A Deep Multi-scale Convolutional Neural Network for Classifying Heartbeats

The electrocardiogram (ECG) is a very important tool to reflect the health of the human heart. There are many cardiac abnormalities which can be diagnosed from ECG data. In our paper, we design a 15-layer multi-scale convolutional neural network (CNN) which can map ECG data and RR intervals to the corresponding rhythm classes. One of the key points of the proposed model is that the multi-scale convolution block enables the network extract scale-relevant features of heartbeats, which is effective in practice. Another key point is that shortcut connections are employed to avoid the loss of information as the network depth increases. Furthermore, we employ RR interval as dynamic features and concatenate them with the morphological features extracted by the multi-scale CNN model as the final heartbeat features for classification. We use the open source PhysioBank MIT-BIH Arrhythmia database to train and evaluate ECG algorithms. In "class-based" strategy, the recognition accuracy rate reaches 98.32%, while in the "subject-based" strategy, the accuracy is 93.9%, which exceed the performance of most existing classification methods.