Detection of Myocardial Infarction Using ECG and Multi-Scale Feature Concatenate

Area of Science:

• Cardiology and Artificial Intelligence
• Medical Signal Processing
• Machine Learning for Healthcare

Background:

• Computer-aided diagnosis systems using convolutional neural networks (CNNs) are crucial for automating myocardial infarction (MI) detection from electrocardiograms (ECGs).
• Existing systems often neglect overfitting and underfitting issues, leaving network complexity uncertain.
• The need for robust and efficient models for early MI diagnosis and prevention is paramount.

Purpose of the Study:

• To develop and evaluate novel CNN architectures for automated MI detection and localization using ECG data.
• To address the challenges of network complexity and optimize model performance for clinical application.
• To create compact models suitable for integration into wearable devices for continuous monitoring.

Main Methods:

• Development of a narrow network (N-Net) with a simple structure (two convolutional layers per lead branch).
• Implementation of a multi-scale features-concatenate network (MSN-Net) for extracting larger features via signal pooling.
• Systematic tuning of convolutional layer filters and signal scales to determine optimal network structures.

Main Results:

• The N-Net achieved 95.76% accuracy in MI detection, while the MSN-Net attained 61.82% accuracy in MI localization.
• Both models demonstrated statistically significant improvements (p < 0.001) compared to state-of-the-art methods.
• The developed networks are notably smaller, making them suitable for offline monitoring on wearable devices.

Conclusions:

• Testing diverse network structures, from simple to complex, is essential for effective MI detection.
• The developed CNN models offer a promising approach for accurate and efficient MI diagnosis.
• Further research is needed to address class imbalance and enhance feature extraction quality.