Explainable detection of myocardial infarction using deep learning models with Grad-CAM technique on ECG signals.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Myocardial infarction (MI) is a leading cause of global mortality.
• Accurate electrocardiography (ECG) is critical for timely MI diagnosis and emergency interventions.
• Machine learning (ML) shows promise for automated computer-aided ECG diagnosis of cardiovascular diseases.

Purpose of the Study:

• To develop and compare DenseNet and Convolutional Neural Network (CNN) models for classifying healthy subjects and ten types of MI based on ECG data.
• To utilize Grad-CAM for visualizing influential ECG leads and signal portions in model predictions.
• To enhance clinical acceptance of ML models in MI diagnosis through explainability.

Main Methods:

• ECG signals from the Physikalisch-Technische Bundesanstalt database were pre-processed.
• ECG beats were extracted using R peak detection and fed into DenseNet and CNN models.
• Grad-CAM was applied to visualize model decision-making processes.

Main Results:

• Both DenseNet and CNN models achieved over 95% classification accuracy.
• DenseNet demonstrated superior performance due to feature reusability, lower computational complexity, and higher accuracy.
• Lead V4 was identified as the most activated lead across both models; specific lead activations for each MI class were established.

Conclusions:

• DenseNet is a preferred model for multiclass MI classification from ECGs, offering high accuracy and efficiency.
• Explainability through Grad-CAM visualization aids in understanding model predictions, potentially increasing clinical trust and adoption.
• These explainable ML models have the potential for implementation in hospital and remote settings for ECG-based MI triage.