Atrial fibrillation classification based on convolutional neural networks.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Global mortality rates for atrial fibrillation (AF) have significantly increased over the past four decades.
• Accurate and timely diagnosis of AF is crucial for patient outcomes.
• Electrocardiogram (ECG) data is a primary tool for diagnosing cardiac conditions like AF.

Purpose of the Study:

• To evaluate the effectiveness of convolutional neural networks (CNNs) in diagnosing atrial fibrillation (AF) using ECG data.
• To compare the performance of Alex networks and residual networks for AF classification.
• To identify the optimal CNN architecture for accurate AF diagnosis in a general hospital setting.

Main Methods:

• Utilized ECG data from 20,000 patients (10,000 normal sinus rhythm, 10,000 AF) from Anam Hospital, Seoul.
• Applied and compared 30 CNN models: 6 Alex networks and 24 residual networks with varying configurations.
• Assessed models based on accuracy, parameter count, and training time for classifying normal sinus rhythm versus AF.

Main Results:

• The best performing residual network achieved an accuracy of 0.999 with 248,418 parameters and a training time of 253 seconds.
• The best Alex network achieved an accuracy of 0.997 with 5,268,818 parameters and a training time of 89 seconds.
• Residual network performance generally improved with increased depth (number of residual blocks).

Conclusions:

• Residual networks demonstrate superior performance in terms of accuracy and parameter efficiency compared to Alex networks for AF diagnosis.
• Deep learning models, particularly residual networks, show significant potential for improving the accuracy and efficiency of AF diagnosis.
• The findings suggest residual networks are a viable and effective tool for clinical AF detection using ECG data.