Electrocardiograms (ECGs) embed subject‑specific waveform patterns that enable reliable identity discrimination, forming the basis of ECG biometrics. Beyond authentication, this property can protect sensitive cardiac data and serve as a pretext task in self‑supervised learning. Most prior work, however, is confined to single‑session, resting recordings, leaving robustness to temporal and physiological variations largely untested. To address this gap, we evaluate ECG biometrics under realistic conditions that include exercise‑induced stress and cross‑session variability. A Siamese ResNet with late multi‑lead fusion is trained on a large ECG dataset extracted from cardiopulmonary exercise tests. Evaluation follows an exercise‑and‑time‑aware protocol and includes public benchmarks. Results show an intra‑session rest‑to‑peak equal error rate (EER) of 1.7% and a state‑of‑the‑art 3.9% on the CYBHi dataset. These findings support the existence of an intrinsic cardiac signature that remains resilient to physiological and temporal drift.
Review