EvoMBN: Evolving Multi-Branch Networks on Myocardial Infarction Diagnosis Using 12-Lead Electrocardiograms

Area of Science:

• Artificial Intelligence
• Cardiology
• Biomedical Engineering

Background:

• Multi-branch Networks (MBNs) are used for myocardial infarction (MI) diagnosis via electrocardiograms.
• Fixed architectures limit the diagnostic accuracy of current MBNs.

Purpose of the Study:

• To propose an evolving neural network (EvoMBN) for optimizing MBN architectures for MI diagnosis.
• To enhance the accuracy and generalization of MI diagnostic models.

Main Methods:

• Utilized a genetic algorithm (GA) for automatic MBN architecture optimization.
• Developed a novel fixed-length encoding for MBN architectures.
• Introduced a Lead Squeeze and Excitation (LSE) block for feature summarization.
• Conducted five-fold cross-validation on PTB and PTB-XL databases.

Main Results:

• EvoMBN demonstrated superior generalization capabilities compared to existing methods.
• The optimized flexible architecture proved efficient for auxiliary MI diagnosis.
• Successfully transferred learned model architecture between PTB and PTB-XL databases.

Conclusions:

• EvoMBN offers a flexible and effective approach to MI diagnosis by optimizing network architectures.
• The proposed method shows significant potential for real-world clinical decision support in cardiology.