HuBERT-ECG: A Self-Supervised Foundation Model for Broad and Scalable Cardiac Applications

Summary

HuBERT-ECG (University of Brescia) adapts the HuBERT speech architecture — which learns discrete acoustic tokens via iterative clustering then predicts masked tokens — to 12-lead ECG. The analogy is apt: ECG signals, like speech, are continuous temporal signals with both local morphological structure (waveform peaks) and long-range temporal patterns (rhythm regularity).