The Impact of Temporal Context Length and Encoding Strategies on Self-Supervised ECG Representation Learning

Summary

Self-supervised ECG models are typically trained on only a few seconds of signal and, increasingly, on discretized token sequences. This paper asks directly whether those two conventions sacrifice information that is needed for rhythm inference and longitudinal consistency in real-world ambulatory recordings. The authors run a controlled ablation on the Icentia11k single-lead dataset, varying exactly two factors while holding the Transformer backbone and the entire training protocol constant.