Comparing self-supervised pretraining methods and transfer configurations for wrist ECG under motion (HumanSys '26)

Compares two SSL pretraining methods, TAP-MAE and BioFAME, across four pretraining→fine-tuning configurations that vary dataset consistency, electrode location and channel configuration. Representations learned on large public chest/multi-lead ECG are transferred to a custom wrist-worn ECG for HR/HRV estimation under motion. TAP-MAE consistently beats the DSP baseline and BioFAME.

• Form factor: wrist

Wrist-worn electrocardiogram (ECG) enables unobtrusive and continuous cardiac monitoring in daily life, but reliable heart rate (HR) and heart rate variability (HRV) estimation remains challenging due to limited labeled data and electrode placement area, and severe motion artifacts. Self-supervised learning (SSL) provides a promising solution by transferring representations learned from large public ECG datasets collected from chest-worn, multi-channel systems to data-scarce wrist-worn systems. However, these data often differ from the target wrist device in electrode location, channel configuration, and signal quality due to motion. We present a study comparing two self-supervised pretraining methods, TAP-MAE and BioFAME, across four pretraining-to-fine-tuning configurations designed to investigate the effects of dataset consistency, electrode location, and channel configuration. TAP-MAE consistently outperforms the digital signal processing baseline and BioFAME for HR/HRV estimation under motion. We further identify the domain gap between public ECG pretraining data and our custom wrist ECG data as the dominant factor affecting transfer performance.

Verbatim from Semantic Scholar (opens in a new tab) · Abstract via Semantic Scholar; © the authors/publisher

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.