MIRAGE: a multimodal deep learning framework for interpretable risk assessment of high myopia from genetic and retinal imaging data.

High myopia (HM) is a complex condition influenced by both genetic and environmental factors, yet its early prediction and clinical intervention remain challenging due to heterogeneous progression patterns. To support early identification of individuals at risk for HM, we developed MIRAGE, a deep learning framework combining exome-wide genotypes with fundus images for personalized prediction. The model combines DeepExGRS for genetic risk modeling and a convolutional network for imaging, fused via a gating attention mechanism. Applied to a cohort of 1991 individuals, MIRAGE achieved high predictive accuracy (AUC = 0.963), outperforming unimodal models. Interpretation using Integrated Gradients and Shapley values identified key genes (e.g. GJD2, FGF1) and image regions (optic disc, macula) contributing to predictions. Further, deep learning-based genetic risk models (DeepcvGRS and DeeprvGRS) outperformed traditional polygenic risk scores, capturing nonlinear interactions and improving HM prediction. Interaction analysis using Shapley interaction scores revealed 314 significant gene-gene interactions, including biologically relevant pairs such as PRSS56-GLI3 and SEMA4D-FRY. Overall, MIRAGE offers a scalable, interpretable, and accurate approach for HM risk stratification, paving the way to early intervention guided by both genetics and retinal imaging.