Aims: This study aimed to design a well-performing model to help physicians better diagnose heart disease with a 12-lead electrocardiogram(ECG). Methods: This study designed a multi-scale model named SE-ECGNet which was modified on ResNet. In this model, three paths were designed to extract the features of ECGs. Paths had a different convolution kernel size, which was 3,5 and 7, respectively. Such multi-scale design enabled the network to obtain different receptive fields and capture information at different scales, which significantly improved the classification effect. And SE-Net was added to every path of the model. The attention mechanism of SE-Net learned feature weights according to the loss, which made the effective feature maps have significant weights and the ineffective or low-effect feature maps have small weights. Since there is a certain relationship between heart disease and age and sex, for example, the premature atrial complex is more common in middle-aged and elderly people, the depth characteristics obtained by the model and the age and sex characteristics of patients were combined and input into the fully connected layer to obtain the classification results. This study designed a special loss function, which was a combination of F_β loss and weighted binary cross-entropy loss. The weight of each disease on the loss was set according to the number of samples of each disease to solve the imbalance of data. Results: Our team name is CQUPT_ECG. This study used 5 fold cross-validation, and the F-score (F_β score) of the cross-validation is 0.79, the Jaccard measure value(G_β score) is 0.57, and the total score is 0.68. The F_β score of the corresponding online submission is 0.78, the G_β score is 0.55, and the total score is 0.66.