A multimodal machine-learning diagnostic model for Alzheimer's disease: integrating OCTA-derived retinal...

Alzheimer's disease (AD) lacks widely accessible, noninvasive approaches for the evaluation of AD remain limited. We developed a multimodal machine-learning model that integrates retinal microvascular features derived from optical coherence tomography angiography (OCTA) with clinical variables to classify AD.

In this cross-sectional case-control study, 160 participants (81 healthy controls and 79 patients with AD) underwent macular OCTA imaging. Twenty-seven superficial capillary plexus features, including vessel density (VD), perfusion density (PD), and foveal avascular zone (FAZ) metrics, were analyzed. Logistic regression, support vector machine, random forest, and XGBoost classifiers were evaluated using stratified five-fold cross-validation. Feature selection was performed with nested cross-validated recursive feature elimination, and model interpretability was assessed with Shapley additive explanations (SHAP).

Patients with AD showed reduced VD and PD across multiple macular subfields. The multimodal random forest model achieved the highest area under the receiver operating characteristic curve (AUC = 0.853), whereas the multimodal XGBoost model achieved the highest sensitivity (0.800) with an AUC of 0.839. SHAP analysis identified whole-macula VD, nasal inner-ring VD, FAZ circularity, and inferior inner-ring PD as influential predictors.

These findings provide proof-of-concept evidence that OCTA-derived retinal microvascular features, particularly when combined with selected clinical variables, may help distinguish clinically diagnosed AD from HC in this selected case-control cohort. Their utility for population screening or clinical diagnosis remains unestablished and requires prospective external validation in consecutive, clinically representative cohorts including mild cognitive impairment (MCI), other dementias, and relevant neurological and ocular differential diagnoses.