The AI-ECG clock as a biomarker of cardiovascular ageing: molecular, genetic and clinical validation in three cohorts

Background Deep learning can estimate age from 12-lead electrocardiograms (ECGs)(1), but whether the discrepancy between predicted and chronological age reflects meaningful cardiovascular aging remains uncertain. We developed an artificial intelligence (AI)-derived ECG clock and tested whether this "clock acceleration" (AI-predicted minus chronological age) captures electrophysiological, structural, and molecular signatures of cardiovascular aging. Methods The ECG clock was derived from a one-dimensional residual neural network (1D-ResNet) trained on 1.2 million 12-lead ECGs from Beth Israel Deaconess Medical Center (BIDMC). Prognostic associations were tested using age- and sex-adjusted Cox models in BIDMC (n=62K), the Airwave Health Monitoring Study (AHMS; n=45K), and Massachusetts General Hospital (MGH; n=1.2M). Contemporaneous echocardiographic, metabolomic, lipoprotein and proteomic correlates were evaluated in the multi-omic subset of AHMS. Genetic validation used Mendelian randomisation across ten cardiovascular outcomes, with false discovery rate (FDR) correction. Results Model performance was robust across cohorts (MAE: 7.48 years in BIDMC, 6.44 in AHMS, and 8.16 in MGH). Higher ECG clock acceleration showed graded risk increases, strongest for electrophysiological outcomes: atrial flutter (hazard ratio [HR] range across cohorts: 1.26–1.62), atrial fibrillation (1.26–1.37), bundle branch block (1.23–1.79), complete heart block (1.29–2.07), and sudden cardiac death (1.09–1.34). Structural associations were also consistent: dilated-cardiomyopathy (1.22–1.84), heart failure (1.24–1.37), and aortic aneurysm (1.21–1.35). Accelerated ECG age correlated with atrial enlargement, conduction prolongation, and reduced ejection fraction on echocardiography. Plasma metabolomics analysis revealed association of heart age acceleration with increased levels of circulatory fatty acids, pro-inflammatory lipids, acylcarnitines, purine metabolism and ethanolamides. In opposition we found higher levels of antioxidants (dietary, ether/plasmalogen phospholipids), steroid metabolites, branched-chain aminoacids and N-acyl ethanolamides, as well as large HDL particles. Phenylalanine and tyrosine metabolism pathway, urea cycle and arginine metabolism presented both pro- and anti- ageing metabolites. Proteomics analysis also revealed an inflammatory pattern (chemokines, IL-1 signalling). Mendelian randomisation showed strongest genetic support for atrial fibrillation, with multiple FDR-significant loci; heart failure and cardiomyopathy signals were more modest, and vascular associations were weaker. Conclusion The AI-ECG clock captures a scalable cardiovascular and electrophysiological ageing phenotype with linked structural, molecular and genetic correlates. Classifying individuals with accelerated versus decelerated ECG ageing improves interpretability and supports use in risk stratification, trial enrichment, and therapeutic target discovery.