SwinDAE: Electrocardiogram Quality Assessment Using 1D Swin Transformer and Denoising AutoEncoder

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Electrocardiogram (ECG) signal analysis is vital across many fields.
• Accurate ECG quality assessment is crucial for reliable diagnostics.
• Current deep learning models for ECG quality assessment lack generalizability across datasets.

Purpose of the Study:

• To propose an effective deep learning model for robust ECG signal quality assessment.
• To enhance the generalization capability of ECG quality assessment models across different sensors and collection scenarios.
• To introduce a novel waveform component localization loss for improved feature learning.

Main Methods:

• Developed the Swin Denoising AutoEncoder (SwinDAE) model, integrating a 1D Swin Transformer into a Denoising AutoEncoder architecture.
• Pre-trained SwinDAE on the PTB-XL dataset using signal reconstruction and quality assessment loss, incorporating a novel waveform component localization loss.
• Fine-tuned the model on the BUT QDB dataset for specialized quality assessment.

Main Results:

• SwinDAE achieved a 0.02-0.13 mean F1 score improvement on the BUT QDB dataset compared to existing deep learning methods.
• Demonstrated significant applicability and superior performance on two additional independent datasets.
• Statistical analysis confirmed the model's performance significance and prediction rationality.

Conclusions:

• The proposed SwinDAE exhibits strong generalization ability, outperforming state-of-the-art deep learning methods in ECG quality assessment.
• The model effectively learns common features of high-quality ECG signals.
• SwinDAE shows excellent potential for cross-sensor and cross-collection scenario applications.