MADNN: A Multi-scale Attention Deep Neural Network for Arrhythmia Classification

The morphological ECG features for arrhythmia diagnosis are usually identified and combined on different scales. For example, morphological ECG features can be identified on the scale of length or amplitude of QRS waves. Professionals can then make a diagnosis based on the combination of these identified features. Attention-based deep neural networks have been proved to boost meaningful features on different scales and suppress weak features. To boost and combine ECG features on different scales for arrhythmia classification, we proposed MADNN: a multi-scale attention deep neural network for arrhythmia classification. Our proposed network was designed combining kernel-wise and branch-wise attention modules based on a backbone of 1-dimensional convolutional neural networks. MADNN with properly tuned hyperparameters was tested for arrhythmia classification in the PhysioNet/Computing in Cardiology Challenge 2020. In this challenge, MADNN officially achieved a validation score of 0.446, and a full test set score of 0.236. Our team named Minibus ranked the 18
^th
out of 41 teams.