A Novel Dilated Causal Convolutional and Transformer Attention Prototype Network with Cut Feature Map for ECG Classification

Deep learning is an important means of artificial intelligence, and ECG classification has received extensive attention in the field of data mining. Dilated causal convolution can handle long-term dependencies with exponentially large receptive fields without computational explosion. Transformer is a neural network that improves model parallelism through an attention mechanism, which breaks the limitation that recurrent neural networks cannot be parallelized, and has powerful feature extraction capabilities. We innovatively propose a novel Transformer neural network based on cutting multi-channel feature map for ECG classification. After the raw univariate ECG is modeled by dilated causal convolutions, cut the feature map into some non-overlapping patches and encode the patch set as a sequence of tokens. Correlations between patches are obtained in parallel, and prototype learning is used to improve model performance. We conduct experiments using ECG datasets selected from the well-known UCR archives and achieve the best accuracy compared to several baselines.