Electrocardiogram derived heart age models agreement, accuracy and predictive ability in the Tromsø study

Convolutional neural networks (CNNs) can estimate electrocardiogram (ECG)-based heart age. We compared three published CNNs in the Tromsø Study cohort (7,108 participants) for accuracy, agreement, and prognostic value. Mean absolute error versus chronological age was 6.8, 7.8, and 6.4 years. Correlations with age were ~0.71-0.73 and agreement across CNNs was high (overall ICC 0.86). Using Cox models, we estimated hazard ratios per SD of δ-age (ECG age minus chronological age) for myocardial infarction, stroke, cardiovascular mortality, and all-cause mortality; discrimination was quantified by cross-validated C-index. δ-age predicted higher risk across outcomes; associations were strongest for δ-age₁ with myocardial infarction and all-cause mortality (HR 1.36 (1.11, 1.67) and 1.27 (1.08, 1.50)) and for δ-age₂ with stroke and cardiovascular mortality (HR 1.45 (1.17, 1.80) and 1.48 (1.07, 2.05)). C-indices were similar across models. Despite architectural and training-set differences, CNNs yielded consistent ECG ages and comparable risk prediction in an external population.