Multimodal Deep Learning Reveals the Modular Genetic Architecture of Cardiovascular Aging

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence
• Genetics

Background:

• Chronological age is a primary cardiovascular disease risk factor.
• Individual cardiovascular aging varies significantly across different organs and biological pathways.
• Existing methods lack a multi-modal approach to assess cardiovascular aging.

Purpose of the Study:

• To develop and validate deep learning models for estimating biological age from diverse cardiovascular data.
• To investigate the distinct disease associations and genetic underpinnings of multi-modal cardiovascular aging.
• To explore the potential of AI-derived cardiovascular age as a personalized risk assessment tool.

Main Methods:

• Deep learning models were trained on over 100,000 UK Biobank participants' data.
• Four cardiovascular data streams were utilized: electrocardiograms, cardiac MRI, carotid ultrasound, and retinal fundus photographs.
• Phenome-wide association, genome-wide association, and single-cell transcriptomic analyses were performed.

Main Results:

• AI-derived biological age gaps showed limited overlap across the four modalities (electrical, structural, macrovascular, microvascular).
• Each modality was linked to distinct disease profiles, including atrial fibrillation, heart failure, and diabetic retinopathy.
• 38 independent genetic loci were identified, with largely modality-specific signals and distinct cellular contexts.

Conclusions:

• AI-derived cardiovascular age is not a singular biomarker but a set of related phenotypes.
• These phenotypes decompose cardiovascular aging into distinct electrical, myocardial, macrovascular, and microvascular modules.
• This multi-modal approach offers a more precise understanding of cardiovascular aging and disease risk.