Predicting stroke risk using retinal imaging with artificial intelligence: a scoping review of current evidence.

backgroundStroke remains a leading cause of mortality and long-term disability worldwide, while currently used clinical risk prediction tools demonstrate only moderate accuracy. Retinal imaging combined with artificial intelligence has emerged as a promising non-invasive approach for improving stroke risk stratification by capturing microvascular changes that reflect cerebrovascular health.

methodsThis scoping review was conducted in accordance with Joanna Briggs Institute methodology and reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews guidelines. Electronic searches were performed in PubMed/MEDLINE, Scopus, and Web of Science to identify studies applying artificial intelligence techniques to retinal imaging for stroke risk prediction. Quantitative human studies reporting predictive performance metrics were eligible for inclusion. Eleven studies were included. Most investigations applied deep learning models to fundus photography or optical coherence tomography angiography.

resultsReported predictive performance varied substantially, with area under the receiver operating characteristic curve values ranging from 0.719 to 0.987. Retinal vascular geometry, morphology, and network complexity were the most consistently identified predictors of stroke risk. Models integrating retinal imaging with clinical variables generally demonstrated improved discrimination compared with image-only approaches. However, the majority of studies relied on internal validation, and more than two-thirds were judged to have a high risk of bias. Several studies were based on large shared datasets, particularly the UK Biobank, which may limit interpretation of the overall sample size.

conclusionOverall, artificial intelligence-based retinal imaging shows considerable potential as a complementary tool for stroke risk prediction, but the current evidence is limited by methodological heterogeneity, insufficient external validation, and concerns regarding generalizability. Robust prospective and externally validated studies are required before routine clinical implementation can be recommended.