c-index) than ocular outcomes (diabetic retinopathy and glaucoma; +3-7% c-index), and improved risk stratification for systemic diseases (1.2-2.1-fold higher hazard-ratio trend). Results were consistent across external multi-regional, multi-ethnic datasets from the UK, US, Singapore, Hong Kong, and Denmark.