Predicting ischemic stroke risk from atrial fibrillation based on multi-spectral fundus images using deep learning

Area of Science:

• Ophthalmology
• Cardiology
• Neurology
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) significantly increases the risk of ischemic stroke (IS) within one year.
• Both AF and IS share early manifestations of ischemia and hypoxia in small blood vessels.
• The potential of using multi-spectrum fundus imaging for predicting IS risk in AF patients remains underexplored.

Purpose of the Study:

• To predict the risk of IS secondary to AF using deep learning on multi-spectrum fundus images.
• To investigate the utility of different spectral wavelengths and their combinations for IS risk prediction.
• To explore a novel approach for the early detection and prevention of IS in AF patients.

Main Methods:

• Recruited 150 AF patients (no IS within 1 year) and 100 IS patients (with AF history).
• Collected fundus images at 548, 605, and 810 nm wavelengths.
• Trained three deep neural network (DNN) models (Inception V3, ResNet50, SE50) using spectral fundus images and clinical data.

Main Results:

• All DNN models achieved >78% accuracy using single or multi-spectral images.
• The 605 nm spectral images provided more stable IS detection performance.
• Multi-spectral combination models demonstrated higher Area Under the Curve (AUC) scores than single-spectral models.

Conclusions:

• Multi-spectrum fundus imaging combined with deep learning can effectively predict IS risk in AF patients.
• Utilizing diverse spectral fundus images aids in the early prevention of cardiovascular and cerebrovascular diseases.
• This study offers a novel preliminary exploration for predicting challenging conditions like IS.