Stroke is a major health issue worldwide, with 15 million people experiencing a stroke each year. Of these cases, 10 million result in death or permanent disability. Stroke not only causes significant mortality and disability, but it also has significant social and economic consequences. There are treatments available for it, but they are most effective when given soon after the stroke begins, and they become less effective over time. This presents a challenge, as patients need to be quickly taken to a stroke center for a CT or MRI scan to confirm the diagnosis and check for bleeding.

In this presentation, I will present a deep learning system that can accurately predict the likelihood of a patient experiencing an ischemic or hemorrhagic stroke. The system utilizes retinal images and basic medical information to extract multiple retinal biomarkers using a combination of various ensembles of models like GANs, LWnets, EfficientNets and ResNets, as well as classical vision methods. These extracted biomarkers, along with the basic medical data, are then fed into a CatBoost model for prediction. The approach leverages the similarity between retinal and cerebral vessels to provide a rapid, non-invasive, and highly effective way to predict strokes.