Abstract: Electrocardiogram (ECG) is important for clinical diagnosis of cardiovascular disease. Extracting effective
features is the key to automatic diagnosis system. Nowadays, most research works only use single-lead or two-lead
signals, which makes the extracted features too incomprehensive to achieve excellent classification performance. In
order to obtain more comprehensive features and better classification results, we proposed a multi-lead heartbeat
signal classification method based on multi-scale feature extraction (MSNet). Firstly, the method received multi-lead
stacked heartbeats signal matrix as input. Then, three feature vectors with different scale information were extracted
by convolution layers. Finally, classification was designed to decide to which category the input heartbeat belonged
based on fused features. Three open source databases including MIT-BIH Arrhythmia Database, MIT-BIH
Supraventricular Arrhythmia Database, and St Petersburg INCART 12-Lead Arrhythmia Database were utilized in this research. Experiments on these databases with 5-fold cross-validation strategy show the excellent performance of the
proposed method. For “normal-abnormal” classification, more than 99% of the accuracy, recall, precision, and F1
score can be achieved. For “multi-category” classification, 99.5% of the average accuracy, average recall, average
precision, and average F1 score can be achieved. Compared with the state-of-the-art methods, the proposed method
exhibits better performance.

Key words: heartbeat classification, multi-scale feature extraction, feature fusion