欢迎访问《中国循证儿科杂志》官方网站，今天是
English Version
中国循证儿科杂志

中国循证儿科杂志 ›› 2024, Vol. 19 ›› Issue (6): 464-468.DOI: 10.3969/j.issn.1673-5501.2024.06.011

• 论著 • 上一篇    下一篇

基于变分模态分解和深度学习融合的心电信号分类诊断模型的构建和验证

张英，党艳，梁雪村   

  1. 国家儿童医学中心，复旦大学附属儿科医院心血管中心上海，201102
  • 收稿日期:2024-04-27 修回日期:2024-08-26 出版日期:2024-12-25 发布日期:2024-12-25
  • 通讯作者: 梁雪村

Construction and validation of an ECG signal classification model based on the fusion of variational mode decomposition and deep learning

ZHANG Ying, DANG Yan, LIANG Xuecun   

  1. National Children's Medical Center Cardiovascular Center in Children's Hospital of Fudan University, Shanghai 201102, China
  • Received:2024-04-27 Revised:2024-08-26 Online:2024-12-25 Published:2024-12-25
  • Contact: LIANG Xuecun

PDF

84

摘要： 背景 传统的心电图（ECG）诊断方法受主观因素和经验影响较大，且ECG信号采集过程中难免会受到各种噪声干扰。近年来，基于人工智能对ECG信号进行诊断已成为研究热点。 目的 探讨基于变分模态分解（VMD）和深度学习融合的ECG信号分类诊断的可行性。 设计 诊断准确性研究。 方法 ECG信号数据来源为2020年美国查普曼大学和绍兴人民医院联合发布的12导联心电信号开源数据库，心律类型归纳为心房纤颤（AFIB）、广义的室上性心动过速（GSVT）、窦性心动过缓（SB）和窦性心律（SR）四类，均按4∶1随机划分为训练集和测试集。利用VMD进行模态分解，剔除高频模态，得到心律失常ECG信号有效特征数据集。先用未降噪ECG数据比较4种典型深度神经网络模型（LeNet、AlexNet、VggNet16、ResNet50）的分类性能，根据损失率和准确率选取最优模型后，再用该分类模型对VMD降噪后的数据进行训练和测试。 主要结局指标 准确率、精准率、敏感度、特异度、F1值和约登指数。 结果 4种典型深度神经网络模型中ResNet50模型的准确率最高且损失率最低，分类效果最好。经VMD滤波后，ECG信号的噪声明显下降。VMD降噪后可以使ResNet50构建的分类诊断网络对4类心律类型总的分类准确率由93.75%±0.24%升高至94.41%±0.18%，使F1值（＞0.92）和约登指数（＞0.90）均提高，精准率除SR外、敏感度除SB外均提高，特异度GSVT和SB均有提高。 结论 基于VMD降噪和深度卷积神经网络融合的ECG信号分类诊断模型的分类诊断性能良好。

关键词: 变分模态分解, 深度学习, 心电信号, 分类诊断

Abstract: Background Traditional electrocardiogram (ECG) diagnostic methods are significantly influenced by subjective factors and experience, and ECG signals are inevitably subject to various noise interferences during the acquisition process. In recent years, the use of artificial intelligence for ECG signal diagnosis has become a research hotspot. Objective To explore the feasibility of ECG signal classification and diagnosis based on the integration of Variational Mode Decomposition (VMD) and deep learning. Design Research on diagnostic accuracy. Methods The ECG signal dataset used in this paper originates from a 12-lead ECG open-source database jointly released by Chapman University in the United States and Shaoxing People's Hospital in 2020. The cardiac rhythms are categorized into four types: atrial fibrillation (AFIB), generalized supraventricular tachycardia (GSVT), sinus bradycardia (SB), and sinus rhythm (SR). This dataset is randomly split into training and testing sets at a 4∶1 ratio. Variational Mode Decomposition (VMD) is employed for modal decomposition, with high-frequency modes removed to obtain an effective feature dataset of arrhythmic ECG signals. Initially, the classification performance of four typical deep neural network models—LeNet, AlexNet,VggNet16, and ResNet50—is compared using the non-denoised ECG data and the optimal network is selected based on loss rate and accuracy. Subsequently, this classification network model is used to train and test on the data after VMD denoising. Main outcome measures Accuracy, precision, sensitivity, specificity, F1 score, and Youden's index. Results Among the four typical deep neural network models, the ResNet50 model exhibits the highest accuracy, the lowest loss rate, and the best classification performance. After VMD filtering, the noise in the ECG signals is significantly reduced. Following VMD denoising, the classification diagnostic network model built with ResNet50 improves the overall classification accuracy of the four types of cardiac rhythms from 93.75%±0.24% to 94.41%±0.18%, with enhancements in both the F1 score (>0.92) and Youden's index (>0.90). Except for SR in precision and SB in sensitivity, all other metrics show improvements, and specificity increases for both GSVT and SB. Conclusion The ECG signal classification and diagnosis model based on the fusion of VMD denoising and deep convolutional neural networks exhibits good classification and diagnosis performance.

Key words: Variational mode decomposition, Deep learning, ECG signal, Classification diagnosis