【摘要】 心电信号分类是医疗保健领域的重要研究内容。心电信号数据是类不平衡数据，不同类别的心律失常依赖于心电图的长期变化特征，局部变化特征及其相对位置。针对大多数方法不能较好地解决数据类不平衡，且未考虑特定波形重要性等问题，提出一种基于位置注意力机制的混合神经网络心电信号分类(DCLB)算法。首先，利用深度卷积生成对抗网络扩充数量少的类别样本，从而解决类不平衡问题；其次，利用二维卷积神经网络和双向长短期记忆网络进行特征提取，从而获得心电信号的局部变化特征和长期变化特征；然后，在每个二维卷积神经网络后嵌入位置注意力机制，从而提高关键位置特征的重要程度；最后，利用全连接网络输出分类结果。对MIT-BIH心律失常数据集中的30 584个样本的实验结果表明，DCLB算法的平均准确率为98.79%,敏感性为94.21%,特异性为98.98%,阳性预测值为93.70%。该模型可以有效提取心电信号特征，适用于监测系统中心律失常疾病的诊断。更多还原

【Abstract】 Electrocardiogram(ECG) signal classification is significant research issue in the healthcare field. Signal data from ECG are classification imbalanced, and different classification of arrhythmias depend on long-term variation features, local variation features and their relative location of electrocardiogram. Most existing methods are not able to solve the classification imbalance problem well and consider the importance of specific waveforms. In this study, a hybrid neural network algorithm based on the location attention mechanisms was proposed for classifying ECG signals, referred to as DCLB algorithm. Firstly, the small-size classification samples were augmented adopting the deep convolutional generative adversarial networks(DCGAN) to solve the classification imbalance problem. Secondly, the local variation features and long-term variation features of ECG signals were extracted utilizing two-dimensional convolutional neural networks(2DCNN) and bi-directional long short-term memory network(BLSTM). Next, the location attention mechanisms(LAM) were nested behind each 2DCNN for enhancing the effects of key location features. Finally, the classification results were output using the fully connected neural networks. Experimental results on 30 584 samples of the MIT-BIH arrhythmia database showed that the proposed algorithm achieved the average accuracy of 98.79%, sensitivity of 94.21%, specificity of 98.98%, and positive predictive value of 93.70%. respectively. The results indicated that DCLB was able to extract effectively ECG signal features and suitable for the diagnosis of arrhythmia in the monitoring system.更多还原