To develop an AI system, DeepRETStroke, that detects silent brain infarctions (SBIs) and predicts stroke risk using retinal images.
Key Findings:
DeepRETStroke achieved 85.2% sensitivity for SBI detection, significantly outperforming traditional screening methods.
The model demonstrated internal AUCs of 0.901 for incident stroke prediction and 0.769 for recurrent stroke prediction.
In a prospective study, DeepRETStroke-guided management led to an 82.44% relative reduction in recurrent strokes compared to conventional screening.
Interpretation:
DeepRETStroke represents a significant advancement in non-invasive stroke risk assessment, leveraging retinal imaging to identify high-risk individuals effectively.
Limitations:
Generalizability concerns due to reliance on Chinese cohorts, potentially introducing biases in diverse populations.
Oversimplified SBI phenotyping may not capture the full complexity of cerebrovascular health.
Conclusion:
DeepRETStroke has the potential to transform stroke prevention strategies by providing an accessible and effective screening tool through retinal imaging.
