A Multi-Level Multiple Contrastive Learning Method for Single-Lead Electrocardiogram Atrial Fibrillation Detection.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) is a prevalent arrhythmia requiring robust automatic detection.
• Supervised learning for AF detection is hindered by the scarcity of labeled electrocardiogram (ECG) data.
• Developing generalizable AF detection models necessitates addressing the limited labeled data challenge.

Purpose of the Study:

• To propose MLMCL, a semi-supervised contrastive learning method for robust single-lead ECG AF detection.
• To overcome limitations posed by insufficient labeled data in AF detection models.
• To enhance model generalization and performance in identifying atrial fibrillation.

Main Methods:

• Developed MLMCL, a semi-supervised contrastive learning approach for AF detection.
• Utilized multi-level feature representations for contrastive learning, exploiting temporal, channel, and label consistency.
• Integrated labeled and unlabeled data for pre-training and employed domain knowledge augmentation for hard sample generation.

Main Results:

• MLMCL demonstrated superior performance and stability in cross-dataset testing for AF detection.
• The method outperformed existing approaches in external validation tests.
• Results indicate MLMCL's effectiveness and robustness in automated arrhythmia detection.

Conclusions:

• MLMCL offers an effective solution for AF detection with limited labeled ECG data.
• The proposed method shows significant potential for improving automated arrhythmia diagnosis.
• MLMCL's framework is adaptable for multi-lead ECG analysis and other arrhythmia detection tasks.