ECGNet: An Efficient Network for Detecting Premature Ventricular Complexes Based on ECG Images.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate preoperative prediction of premature ventricular complexes (PVCs) origin is crucial for surgical success.
• Current prediction methods for PVCs lack efficiency and accuracy, especially those using electrocardiogram (ECG) images and deep learning.
• There is a need for improved deep learning-based prediction models for ECG images.

Purpose of the Study:

• To develop and evaluate ECGNet, a novel neural network for classifying 12-lead ECG images to predict PVCs origin.
• To improve the accuracy and efficiency of PVCs origin prediction using deep learning on ECG data.
• To address the challenge of sample imbalance in medical datasets.

Main Methods:

• ECGNet was developed using 609 ECG images from 310 patients.
• The network incorporates dense blocks, specialized convolution kernels, and divergent paths for enhanced performance.
• A novel loss function was designed to mitigate issues arising from imbalanced sample distribution.

Main Results:

• ECGNet achieved a high prediction accuracy of 91.74% even with small datasets.
• The proposed loss function effectively handled sample imbalance during training.
• ECGNet demonstrated superior performance compared to established networks like ResNet and DarkNet.

Conclusions:

• ECGNet offers a rapid and accurate method for multiclassification of PVCs origin from limited ECG data.
• The developed model shows potential as a valuable tool for preoperative diagnosis by clinicians.
• Future work will focus on expanding the dataset and refining the network architecture to further enhance prediction accuracy.