Capturing population differences in rates of vascular aging using a deep learning electrocardiogram algorithm: a cross-sectional study

Background Cardiovascular event rates increase with age in all populations. This is thought to be the result of multiple underlying molecular and cellular processes that lead to cumulative vascular damage. Apart from arterial stiffness based on pulse wave velocity there are few other non-invasive measures of this process of vascular aging. We have developed a potential biomarker of vascular aging using deep-learning to predict age from a standard 12-lead electrocardiogram (ECG). The difference between ECG predicted and chronological age (δ-age) can be interpreted as a measure of vascular aging. Methods We use data collected in two cross-sectional studies of adults aged 40-69 years in Norway and Russia to test the hypothesis that mean levels of δ-age, derived from a deep-learning model trained on a US population, correspond to the known large differences in cardiovascular mortality between the two countries. Findings Substantial differences were found in mean δ-age between populations: Russia-USA (+5·2 years; 0·7, 10 IQR) and Norway-USA (-2·6 years; -7, 2 IQR). These differences were only marginally explained when accounting for differences in established cardiovascular disease risk factors. Interpretation δ-age may be an important biomarker of fundamental differences in cardiovascular disease risk between populations as well as between individuals.