BioFPT: Biosignal Feature Pyramid Transformer for self-supervised representation learning from ECGsignals.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Deep learning in electrocardiogram (ECG) analysis offers diagnostic potential but faces challenges with labeled data, processing efficiency, and signal quality.
• Current methods often require extensive labeled datasets, limiting practical clinical application.
• Limitations in computational efficiency and robustness to varying signal quality hinder widespread adoption.

Purpose of the Study:

• To introduce BioFPT (Biosignal Feature Pyramid Transformer), a novel self-supervised learning framework for ECG signal analysis.
• To address limitations of existing deep learning models in ECG analysis, particularly concerning data requirements and efficiency.
• To develop a versatile architecture for biosignal processing applicable in data-scarce environments.

Main Methods:

• Developed BioFPT, a self-supervised learning framework utilizing a Split Mask-Join (SMJ) transformation for pre-training.
• Incorporated an overlapping embedding mechanism to eliminate the need for positional encoding.
• Enhanced architectural efficiency with a Spatial Reduction Attention (SRA) transformer to reduce computational complexity without performance loss.

Main Results:

• BioFPT achieved a 4.2% accuracy improvement and a 14.8% parameter reduction compared to state-of-the-art models across seven public ECG datasets (over 94,000 subjects).
• Demonstrated robust performance across diverse pathological conditions and varying signal qualities.
• The framework proved effective in scenarios with limited labeled data availability.

Conclusions:

• BioFPT represents a significant advancement in self-supervised ECG analysis, overcoming key limitations of current deep learning approaches.
• The framework's efficiency and effectiveness make it suitable for clinical diagnostics, especially where labeled data is scarce.
• The versatile architecture of BioFPT shows potential for broader applications in analyzing various biosignals.