Scaling Representation Learning From Ubiquitous ECG With State-Space Models.

Area of Science:

• Computational biology
• Machine learning for healthcare
• Wearable biosensing

Background:

• Wearable devices offer continuous health monitoring but generate vast, complex data challenging traditional methods.
• Representation learning from biological signals is advancing, yet ECG studies often use limited data and complex models.

Purpose of the Study:

• To develop a robust representation learning model for electrocardiogram (ECG) signals collected in real-world settings.
• To address challenges posed by large-scale, heterogeneous data from wearable biosensors.

Main Methods:

• Introduced WildECG, a pre-trained state-space model for self-supervised representation learning.
• Trained the model on 275,000 10-second ECG recordings obtained 'in the wild'.
• Evaluated the model's performance on diverse downstream tasks.

Main Results:

• WildECG demonstrates competitive performance across various ECG analysis tasks.
• The model shows particular efficacy in low-resource scenarios.
• Established WildECG as a robust backbone for ECG analysis.

Conclusions:

• Self-supervised representation learning with large, real-world ECG datasets is effective.
• WildECG provides a powerful and adaptable tool for advancing health monitoring and diagnostics.
• The model's performance highlights the potential of large-scale, in-the-wild data for biosignal analysis.