Key result
The 20-layer squeeze-excitation residual network (SE-ResNet) achieved an overall accuracy of 98.71% for arrhythmia detection, with a sensitivity of 99.12% and specificity of 99.59% for premature ventricular contractions.
Why the study?
Deep learning methods for automatic ECG detection face challenges like overfitting and gradient issues, while dynamic ECGs are often contaminated by noise and morphological variation.
Does a 20-layer squeeze-extraction residual network (SE-ResNet) improve the detection accuracy of premature ventricular contractions on 12-lead dynamic ECGs compared to other deep learning models?
Does a 20-layer squeeze-extraction residual network (SE-ResNet) improve the detection accuracy of premature ventricular contractions on 12-lead dynamic ECGs compared to other deep learning models?
Absolute Event Rate: 98.71% vs 97.98%
A novel 20-layer squeeze-extraction residual network (SE-ResNet) demonstrates high accuracy (98.71%) for real-time detection of premature ventricular contractions on 12-lead dynamic ECGs, outperforming standard deep learning models.
Promising SE-ResNet accuracy for arrhythmia detection warrants caution in deployment; leaves open prospective validation before clinical use.
Premature ventricular contraction (PVC) is a very common arrhythmia that can originate in any part of the ventricle and is one of the important causes of sudden cardiac death. Timely and rapid detection of PVC on dynamic electrocardiogram (ECG) recording for patients with cardiovascular diseases is of great significance for clinical diagnosis. Furthermore, it can facilitate the planning and execution of radiofrequency ablation. But the dynamic ECGs can be easily contaminated by various noises and its morphological characteristics show significant variations for different patients. Though the deep learning methods achieved outstanding performance in ECG automatic recognition, there are still some limitations, such as overfitting, gradient disappearance or gradient explosion in deep networks. Therefore, a residual module is constructed using the squeeze-excitation method to alleviate the problems. A 20-layer squeeze-extraction residual network (SE-ResNet) containing multiple squeeze-extraction modules was designed for real-time PVC detection on 12-lead dynamic ECG. The algorithm was evaluated using the dynamic 12-lead ECGs in INCART database (168,379 heartbeats in total). The experimental results show that the test accuracy of the method proposed in this paper is 98.71%, and the specificity and sensitivity of PVC are 99.12% and 99.59%, respectively. Under the same dataset and experimental platform, the average recognition accuracy of our proposed method is increased by 0.73%, 1.55%, 2.9% and 1.65% compared with the results obtained by CNN, Inception, AlexNet and deep multilayer perceptron, respectively. The proposed scheme provides a new method for real-time detection of PVC on dynamic 12-lead ECGs. The experiment results show that the proposed method outperforms state-of-the-art methods, and has good potential for clinical applications.
No takes yet. Share an insight, caveat, or question.
Li et al. (2022) studied Premature Ventricular Contractions (n=168,379). Squeeze-Excitation Residual Network (SE-ResNet) vs. CNN, Inception, AlexNet, and deep multilayer perceptron was evaluated on Test accuracy for arrhythmia detection. The 20-layer squeeze-excitation residual network (SE-ResNet) achieved an overall accuracy of 98.71% for arrhythmia detection, with a sensitivity of 99.12% and specificity of 99.59% for premature ventricular contractions.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: