The electrocardiogram (ECG) has been widely used to diagnose a variety of cardiac abnormalities. However, the automatic detection of cardiac ab-normalities based on reduced-lead ECGs is difficult to demonstrate the early and correct diagnosis of cardiovascular diseases. We have developed an en-semble machine learning model to detect and classify cardiac abnormalities from either twelve-lead, six-lead, three-lead, and two-lead ECGs. The basic classifier models extract convolution features from ECG records by using a convolution neural network and combine bi-directional gated recurrent unit (GRU) for temporal aggregation of features. Especially, the self-attention (SA) is introduced to prevent the loss of information in a long-time sequence. Five folds cross-validation is used to produce five best validation models as the basic models. Meanwhile, a framework for joint iterative optimization of parameters and labels is proposed to solve the problem of noisy labels result-ing from the preprocess of ECGs segmentation to fixed length. During the training procedure, we introduce a data augmentation scheme of over-sampling to deal with the class imbalance issue. Finally, we ensemble models of different reduced-lead ECGs to predict cardiac abnormalities. By cross validating on the training set, it achieved an average score of 0.538 in the 12 leads test set, 0.531 in the 6 leads set, 0.527 in the 3 leads set, 0.524 in the 2 leads set, respectively. The name of Our team is F-team. Unfortunately, be-cause the data type of binary output is set as float instead of bool, it cannot be recognized by the evaluation function. The score from the unofficial phase is -0.406. The true accuracy of our model has not been present.