Automatic 12-Leading Electrocardiogram Classification Network with Deformable Convolution.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Electrocardiogram (ECG) is crucial for monitoring heart physiology, utilizing 12 leads for comprehensive assessment.
• Conventional Convolutional Neural Networks (CNNs) struggle to capture inter-lead correlations in ECG data.
• Limitations in existing deep learning models necessitate advanced feature extraction for accurate ECG analysis.

Purpose of the Study:

• To propose a novel neural network structure, DCNet, based on Deformable Convolution Networks (DCN) for enhanced ECG feature detection.
• To improve the extraction of data correlations across multiple ECG leads.
• To achieve superior diagnostic accuracy in ECG signal classification tasks.

Main Methods:

• Development of DCNet, a neural network architecture comprising four DCN blocks and a classification layer.
• Integration of deformable convolution with normal convolution within DCN blocks to optimize feature learning.
• Evaluation using the public CPSC-2018 dataset for 12-leading ECG classification.

Main Results:

• DCNet demonstrated superior capability in extracting inter-lead characteristics from ECG signals.
• The proposed architecture achieved the highest diagnostic accuracy on the CPSC-2018 dataset, reaching 86.3%.
• DCNet outperformed other common network architectures in ECG signal classification.

Conclusions:

• DCNet offers an effective automatic ECG classification model by leveraging the feature learning capabilities of DCN.
• The model's ability to capture inter-lead correlations significantly enhances diagnostic accuracy.
• This approach has the potential to reduce the workload for medical staff in ECG interpretation.