用于心律失常识别的LSTM和CNN并行组合模型

心脏病是导致人类死亡的首要原因,而大部分心血管疾病往往伴随着心律失常出现.为实现对不同类型心电图信号的自动分析、识别异常心律,研究并提出了一种基于深度学习的心拍分类算法.考虑到心电图自身的特性,使用长短时记忆网络来捕捉心电序列数据的前后依赖关系,结合卷积神经网络提取局部相关特征,自动识别5种不同类型的心拍.基于LSTM和CNN的深度学习方法将经过预处理的心电信号后直接作为网络的输入,将心电分类的特征提取、分类两个步骤融合为单个学习器.针对类别数据不平衡问题,通过对少数类样本滑窗采样获得更多训练数据.使用MIT-BIH数据集验证模型的有效性,最终在测试集2万多个心拍记录中分类结果准确率达到99.11%,特异性为99.44%,灵敏度为97.27%,此外滑窗采样操作对少数类样本的灵敏度有明显改善.实验结果表明,相比传统的SVM和随机森林等方法,基于LSTM和CNN的并行组合模型不需要人工提取复杂特征,且达到了更好的分类性能,适合用于可穿戴式心电设备和远程监护领域.

Arrhythmia classification using parallel combination of LSTM and CNN

Heart disease is the leading cause of death in humans, and most cardiovascular diseases are accompanied by arrhythmias. In order to realize the automatic analysis of different types of electrocardiogram (ECG) signals and recognize abnormal heart rhythm, a new classification algorithm based on deep learning was studied and proposed. Considering the characteristics of the EGG, the convolutional neural network (CNN) was used to extract the local correlation features, and the long-short term memory (LSTM) network was used to capture the long-term dependence of ECG sequence data to identify five different types of heart beats automatically. The deep learning method based on LSTM and CNN directly took the preprocessed ECG signals as the input of the network, and integrated the feature extraction and ECG classification into a single learner. In terms of the problem of imbalance, sampling by sliding window was performed on minority class data to get more training data. The effectiveness of the algorithm was evaluated with the MIT-BIH arrhythmia dataset, and the accuracy, specificity, and sensitivity of the classification results in more than 20 000 cardiac beats recorded in the test set reached 99.11%, 99.44%, and 97.27%, respectively. In addition, the operation of sliding window sampling significantly improved the sensitivity of minority class. The experimental results show that compared with the traditional methods, the parallel combination model based on LSTM and CNN did not require separate feature extraction steps and achieved better classification performance, which is suitable for wearable ECG devices and remote monitoring field.