MMDN: Arrhythmia detection using multi-scale multi-view dual-branch fusion network.

• Our proposed MMDN can extract features from ECG signals in multiple dimensions, which can improve arrhythmia classification performances effectively by fusing these features. • The proposed MSTA can extract multi-timescale features from ECG signals, leading to improve arrhythmia classification accuracy. • The auxiliary information was introduced into the classification network to enhance the arrhythmia classification performances through using the relationship between different types of arrhythmias and auxiliary information. Automatic arrhythmia classification plays an important role in preventing cardiac death. Due to the intricate multi-periodic patterns inherent in arrhythmias, how to improve the classification accuracy is a challenging problem. In this paper, a multi-scale multi-view dual-branch fusion network (MMDN) is proposed to implement accurate and interpretable arrhythmia classification by fusing features at different levels. The proposed MMDN method consists of three parts: a multi-view block, an additional information fusion block, and a feature fusion block. The multi-view block employs channel, spatial, and the proposed multi-scale temporal attention module to extract anomalous features in raw data from diverse perspectives. Subsequently, the output of the multi-view block is fed into an additional information fusion block, which enhances features by incorporating auxiliary information such as age and gender. The feature fusion block combines the output to produce recognition results using a multi-layer perceptron. Signal Challenge 2018 database (CPSC 2018 DB) is used to validate the classification performances of the proposed MMDN method. Experimental results demonstrate that MMDN outperforms current state-of-the-art methods for ECG classification tasks, with an accuracy of 0.861, a recall of 0.844, and an F1 score of 0.850. [ABSTRACT FROM AUTHOR]