Detection Of Arrhythmia Based On Multi-scale Fusion Neural Network

Arrhythmia classification plays an important role in the diagnosis of heart disease,and there are common early detection methods for arrhythmia such as electrocardiogram（ECG）detection.Manual interpretation of ECG information is not only timeconsuming and labor-intensive,but also requires professional ECG knowledge and rich experience in ECG disease diagnosis,so it is very important to detect arrhythmia automatically.In addition,Convolutional Neural Network（CNN）is an automatic classification algorithm for multi-class arrhythmias that is simpler and more noise-immune than traditional methods.Since only the use of two-dimensional CNN does not effectively consider the relationship between ECG leads,and the size of the convolution kernel is single,resulting in low accuracy,it has become a challenge to design an algorithm with better performance for ECG arrhythmia detection.Aiming at the shortcomings of manual detection of arrhythmia and two-dimensional CNN,as well as the characteristics of 12 lead ECG signal,this paper proposes an automatic arrhythmia classification method based on multi-scale fusion convolution neural network structure to detect and classify 12 lead ECG.The classification target has a total of 9 categories,including normal ECG and ECG of 8 different arrhythmia diseases.The main research contents of the paper are summarized as follows:1.By observing the electrocardiogram of the 12 lead,its tip and waveform have similarities and differences,in response to this phenomenon,a convolutional neural network structure integrating 1D and 2D is proposed,in which 2D convolutional neural network convolutes each lead to highlight the difference between 12 leads,1D convolutional neural network convolutes the whole ECG signal to highlight the similarity between 12 leads.2.The ECG signal data length value of the 12 lead used in this paper is relatively large,in view of this phenomenon,a multi-scale fusion method is used to capture the information contained in the ECG signal,specifically,different sizes are used in the convolution process.Finally,the average F1-score of the method can reach 80.3%in this paper.