Title : The role of AI and machine learning in retinal imaging for Alzheimer’s disease: A systematic review

Introduction: Alzheimer’s disease (AD) is a progressive neurodegenerative disorder and leading cause of dementia, often diagnosed at advanced stages when treatment options are limited. Recently, the retina has emerged as a promising site for non-invasive biomarker detection due to its anatomical and embryological connection to the brain. Advances in retinal imaging have enabled detailed visualization of structural and vascular changes associated with AD. Concurrently, the application of artificial intelligence (AI) and machine learning (ML) has enhanced the ability to analyse complex retinal data and identify subtle patterns indicative of disease. This systematic review aims to evaluate how AI-driven analysis of retinal biomarkers can improve the prediction and diagnosis of AD.

Methods: A systematic review was performed using the databases PubMed, Embase (Ovid), and Web of Science. Search strategies combined keywords and synonyms related to Alzheimer’s disease, retinal biomarkers and AI/ ML methods. Studies were screened using predefined inclusion and exclusion criteria, and risk of bias assessment using the STROBE recommendations were used. 

Results: Twenty-seven studies were identified from the search. The results found that several studies have investigated which retinal biomarkers, when integrated into AI and ML models, demonstrate strong potential for AD, and thus utilisation in disease prediction. Structural biomarkers, including retinal nerve fibre layer (RNFL) and ganglion cell layer (GCL) thickness, consistently showed high discriminative ability, with AUC values ranging from 0.79 to 0.95. Models combining retinal features with neuroimaging data achieved optimal performance (AUC up to 0.936), highlighting the benefit of multimodal approaches. GCL-derived measures, particularly GC-IPL maps, often outperformed RNFL alone, indicating their superior diagnostic utility. Across studies, ML algorithms consistently outperformed traditional statistical methods, achieving moderate-to-high diagnostic accuracy (AUC ~0.69–0.73), with performance dependent on input features. Deep learning (DL) models using fundus photography and OCT data showed improved performance, with AUCs up to 0.91 and accuracies exceeding 90%, and demonstrated the ability to distinguish amyloid-β status. Ensemble and multimodal DL approaches further enhanced performance, achieving AUCs up to 0.94 for AD and ~0.74–0.79 for earlier stages such as MCI. Vascular and microstructural biomarkers, including retinal vessel density and foveal avascular zone (FAZ) radiomic features, provided additional diagnostic value. Incorporating multiple radiomic features improved AUC from ~0.60 to ~0.72, while vascular parameters increased sensitivity and achieved AUCs up to 0.86. Novel imaging techniques, such as tri-spectral retinal imaging, identified significant spectral alterations in AD, particularly within the fovea–optic disc region, where blue-to-green reflectance ratios were highly discriminative (p < 0.001). While spectral imaging alone achieved moderate performance (AUC ~0.74), integration with ML models improved accuracy to AUC 0.91. 

Conclusion: Overall, AI and ML applications leveraging retinal biomarkers show considerable promise as non-invasive, accessible tools for the detection of AD, with potential to facilitate earlier diagnosis and improve clinical outcomes. Across the reviewed studies, diagnostic performance was highest when multimodal data—combining structural, vascular, and emerging optical biomarkers—were integrated within advanced analytical models. However, performance remains comparatively lower in early disease detection.

Dr Zain Girach is a current F1 doctor at Kettering General Hospital, UK, with a special interest in ophthalmology.