Artificial Intelligence-Driven Multimodal Prediction of 10-Year Incident Glaucoma Integrating Genetic and Deep Learning-Derived Imaging Features.

Purpose: Glaucoma is the leading cause of irreversible blindness worldwide. It often remains asymptomatic until advanced stages. Hence, accurate prediction of glaucoma is crucial for timely intervention to prevent vision loss. Design: Population-based prospective cohort study. Subjects: The UK Biobank participants had available color fundus photographs (CFPs), genetic data, and ocular measurements and were free of glaucoma at baseline. The primary analytic cohort for strictly defined incident primary open-angle glaucoma (POAG) comprised 340 cases and 9374 controls; the secondary broadly defined POAG cohort comprised 1241 cases and 34 216 controls. Methods: We developed an interpretable, multimodal machine learning (ML) framework to predict 10-year incident POAG. The framework integrated 5 feature domains: deep learning (DL)-derived CFP features, ocular measurements, polygenic risk scores (PRS), physical and lifestyle factors, and electronic health records. Deep learning-derived imaging features included a CFP glaucoma score and automated vertical cup-to-disc ratio estimation. We benchmarked ten ML algorithms to identify the optimal model. We applied SHapley Additive exPlanations (SHAP) for model interpretability and feature contribution. Main Outcome Measures: The primary outcome was incident POAG defined by International Classification of Diseases 10 code H40.1. A secondary broad POAG phenotype incorporated H40.1, H40.0, and H40.9, while excluding other glaucoma subtypes. Results: In the strictly defined POAG analysis, XGBoost achieved the strongest overall performance for 10-year incident glaucoma prediction, with an area under the receiver operating characteristic curve of 0.927 (95% confidence interval, 0.895-0.959), high sensitivity (0.685) at a fixed specificity of 0.95, and excellent calibration (Brier score, 0.024). The corresponding XGBoost model for broadly defined POAG showed modestly lower discrimination while maintaining similarly strong calibration. Incremental modeling demonstrated that both PRS and DL-derived imaging features provided complementary predictive value beyond traditional clinical factors. SHapley Additive exPlanations analysis identified the CFP DL score, age, PRS, and intraocular pressure as the most influential predictors. A reduced model using only the top 4 SHAP-ranked features retained performance comparable to the all-features multimodal model. Conclusions: Our multimodal ML framework integrating genetic and DL-derived imaging features enables accurate and interpretable prediction of incident POAG. Financial Disclosures: The author has no/the authors have no proprietary or commercial interest in any materials discussed in this article.