A Machine Learning Prediction Model to Identify Individuals at Risk of 5-Year Incident Stroke Based on Retinal...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Cardiovascular Disease Prediction

Background:

• Stroke is a major cause of death and disability globally.
• Current stroke risk prediction models have limitations in accuracy and scope.
• Retinal imaging offers a novel, non-invasive window into systemic health.

Purpose of the Study:

• To develop and validate an AI-based model for predicting 5- and 10-year incident stroke risk.
• To assess the efficacy of incorporating retinal imaging data into stroke risk prediction.
• To compare the AI model's performance against traditional stroke risk scores.

Main Methods:

• Trained an AI model on a proprietary dataset (>6500 participants) including retinal images, socio-demographics, and risk factors.
• Validated the model externally on the UK Biobank dataset.
• Utilized fundus photography and ophthalmoscopy for retinal image acquisition.
• Compared AI model performance (AUC, sensitivity, specificity) with Framingham and CHADS2 scores.

Main Results:

• The AI model achieved 80% sensitivity and 82% specificity for 5-year stroke prediction (AUC 0.83).
• For 10-year prediction, the model yielded 72% sensitivity and 78% specificity (AUC 0.79).
• Retinal features significantly enhanced prediction accuracy compared to models without them and traditional scores (Framingham AUC 0.73, CHADS2 AUC 0.74).

Conclusions:

• AI models incorporating retinal imaging data significantly improve the prediction of 5- and 10-year incident strokes.
• This approach offers a promising, non-invasive method for early identification of individuals at high stroke risk.
• The validated model demonstrates potential for widespread clinical application in stroke prevention strategies.