Abstract
Alzheimer's disease (AD) is the most prevalent form of dementia, yet its early detection remains challenging due to the invasiveness, cost, and limited accessibility of current diagnostics. Increasing evidence suggests that retinal changes mirror cerebral pathology in AD, making the eye a promising site for non-invasive biomarker discovery. Here, we present a technique employing a custom-built tri-spectral retinal imaging module, designed to be integrated with existing fundus imaging systems, that captures retinal reflectance across three optimized spectral bands to quantify spectral alterations linked to AD. We validate the system in a case-control study of 38 mild AD patients and 28 age-matched controls, revealing spatially resolved differences in a fundus map derived from the blue-to-green ratiometric channel (p < 0.001). Our analysis identifies specifically the fovea-to-optic disc region as the most discriminative for AD, with an AUC of 0.74. Building on this, we developed a biologically informed machine-learning classification model incorporating spectral, clinical, and demographic data. On an independent validation test, the model achieved an AUC of 0.91, matching or slightly outperforming the most advanced spectral retinal measurements, yet using a simpler, more stable, and cost-effective setup that further facilitates clinical translation. The demonstrated technology, thanks to its non-invasiveness and its integrability with both existing medical technologies and advanced quantitative statistical methods, holds the potential to drive a significant leap forward in the early detection of AD, opening a window for timely intervention and thus profoundly impacting patient care.