Electrocardiogram Beat-Classification Based on a ResNet Network.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electrocardiography (ECG) is crucial for analyzing heart electrical activity.
• Deep learning models have shown significant promise in classifying heartbeats accurately.
• Existing methods require robust algorithms for precise arrhythmia detection.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying heartbeats into four categories: normal, atrial premature contraction, premature ventricular contraction, and others.
• To leverage a ResNet architecture with 1D convolutional layers for enhanced ECG signal processing.
• To assess the model's performance using established arrhythmia databases.

Main Methods:

• Implementation of a 1D convolutional ResNet deep learning architecture.
• Training and validation using the MIT-BIH Arrhythmia Database.
• Comparative analysis of optimization algorithms, specifically Stochastic Gradient Descent (SGD) and Adaptive Moment Estimation (Adam).

Main Results:

• The ResNet model achieved 96% accuracy with SGD and 83% accuracy with Adam.
• SGD optimization resulted in F1-scores exceeding 90% across all four classified heartbeat types.
• Testing on a larger, unseen dataset indicated a need for further model refinement.

Conclusions:

• The proposed deep learning model demonstrates strong performance in classifying cardiac arrhythmias from ECG data.
• Stochastic Gradient Descent (SGD) proved to be a more effective optimizer for this specific task compared to Adam.
• Further research and dataset expansion are recommended to enhance the model's accuracy and generalizability for real-world clinical applications.