Parsa Khalafi ¹ Soroush Morsali ^2,3 Sana Hamidi ² Hamidreza Ashayeri ² Navid Sobhi ⁴ Siamak Pedrammehr ^5,6* Ali Jafarizadeh ^4*

• ¹ School of Medicine, Tehran University of Medical Sciences, Tehran, Alborz, Iran
• ² Tabriz University of Medical Sciences, Tabriz, Iran
• ³ Neurosciences Research Center, Tabriz University of Medical Sciences, Tabriz, Iran
• ⁴ Nikookari Eye Center, Tabriz University of Medical Sciences, Tabriz, Iran
• ⁵ Institute for Intelligent Systems Research and Innovation, Faculty of Science, Engineering and Built Environment, Deakin University, Geelong, Australia
• ⁶ Tabriz Islamic Art University, Tabriz, Iran

Retinal imaging, used for assessing stroke-related retinal changes, is a non-invasive and cost-effective method which can be enhanced by machine learning and deep learning algorithms, showing promise in early disease detection, severity grading, and prognosis evaluation for stroke patients. This review explores the role of artificial intelligence (AI) in stroke patient care, focusing on retinal imaging integration into clinical workflows. A decrease in the central retinal artery diameter and an increase in the central retinal vein diameter are associated with both lacunar stroke and intracranial hemorrhage. Additionally, microvascular changes like arteriovenous nicking, increased vessel tortuosity, enhanced arteriolar light reflex, straighter retinal arterioles, decreased retinal fractals, and thinning of retinal nerve fiber layer are also reported to be associated with higher stroke risk. For stroke risk stratification using established risk assessment tools, AI analysis of retinal vascular distortions using EfficientNetV2-B3 successfully categorized patients into stroke risk groups, although AI-assisted retinal viscometry did not significantly improve the predictive performance of the Framingham Risk Score. The introduction of "retinal age gap" as a novel predictor revealed that each additional year in this gap was associated with a 4% increase in stroke risk. Moreover, a Random Forest model effectively distinguished between ischemic and hemorrhagic stroke subtypes based on retinal features, showing superior predictive performance compared to traditional clinical characteristics. While evaluating the stroke diagnostic tools, the use of vasculature embeddings combined with gradient boosting outperformed both the VGG19 and Inception-v3 models. A self-supervised learning model also demonstrated effective feature extraction from multimodal retinal images, achieving high performance for microvasculature density. The researchers confront several challenges, including the lack of standardized protocols for imaging modalities, hesitance in trusting AI-generated predictions, insufficient integration of retinal imaging data with electronic health records, the need for validation across diverse populations, and ethical and regulatory concerns. Incorporating quantum AI, explainable AI, and transfer learning algorithms would enhance the outcomes in the future.