Classification Method of ECG Signals Based on RANet.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electrocardiograms (ECG) are crucial for assessing heart health and diagnosing arrhythmias.
• Deep learning models for ECG classification face challenges like gradient vanishing in deeper networks.
• Varying significance of ECG signal channels and periods complicates abnormality identification.

Purpose of the Study:

• To develop an advanced ECG classification method mitigating gradient vanishing and improving feature relevance.
• To enhance the accuracy of detecting various ECG abnormalities using deep learning.
• To address data imbalance issues in ECG datasets.

Main Methods:

• A residual attention neural network was proposed, combining Residual Network (ResNet) for gradient control and an attention mechanism.
• ResNet was employed for its ability to prevent gradient vanishing and its simpler structure with fewer parameters.
• An attention mechanism was integrated to focus on critical ECG information and channel features, alongside improved voting for data imbalance.

Main Results:

• Experiments on the PhysioNet/CinC Challenge 2017 dataset demonstrated the model's effectiveness.
• The proposed method achieved an average F1 score of 0.817, a 0.064 improvement over the standard ResNet model.
• The model's performance was found to be excellent compared to other mainstream methods.

Conclusions:

• The residual attention neural network effectively addresses key challenges in deep learning-based ECG classification.
• This approach offers superior performance in arrhythmia detection compared to existing methods.
• The findings suggest a promising direction for improving automated ECG analysis and diagnosis.