Learning Cardiac Features: ECG Biometrics Across Time and Exercise

This research explores the viability of using electrocardiograms (ECGs) as a reliable biometric tool for individual identification. While ECGs are known to contain unique, subject-specific patterns, most existing research has focused on data collected while a person is at rest during a single session. This paper addresses the challenge of whether these cardiac signatures remain consistent when a person is under physical stress or when their data is collected at different times, which is essential for developing robust, real-world authentication systems.

The Challenge of Cardiac Variability

The primary obstacle in ECG biometrics is that cardiac signals are not static; they fluctuate based on physical exertion and the passage of time. Previous studies have largely ignored these "temporal and physiological drifts," leaving the reliability of ECG-based identification in doubt for practical applications. By testing how well an individual can be identified during exercise-induced stress and across different time sessions, the authors aim to determine if there is an "intrinsic cardiac signature" that persists despite these changes. The same ai evaluation question is explored in LimiX-2, which adds a research perspective.

A New Approach to ECG Analysis

To tackle this, the researchers employed a Siamese ResNet model, a type of neural network architecture well-suited for comparing pairs of inputs to determine similarity. They enhanced this model with a "late multi-lead fusion strategy," which allows the system to integrate information from multiple ECG leads effectively. The model was trained on a large dataset derived from cardiopulmonary exercise tests, ensuring that the system learned to recognize features that remain stable even as the heart rate and physical state of the subject change.

Key Findings and Performance

The study demonstrates that an intrinsic, resilient cardiac signature does indeed exist. The researchers achieved an Equal Error Rate (EER) of 1.7% for intra-session identification, measuring the transition from resting to peak exercise. Furthermore, when tested against the public CYBHi dataset, the model achieved a state-of-the-art performance of 3.9%. These results suggest that ECG biometrics can be a robust method for authentication, even when accounting for the significant physiological shifts caused by exercise and the natural variations that occur over time. The same ai evaluation question is explored in Q&A on Any Spreadsheet Requires Interpreting..., which adds a research perspective.

Implications for Future Technology

Beyond simple authentication, the authors highlight that this research has broader implications for data security. Because these cardiac patterns are unique and resilient, they could be used to better secure sensitive medical data. Additionally, the ability to extract these stable features provides a promising foundation for self-supervised learning, where the heart's electrical activity could serve as a reliable pretext task for training more advanced AI models in the healthcare domain. The same reasoning question is explored in MAPLE, which adds a research perspective. as detailed in the full paper on Arxiv