生成对抗神经网络在心电异常识别中的应用研究

Application of Generative Adversarial Networks in Abnormal ECG Recognition

针地在使用深度学习方法构建心电异常识别模型,常常由于心电异常事件样本分布不平衡,造成机器识别心电异常事件模型性能表现差的问题,论文提出一种基于生成对抗网络的数据增强方法来获取均匀分布的训练数据集,其主要过程如下:首先使用小波变换消除心电信号中噪声,然后使用压缩感知模型压缩心电信号来减少网络中的参数,利用生成对抗神经网络模型扩充数据集,最后使用卷积神经网络建立分类模型。实验结果表明,使用对抗神经网络能够显著改善数据集中样本分布不均衡的问题,平均F1达到了98.73%。引入压缩感知模型后,在不影响模型性能表现的情况下,将模型参数量减少了28.30%。基于对抗神经网络的数据增强方法可以有效地解决心电异常分类过程样本分布不均衡,利用压缩感知模型方法不仅可以保证模型性能,同时降低了模型的复杂程度。

To solve the problem that the deep learning method is challenging to obtain high precision ECG anomaly recogni⁃tion model due to unbalanced training dataset of abnormal ECG events,the dataset augmentation method based on GANs is pro⁃posed to solve the problem caused by uniformly distributed training data sets.The main framework are as follows.Firstly,the ECG signals are de-noised and compressed by the compressed sensing model,then the data set is expanded by GANs.Finally,the neu⁃ral network model is applied for classification.The results show that the adversarial neural network can significantly improve the problem of unbalanced sample distribution in the data set,and the average F1 reaches 98.73%.With the help of the compressed sensing model,the total parameters in the proposed model reduce by 28.30%without the loss of the model performance.The data en⁃hancement method based on the adversarial neural network can effectively solve the uneven distribution of samples in the ECG ab⁃normal classification process.Using the compressed sensing model method can not only ensure the performance of the model,but al⁃so reduce the complexity of the model.