AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
ECG-Derived Cardiac Biological Age Gap: A Deep Learning Analysis of the PTB-XL Dataset
By Paruchuri A, Paruchuri R
Background: Organ-specific biological age gap metrics, computed as the difference between deep learning-predicted and chronological age, have demonstrated clinical utility as sensitive markers of subclinical pathology in neural and retinal systems. Extension of this paradigm to cardiac electrophysiology offers a non-invasive window into accelerated cardiac aging, with limited prior work characterizing how such a gap varies across major diagnostic categories, its sex-specific patterns, or its waveform-level substrates. Methods: A 1D residual neural network was applied to the PTB-XL dataset (n=21,837 ECGs; mean age 59.8±17.0 years; 52.2% female). The model was trained exclusively on normal ECGs (n=9,522; 80:20 split) to predict chronological age from raw 12-lead waveforms at 100Hz. The ECG age gap was computed for all subjects and age-adjusted by regressing chronological age against the gap in the normal cohort and computing residuals. Group comparisons across myocardial infarction (MI, n=5,403), ST/T-wave changes (STTC, n=2,807), conduction disorders (CD, n=2,304), and hypertrophy (HYP, n=1,309) used independent samples t-tests with effect size estimation. Per-lead waveform features including amplitude range, standard deviation, energy, and skewness were extracted and correlated with the age-adjusted gap using Pearson correlation. Results: The model achieved a mean absolute error of 10.44 years on the normal validation set. After age-adjustment, CD, STTC, and MI showed significantly elevated gaps versus normals (NORM: 0.00±9.67 years): CD (+2.55±11.53, p=1.16e-27, Cohen's d=0.24), STTC (+1.87±11.35, p=6.96e-18, d=0.18), and MI (+1.08±12.08, p=2.35e-09, d=0.10). Hypertrophy did not reach significance (-0.18±11.46, p=0.55). Sex-stratified analysis revealed females demonstrated consistently larger positive age-adjusted gaps versus males across pathological categories (MI: +2.67±12.00 vs -1.61±11.73, p=6.82e-37, d=0.36; CD: +3.24±11.56 vs +1.55±11.42, p=5.22e-04, d=0.15). Waveform analysis identified lead skewness as the dominant correlate of the age-adjusted gap, with V2 skewness (r=0.350) and aVL skewness (r=0.342) showing the strongest associations, while amplitude-based features showed weaker correlations. Conclusions: Age-adjusted ECG biological age gap is significantly elevated in conduction disorders, ST/T-wave changes, and myocardial infarction, with females showing greater positive deviation than males. Waveform skewness rather than amplitude emerges as the primary morphological substrate of cardiac electrical aging, a finding with implications for targeted feature engineering in future models. These results support prospective validation of ECG biological age gap as a cardiac risk stratification tool and its integration within multi-organ biological age frameworks alongside neural and retinal aging biomarkers.