Document resource
Background Deep learning can estimate age from 12-lead electrocardiograms (ECGs), but whether this reflects true ageing biology is uncertain. We developed an artificial intelligence (AI)-derived ECG clock and tested whether ECG clock acceleration (bias-corrected ECG clock age minus chronological age) captures electrophysiological, structural, molecular and genetic ageing signals.Methods A one-dimensional residual neural network (1D-ResNet) ECG clock was trained on 1.2 million Beth Israel Deaconess Medical Center (BIDMC) ECGs. Prognostic associations were tested using age- and sex-adjusted Cox models in BIDMC (n=62K), 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. Genetic validation used Mendelian randomisation (MR) across 10 outcomes: atrial fibrillation (AF), dilated cardiomyopathy (DCM), heart failure (HF), diastolic blood pressure, low-density lipoprotein cholesterol, type 2 diabetes, ischaemic heart disease, ischaemic stroke, venous thromboembolism (VTE), and thoracic aortic aneurysm (TAA), using single-nucleotide polymorphism (SNP)-level Wald ratio estimates with false discovery rate (FDR) correction.Results The ECG clock performance was robust across cohorts (mean absolute error (MAE) in age estimation 7.48 years in BIDMC, 6.44 in AHMS, and 8.16 in MGH). Higher ECG clock age acceleration showed graded risk increases, strongest for electrophysiological outcomes: atrial flutter (hazard ratio (HR) 1.62; 1.26; 1.26), AF (1.37; 1.35; 1.26), sick sinus syndrome (1.55; NA; 1.29), bundle branch block (1.23; 1.79; 1.23), complete heart block (1.65; 2.07; 1.29), ventricular tachycardia (VT; 1.34; NA; 1.21), ventricular fibrillation (VF; 1.41; NA; 1.20), and sudden cardiac death (1.34; 1.09; 1.21). Structural associations were also consistent: DCM (1.59; 1.84; 1.22), HF (1.27; 1.37; 1.24), aortic aneurysm (1.27; 1.35; 1.21), and hypertrophic cardiomyopathy (HCM; 1.33; 1.67; 1.19). Multi-omic profiling identified convergent dysregulation of lipid metabolism, arginine-urea pathways, mitochondrial beta-oxidation and glutathione-redox biology, with a high-density lipoprotein-depleted, apolipoprotein B-enriched lipoprotein profile. MR showed the strongest and most coherent genetic support for AF, including multiple FDR-significant loci; HF/cardiomyopathy signals were more modest, and vascular/cardiometabolic associations were weaker and likely indirect, consistent with predominantly electrophysiological clinical risk.Conclusion The ECG clock captures a scalable cardiovascular and electrophysiological ageing phenotype with linked structural, molecular and genetic correlates. Classifying individuals as accelerated versus decelerated ECG ageing improves interpretability and supports use in risk stratification, trial enrichment and therapeutic target discovery.