12-lead resting electrocardiogram(ECG) is an important method for clinical diagnosis of cardiovascular disease. There are hundreds of millions of electrocardiogram examinations every year worldwide. However, accurate analysis of ECG requires very specialized experience, which leads to a shortage of doctors who are proficient in ECG analysis, especially in areas with poor medical conditions. Although there are many methods based on deep learning to detect ECG abnormalities, such a problem is still a huge challenge due to the variety of ECG morphological changes. Some abnormalities may last only for a short segment, while others may be longer and even last for the entire record. To address this issue, we propose an attention-based multi-scale convolutional neural network(AMS-CNN), which can effectively extract multi-scale features and focus on abnormal ECG segments. Although our model is currently not performing well (Team Name: DeepECG, Rank: 228, F_2 Score: 0.555, G_2 Score:0.333, Geometric Mean: 0.430), we believe that with subsequent improvements, it can achieve more excellent performance.