Electrocardiogram (ECG) is widely used in the clinical diagnosis of arrhythmia, numerous algorithms based on ECGs and deep learning have been proposed in recent years to achieve automatic arrhythmia detection. In this challenge, we designed a novel algorithm combining the residual convolutional neural network (RCNN) and gated recurrent unit (GRU). Specifically, we first extracted up to 60 seconds of record, considering that the shortest record is 6s in the dataset, we set a constant window with a length of 6s and constructed a total of 10 segments for each record (if the current record is less than 60s, there was overlap with equal size between all adjacent segments), and then each segment was separately input into a single shared-weight RCNN model. The RCNN model was consist of 9 convolutional layers with RELU activation and spatial dropout strategy, the introduction of residual connections was used to alleviate the problem of gradient vanishing, and the shared-weight strategy was adopted to reduce the number of parameters and enhance model generalization. A 1-layer bidirectional GRU network was then stacked to further integrate the information obtained in 10 time steps. During the model training, 2 data augmentation algorithms were designed to reduce over-fitting. The first one was randomly compressing the signal (ratio: 1.1-1.5) and adding Gaussian white noise, the second one was randomly removing N 50ms fragments from the signal and adding Gaussian white noise. Finally, we trained 10 diverse models under the aforementioned framework with slightly adjusting (kernel size, the number of layers, etc.), and the simple averaging results were used as final prediction probability for 9 classes, a threshold of 0.2 was applied to get binary labels. The best F_2 and G_2 score our team (MetaHeart) have obtained in the unofficial phase were 0.818 and 0.602, and the geometric mean value was 0.702.