A novel multimodal self-supervised framework for ECG arrhythmia classification.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Electrocardiogram (ECG) is crucial for cardiovascular disease diagnosis.
• High costs of annotated ECG data limit supervised learning.
• Self-supervised learning (SSL) can utilize abundant unlabeled ECG data.

Purpose of the Study:

• To propose a novel multimodal self-supervised framework for ECG arrhythmia classification.
• To enhance ECG signal feature learning using time and frequency domain representations.
• To improve model initialization for ECG classification tasks.

Main Methods:

• Developed a simple multimodal self-supervised framework for ECG pre-training.
• Utilized contrastive learning with time-domain and time-frequency losses.
• Evaluated the method on multi-lead and single-lead ECG datasets.

Main Results:

• The proposed pre-training method significantly improved downstream classification performance.
• Outperformed standard contrastive learning paradigms in accuracy (ACC) and area under the curve (AUC).
• Achieved superior results compared to traditional supervised learning methods.

Conclusions:

• The multimodal self-supervised framework effectively learns ECG signal features.
• This approach mitigates the need for extensive labeled ECG data.
• Offers a promising direction for enhancing ECG-based disease classification.