Explainable Multimodal OCT and OCTA Analysis Reveals Structural and Vascular Features of Mild Glaucoma Across Ethnicities

Purpose: To develop and externally validate an explainable artificial intelligence model integrating optical coherence tomography (OCT) and OCT angiography (OCTA) for mild primary open-angle glaucoma (POAG) and to determine the structural and vascular features contributing to classification across ethnically distinct cohorts. Methods: This cross-sectional, multicenter diagnostic study included participants with mild POAG (visual field mean deviation ≥ -6 dB) and normal controls. Quantitative structural OCT and vascular OCTA features from the optic disc and macula were extracted. An Asian cohort (n = 621; 429 glaucoma, 192 controls) was used for model development and testing, and an independent Caucasian cohort (n = 131; 62 glaucoma, 69 controls) was used for external validation. Random Forest models using OCT-only, OCTA-only, and combined OCT/OCTA features were compared using area under the receiver operating characteristic (AUC) curve, sensitivity, and specificity. Feature contributions were assessed using SHapley Additive exPlanations (SHAP). Results: Among 752 participants (mean age, 60 years; 50% male), the multimodal model achieved high discrimination in the internal test set (AUC = 0.975) and comparable performance to OCT-only model (AUC = 0.965; P = 0.248) but outperformed the OCTA-only model (AUC = 0.870; P < 0.001). In the external test set, the multimodal model achieved an AUC of 0.884, outperforming both OCT-only (AUC = 0.801; P < 0.001) and OCTA-only (AUC = 0.792; P = 0.005) models. SHAP analysis identified inferior peripapillary retinal nerve fiber layer thickness and peripapillary perfusion density as key contributors. Conclusions: Explainable multimodal OCT/OCTA analysis identified consistent structural and vascular features of mild glaucoma across ethnically distinct cohorts. OCTA-derived vascular metrics provided complementary information to structural OCT, particularly in the external validation cohort.