Predicting Arterial Stiffness from Retinal Microvasculature: A Machine Learning Approach Integrating OCTA Imaging and Clinical Parameters

BACKGROUND: Arterial stiffness is an independent predictor of cardiovascular morbidity and mortality, but its gold-standard measurement, pulse wave velocity (PWV), requires specialized equipment and is not routinely available in clinical practice. Optical coherence tomography angiography (OCTA) provides a non-invasive window into the systemic microvasculature through retinal imaging. However, the utility of OCTA-derived features for predicting PWV using explainable machine learning (ML) remains underexplored. This study aimed to develop and internally evaluate machine learning models that integrate quantitative OCTA retinal features with clinical parameters to predict mean PWV and to identify key predictors using SHapley Additive exPlanations (SHAP)-based interpretability. METHODS: This cross-sectional study included 653 participants from the PERSIAN Cohort who underwent OCTA imaging and oscillometric PWV assessment. Participants were excluded if OCTA signal strength was below 7 or if PWV recordings showed excessive variability (standard deviation > 1.5 m/s). Quantitative retinal microvascular and texture features from macular and optic disc regions were combined with demographic and metabolic variables. An ensemble feature selection framework combining Boruta, recursive feature elimination with cross-validation (RFECV), and least absolute shrinkage and selection operator cross-validation (LassoCV) was applied. The dataset was split 80:20 into training and test sets. Seven base regression models were trained, with an additional stacking ensemble, using Optuna-based hyperparameter optimization and evaluated on the held-out test set. Internal validation was performed using 10-fold cross-validation. RESULTS: Random Forest achieved the best test performance (R² = 0.538, 95% CI [0.420, 0.645]; RMSE = 1.020 m/s; Pearson r = 0.749, p < 0.001).. The final model retained 21 predictors: nine clinical variables and twelve OCTA‑derived imaging descriptors. SHAP analysis identified waist circumference, age, and glucose as the three most influential features, with retinal texture and morphological descriptors contributing additional predictive signal. CONCLUSIONS: Integrating OCTA‑derived retinal microvascular features with clinical variables using machine learning enables meaningful prediction of arterial stiffness. SHAP analysis highlights both established cardiovascular risk factors and novel retinal imaging biomarkers. These findings suggest that OCTA may serve as a complementary tool alongside clinical data for vascular risk assessment, though prospective external validation is required before clinical translation.