Toward robust automated cardiovascular arrhythmia detection using self-supervised learning and 1-dimensional vision...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Cardiovascular diseases are a leading global cause of mortality.
• Electrocardiogram (ECG) monitoring is increasingly accessible for passive arrhythmia detection.
• Challenges in ECG analysis include class imbalance and noise, hindering traditional machine learning models.

Purpose of the Study:

• To leverage self-supervised learning on large-scale unlabeled ECG data for improved arrhythmia detection.
• To introduce and evaluate PatchECG, a novel 1D Transformer model, for ECG analysis.
• To enhance model performance, efficiency, and robustness against common ECG data issues.

Main Methods:

• Utilized Masked Patch Modelling (MPM) for self-supervised pre-training on 8.2 million unlabeled ECGs.
• Developed PatchECG, a 1D Transformer architecture, for various ECG classification tasks.
• Fine-tuned PatchECG on standard datasets (e.g., PTB-XL) and a large, high-quality multi-label dataset.

Main Results:

• PatchECG achieved state-of-the-art performance on benchmark datasets, setting new records on a large multi-label dataset.
• The model demonstrated a 5x increase in computational efficiency and a 14x increase in model capacity compared to existing methods.
• PatchECG outperformed a state-of-the-art 2D vision Transformer (HeartBEiT) and showed a 2% improvement in handling data challenges like class imbalance and noise.

Conclusions:

• Self-supervised learning, particularly with PatchECG, significantly advances automated arrhythmia detection.
• The PatchECG model offers a computationally efficient and highly effective solution for analyzing ECG data.
• This approach holds substantial potential for improving cardiovascular disease diagnosis and patient outcomes through automated monitoring.