Automatic arrhythmia detection based on electrocardiogram (ECG) is of great significance for early prevention and diagnosis of cardiac diseases. Recently, deep learning methods have been applied to arrhythmia detection and obtained great success. Among them, convolutional neural network (CNN) is an effective method for extracting features due to its local connectivity and parameter sharing. In addition, recurrent neural network (RNN) is another commonly used method, which is applied to process time-series signal. The stacking of both CNN and RNN has been proved to be more effective in multi-class arrhythmia detection. However, these networks ignored the fact that different channels and temporal segments of a feature map extracted from the 12-lead ECG signal contribute differently to cardiac arrhythmia detection, and thus, the classification performance could be greatly improved. To address this issue, spatio-temporal attention-based convolutional recurrent neural network (STA-CRNN) is proposed to focus on representative features along both spatial and temporal axes. STA-CRNN consists of CNN subnetwork, spatio-temporal attention modules and RNN subnetwork. The experiment result shows that, STA-CRNN reaches an average F₁ score of 0.835 in classifying 8 types of arrhythmias and normal rhythm. Compared with the state-of-the-art methods based on the same public dataset, STA-CRNN achieves an obvious improvement on identifying most of arrhythmias. Also, it is demonstrated by visualization that the learned features through STA-CRNN are in line with clinical judgement. STA-CRNN provides a promising method for automatic arrhythmia detection, which has a potential to assist cardiologists in the diagnosis of arrhythmias.

基于心电图（ECG）的自动心律失常检测对于心脏疾病的早期预防和诊断具有重要意义。最近，深度学习方法已应用于心律失常检测并取得了巨大成功。其中，卷积神经网络（CNN）由于其局部连通性和参数共享，是一种有效的特征提取方法。此外，循环神经网络（RNN）是另一种常用的方法，用于处理时间序列信号。CNN 和 RNN 的叠加已被证明在多类心律失常检测中更有效。然而，这些网络忽略了一个事实，即从 12 导联 ECG 信号中提取的特征图的不同通道和时间段对心律失常检测的贡献不同，因此，分类性能可以大大提高。为了解决这个问题，提出了基于时空注意力的卷积递归神经网络（STA-CRNN）来关注空间和时间轴上的代表性特征。STA-CRNN 由 CNN 子网络、时空注意力模块和 RNN 子网络组成。实验结果表明，STA-CRNN达到了平均F ₁分 0.835 分 8 种类型的心律失常和正常心律。与基于相同公共数据集的最新方法相比，STA-CRNN 在识别大多数心律失常方面取得了明显的进步。此外，通过可视化证明，通过 STA-CRNN 学习到的特征符合临床判断。STA-CRNN 为自动心律失常检测提供了一种很有前景的方法，它有可能帮助心脏病专家诊断心律失常。