TFDF: Self-Supervised Time-Frequency Dynamic Fusion with Dual Constraints for Atrial Fibrillation Detection.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) is a common cardiac arrhythmia requiring continuous ECG monitoring for early detection.
• Supervised AF detection methods face challenges due to the high cost of acquiring annotated ECG data.
• Existing self-supervised learning (SSL) methods struggle to model cross-domain dependencies in AF detection.

Purpose of the Study:

• To propose a novel self-supervised Time-Frequency Dynamic Fusion (TFDF) method for label-efficient AF detection.
• To address the limitations of current SSL methods in modeling temporal and spectral characteristics of AF.
• To develop a generalizable solution for AF detection using unlabeled ECG data.

Main Methods:

• Developed a Time-Frequency Dynamic Fusion (TFDF) model integrating temporal RR interval features and multi-scale spectral representations.
• Implemented a directional consistency constraint for adaptive cross-domain feature fusion.
• Utilized a cluster-guided constraint for stabilizing feature alignment during unsupervised pretraining.
• Pretrained TFDF on the MIT-BIH AF Database and fine-tuned on the CPSC2018 dataset.

Main Results:

• TFDF achieved an average F1-score of approximately 0.920 and AUC of approximately 0.979 on the Chapman-Shaoxing 12-lead ECG dataset.
• The proposed method outperformed existing state-of-the-art self-supervised learning baselines.
• Demonstrated effective fusion of temporal and spectral ECG features under label-free conditions.

Conclusions:

• The TFDF method offers a generalizable and label-efficient approach for AF detection.
• Self-supervised learning with dual constraints can effectively model cross-domain dependencies in ECG data.
• This approach reduces the reliance on large annotated datasets for AF detection.