Multimodal deep learning reveals the modular genetic architecture of cardiovascular aging

Background Accelerated biological aging is strongly associated with cardiovascular morbidity and mortality. Deep learning (DL) enables estimation of biological age across diverse cardiovascular modalities, generating quantitative biomarkers of system-level aging. This framework provides an opportunity to systematically evaluate the heritability, genetic overlap, and disease correlations of modality-specific aging signatures, providing biological context with potential relevance to cardiovascular risk. Purpose To develop system-specific DL-derived cardiovascular age models across electrocardiography (ECG), cardiac magnetic resonance (CMR), carotid ultrasound, and retinal imaging, and to quantify the heritability, genetic overlap, and disease correlations of their modality-specific aging signatures. Methods We leveraged data from 163,721 UK Biobank participants (median age 67 years [IQR: 60-73]), including 91,057 ECGs, 82,615 CMRs, 82,929 carotid ultrasounds, and 96,164 retinal images. We trained modality-specific aging models using a ResNet architecture to predict chronological age, leveraging a development cohort of 5000 individuals, and inferred the DL age on the remaining held-out test cohort. We performed multiple modality-specific genome-wide association studies (GWAS) using the delta between chronological and DL-derived biological age in participants of European ancestry. Furthermore, we conducted functional annotation and linkage disequilibrium score regression (LDSC) to estimate heritability (h2). Finally, genetic correlations (rg) were calculated between modalities and with various clinical phenotypes including cardiovascular diseases, chronic conditions, and biometric traits. Results The modality-specific age models predicted chronological age with mean absolute errors (MAE) ranging from 4.27 to 5.90 for retinal imaging and ECG, respectively (Figure 1A). In participants with available data across all four modalities, we observed largely distinct patterns with limited concordance across modalities (Figure 1B). Modality-specific delta-age GWAS identified 60 independent genome-wide significant loci (ECG: 20, CMR: 12, carotid ultrasound: 10, retinal imaging: 18). SNP-based heritability ranged from 9.8% to 12.4%. Genetic correlation between the four aging metrics was low across ECG, CMR, and carotid ultrasound (rg range: 0.2–0.36), and minimal with retinal imaging (Figure 2). Clinically, delta-ages demonstrated distinct genetic correlations; for example, ECG age correlated most strongly with atrial fibrillation (rg = 0.37), while carotid ultrasound age correlated with heart failure (rg =0.22). Conclusion Deep learning–derived biological age models across cardiovascular modalities capture system-specific aging signatures with limited genetic overlap, supporting a heterogeneous model of cardiovascular aging. Multimodal aging biomarkers may therefore improve organ-specific risk stratification.Figure 1Figure2