PRA-Net: Polymorphic Representation-Aware Network for Accurate Cardiac Arrhythmia Detection From Electrocardiogram Signals

Electrocardiogram (ECG) arrhythmia classification is of great significance for cardiologist-level disease diagnosis. However, it is challenging to perform cardiac arrhythmia detection because of these limitations: 1) the abnormal ECG signals vary in different scales; 2) the temporal and spatial dependencies in ECG sequences are always ignored; and 3) the waveform semantic information, for example, wavelet coefficients, higher-order statistics, and slope relationship are often missing. To address these issues, a polymorphic representation-aware network (PRA-Net) is proposed for cardiac arrhythmia detection from ECG signals, where the adaptive wavelet-aware (AWA) module is developed to identify ECG signals at different scales by learning multiscale features, while the adaptive temporal-space-aware (ATA) module is designed to model the temporal and positional dependencies in the sequence, and the collaborative semantic-aware (CSA) module is devised to encode semantic information from waveforms, finally, the adaptive feature attention fusion (AFAF) module is devised to select effective information for cardiac arrhythmia classification. Experiments demonstrate that the proposed PRA-Net outperforms the state-of-the-art methods, achieving average areas under the receiver-operating characteristic curve (AUC) of 99.23% and 98.86% on the CinC17 and MIT-BIH datasets, respectively.