Multimodal deep learning of fundus abnormalities and traditional risk factors for cardiovascular risk prediction

Area of Science:

• Ophthalmology
• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Cardiovascular disease (CVD) is a leading global cause of mortality, driven by complex risk factors.
• Early identification and risk prediction are crucial for managing CVD.
• Traditional risk factor assessment can be enhanced with novel predictive markers.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) model for identifying CVD using multimodal data.
• To assess the predictive capability of fundus photography as a non-invasive marker for CVD.
• To compare model performance across different datasets for robust validation.

Main Methods:

• Development of an AI model integrating clinical risk factors and fundus photographs.
• Model training and internal validation using data from Samsung Medical Center (SMC).
• External validation using data from the UK Biobank.
• Analysis of feature importance for both photographic and clinical data.

Main Results:

• The multimodal AI model achieved an AUROC of 0.781 in the SMC dataset and 0.872 in the UK Biobank dataset.
• A significant association was found between predicted CVD risk and actual incidence in the UK Biobank (HR 6.28).
• Feature importance analysis highlighted the predictive value of specific features in fundus photography.

Conclusions:

• AI models integrating multimodal data, including fundus photography, can effectively identify and predict cardiovascular disease risk.
• Non-invasive fundus photography shows potential as a valuable predictive marker for CVD.
• This approach may offer a novel, accessible method for CVD risk assessment.