A Weighted Graph Attention Network Based Method for Multi-label Classification of Electrocardiogram Abnormalities.

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Multi-label electrocardiogram (ECG) classification is crucial for diagnosing concurrent cardiac abnormalities.
• Accurate modeling of dependencies between different cardiac abnormalities is essential for improving classification performance.

Purpose of the Study:

• To propose a novel multi-label ECG classification method that effectively models cardiac abnormality dependencies.
• To enhance the accuracy of automated ECG interpretation through improved dependency modeling.

Main Methods:

• Developed a weighted graph attention network approach for multi-label ECG classification.
• Represented cardiac abnormality dependencies as edge weights in a graph where each node is a class.
• Introduced a new method for generating edge weights by combining self-attentional weights and prior co-occurrence knowledge.

Main Results:

• Evaluated the algorithm on a dataset for 34 types of ECG abnormalities.
• Achieved a micro-f1 score of 91.45% and a macro-f1 score of 44.48% in cross-validation.
• Demonstrated that the proposed method successfully models class dependencies.

Conclusions:

• The proposed weighted graph attention network method significantly improves multi-label ECG classification performance.
• Effective modeling of cardiac abnormality dependencies is key to advancing automated ECG diagnosis.
• This approach offers a promising direction for clinical diagnosis support using ECG data.