Abstract:  Cardiovascular disease, with the highest mortality rate across the globe, kills over ten million people every year. Thanks to the continuous development of artificial intelligence, patients’ heart conditions can now be quickly and accurately diagnosed with the assistance of automatic electrocardiogram anomaly classification technology. This paper proposes an electrocardiogram classification algorithm based on CPSC-2018 twelve lead data, which combines U-Net network and attention mechanism. First, the data are processed for equal length and normalization to address their varied lengths. Then, the preprocessed data with longer lengths are reprocessed by the skip layer connection and encoding and decoding methods in the U-Net network. An attention mechanism is added to the last layer of U-Net network decoding to combat noise and improve the effective information attention and accuracy of the model. Finally, CPSC-2018 dataset is employed to verify the model. Our experimental results show the proposed model delivers fairly satisfying classification performance, recording its accuracy, recall, and F1 values of over 90% in identifying atrial fibrillation (AF) and right bundle branch block (RBBB) arrhythmias, and an average F1 value of 82.5%.