MSGNet: Multi-Scale Grid Based Automatic Cardiac Arrhythmia Detection in 12-Lead ECG

The electrocardiogram (ECG) is a commonly used medical diagnostic tool for detecting various cardiac arrhythmias. Abnormal ECG signals are identified by distorted heartbeat morphology and irregular intervals. Traditional ECG analysis methods mainly rely on single-lead or single-scale signal segments, which overlook the potentially complementary information across the 12 leads and different scales. In this paper, we propose a novel Multi-Scale Grid based Network (MSGNet) for automatic cardiac arrhythmia detection in 12-lead ECG. MSGNet can extract spatial features of 12-lead ECG signals on different channels and temporal features of different scales on the same channel to effectively capture the features of distorted heartbeat morphology and irregular intervals. By fusing the morphological features of different channels, MSGNet extracts more diverse features from different spatial dimensions. Furthermore, we designed a multi-scale grid based feature extraction strategy to extract features of signal segments of various sizes at different scales. MSGNet integrates these two feature extraction strategies to simultaneously focus on both the morphological features of different leads and the temporal features within the same lead. We evaluated the performance of MSGNet on the publicly available ECG dataset CPSC 2018 and compared it with other existing ECG classification models. The experimental results show that MSGNet outperforms other existing ECG classification models, achieving an F1 score of 0.858 on this dataset.