Artificial intelligence derived electrocardiographic aging and the risk of new and early onset cardiovascular disease

Abstract Background Artificial intelligence (AI) derived electrocardiographic aging and sex misclassification might be a useful risk marker for future major cardiovascular events (MACE). Methods We developed (1,730,222 ECGs from 662,246 participants) and validated a residual network (ResNet)-based model for age prediction on independent multinational datasets. We then calculated AI-derived ECG age and sex in the Severance hospital dataset (n=578,854) to assess whether (1) the gap between AI-derived ECG age and chronological age (AI-ECG age gap) and (2) ECG sex misclassification would predict new and early-onset (≤65years) cardiovascular events. AI-ECG age gap was categorized into two groups: aged-ECG (≥10 years) and normal ECG age (<10 years) groups based on the mean absolute ECG age gap error from the validation datasets. AI-ECG sex misclassification was defined as ECG sex probability of more than 50% for the opposite sex. Results In the Severance hospital dataset, individuals without pre-existing cardiometabolic conditions (n=533,788) had a mean chronological age, AI-ECG age, and AI-ECG age gap of 50.0 (standard deviation [SD] 16.4), 50.3 (SD 15.8), -0.3 (SD 9.8), respectively. Compared with normal ECG age and ECG sex non-discrepant, those with aged ECG and ECG sex discrepant was associated with significant increased risk of myocardial infarction (hazard ratio [HR] 8.39, 95% confidence interval [CI] 3.58-18.73), stroke (HR 4.20, 95% CI 1.71-10.27), heart failure (HR 9.47, 95% CI 5.39-16.64), cardiovascular death (HR 7.15, 95% CI 5.25-9.73), and MACE (HR 7.41, 95% CI 5.74-9.56). The lifetime risk of MACE among male with aged ECG and ECG sex discrepancy was 27.1% (95% CI, 5.9-32.4) whereas those with normal ECG age and ECG sex non-discrepancy was 10.6% (95% CI, 9.4-11.1). Among female with aged ECG and ECG sex discrepancy, the lifetime risk of MACE was 31.8% (95% CI, 9.6-45.7) compared to 11.4% (10.4-12.0) among those with normal ECG age and ECG sex non-discrepancy. Aged ECG and ECG sex discrepancy had additive effects on MACE events. Conclusions AI-ECG age gap and ECG sex misclassification might be a useful risk marker for future cardiovascular events. Future research is needed to determine whether an AI-derived ECG aging would be useful in clinical practice.