Developing Graph Convolutional Networks and Mutual Information for Arrhythmic Diagnosis Based on Multichannel ECG...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cardiovascular diseases, including arrhythmia, are leading global causes of death.
• Electrocardiogram (ECG) analysis is crucial for diagnosing these conditions and reducing human error.
• Existing ECG diagnostic methods require improvement in accuracy and classification capabilities.

Purpose of the Study:

• To enhance the accuracy of arrhythmia diagnosis using a novel graph convolutional network (GCN).
• To classify various types of arrhythmias by leveraging mutual information (MI) indices from ECG leads.
• To introduce a new GCN methodology that incorporates MI-derived relationships between ECG leads.

Main Methods:

• A novel graph convolutional network (GCN) model was developed, integrating mutual information (MI) indices.
• MI was used to represent the relationships between 12 ECG leads as an adjacency matrix for the GCN.
• The methodology was validated on a large ECG database using cross-validation for training and testing.
• A 15-layer GCN-MI structure was identified as the optimal model.

Main Results:

• The proposed GCN-MI model achieved high classification accuracy for heart rhythm types.
• Key performance indicators included sensitivity (98.45%), precision (97.89%), specificity (99.85%), and accuracy (99.71%).
• The GCN-MI approach demonstrated superior performance compared to a GCN using an identity adjacency matrix (GCN-Id).

Conclusions:

• Incorporating mutual information (MI) to measure cardiac lead relationships significantly improves GCN performance in arrhythmia detection and classification.
• The novel GCN-MI method offers a more accurate and effective approach for diagnosing arrhythmias from ECG data.
• This research highlights the potential of advanced AI techniques for improving cardiovascular disease diagnostics.