ECG autoencoder based on low-rank attention.

Area of Science:

• Cardiology
• Machine Learning
• Signal Processing

Background:

• Cardiovascular disease (CVD) is a leading cause of mortality globally.
• Electrocardiogram (ECG) analysis is crucial for diagnosing CVD.
• Current machine learning models often overlook spatial features in ECG signals.

Purpose of the Study:

• To propose a novel ECG autoencoder network incorporating low-rank attention (LRA-autoencoder).
• To capture and leverage spatial dimension features in ECG signals for improved diagnostic accuracy.
• To enhance the differentiation of features among different cardiovascular disease categories.

Main Methods:

• Developed an LRA-autoencoder architecture to interpret ECG signals spatially.
• Utilized a low-rank attention block (LRA-block) with singular value decomposition to extract and weight spatial features.
• Employed a ResNet-18 network classifier for performance evaluation on benchmark datasets.

Main Results:

• The LRA-autoencoder achieved a mean accuracy of 0.997 on the MIT-BIH Arrhythmia dataset.
• On the PhysioNet Challenge 2017 dataset, the method obtained a mean accuracy of 0.850 and a mean F1-score of 0.843.
• Experimental results demonstrate superior classification performance compared to existing methods.

Conclusions:

• The proposed LRA-autoencoder effectively captures spatial features in ECG signals.
• This approach significantly enhances the accuracy of cardiovascular disease classification using machine learning.
• The method shows promise for improving automated ECG diagnostic tools.