PearsonUnet: ECG Classification Model Integrating Pearson Related Beats

Cardiovascular disease is a significant cause of global mortality, and electrocardiogram (ECG) is a commonly used clinical tool for detecting heart health and cardiovascular disease. Considering the discontinuity of ECG signal leads, the long temporal nature within leads, and the intermittency of disease outbreaks, this article proposes an ECG classification model that integrates Pearson related beats. In response to the intermittent characteristics of ECG signal disease outbreaks, this model calculates the Pearson correlation between ECG segments to obtain the least correlated segment in ECG data, in order to find a special rhythm (possible onset rhythm). Second, in order to enhance the features of special beats, the model uses a parallel-based hole Unet to fuse special beats with raw data to obtain enhanced local features of the data. In response to the long temporal characteristics of ECG signals, this model uses a transformer model that integrates local feature prediction to obtain the long sequence features of ECG. We evaluated the proposed method on the China Physical Signal Challenge 2018 (CPSC2018) dataset and physikalisch-technische bundesanstalt extra large (PTB-XL) dataset, where the average $F1$ for nine types of arrhythmia diseases on the CPSC2018 dataset was 0.84 and the average $F1$ for five super types and 23 subtypes of diseases on the PTB-XL dataset was 0.812 and 0.486, respectively.