Four recent arXiv preprints tackle deep learning for electrocardiograms (ECGs), but from different directions: one examines vision-language models in emergency-department prediction, one explores ECG biometrics, and two propose architectural changes for diagnosis and classification. Despite their different tasks, they converge on a shared concern: generic representations do not fully exploit the structure of ECG signals.
ECG Mirage (arXiv:2609.21755) reports that vision-language models underutilise ECGs when making clinical predictions, and proposes ways to mitigate that gap. BEAT-Net (arXiv:2601.07316) makes a similar argument for diagnosis, criticising signal-agnostic approaches that treat multi-lead recordings as undifferentiated time-series or images, and instead injects biomimetic spatio-temporal priors for interpretability. ECG-Mamba-V2 (arXiv:2510.13046) refines a bidirectional state space model for multi-label 12-lead classification, noting that the design choices behind such models' accuracy remain unclear.
The outlier is Learning Cardiac Features (arXiv:2609.21962), which uses ECGs for biometric identification rather than clinical diagnosis. It shows that ECGs carry subject-specific patterns that remain reliable across time and exercise, suggesting a role in authentication and securing sensitive data. This paper agrees with the others that ECG-derived features are valuable, but it applies them to a different problem.
Taken together, the papers indicate a shift toward ECG-specific inductive biases in deep learning, whether for interpretability, efficiency, or identity. They differ in task and method, but all push back against treating ECGs as just another time series.