Cross-Site Generalizability of ECG-Derived Biological Age for Cardiovascular Risk Stratification

Biological age reflects the functional state of the heart and can differ from chronological age, providing a useful indicator of cardiovascular health. This difference, estimated from electrocardiogram (ECG) signals as ECGderived biological age, has emerged as a promising digital biomarker for cardiovascular risk stratification. However, the ability of such models to generalise across hospitals and to leverage complementary clinical context such as electronic health records (EHR) remains unclear. This work presents a multi-task convolutional transformer that jointly estimates ECG-age, predicts ICD-coded risk for myocardial infarction (MI), stroke, and heart failure (HF), and computes a mortality hazard score from 12-lead ECG signals combined with temporal ICD histories, an aspect largely unexplored in prior ECG-derived biological age. The model is trained exclusively on Massachusetts General Hospital (MGH) data from the Harvard-Emory ECG Database and evaluated on the complete Emory University Hospital (EUH) cohort (n = 998,844) without any domain adaptation. On this unseen cohort, the model achieves an age prediction mean absolute error of 9.97 years (Pearson r = 0.64), with AUROCs of 0.675 (MI), 0.676 (HF), 0.656 (stroke), and 0.580 (mortality C-index = 0.582). Saliency analysis highlights QRS-complex and T-wave regions, consistent with known cardiac ageing patterns. Subgroup analysis shows broadly equitable performance across race and age groups (AUROC range 0.505-0.721). A clinical reasoning module translates model outputs into structured risk narratives for patient-level interpretation. These findings position ECG-derived age and risk prediction as a generalisable digital biomarker for large-scale cardiovascular screening and early risk stratification.