Multi-label Classification of ECG using 1-D ResNet

Wei Zhao


Aims: In this study, we aimed to build an effective model that can automatically detect cardiac abnormalities from 12-lead ECG recordings using 1D RestNet. Methods: Data preprocessing: We divided along ECG signal into several signal segments with the same length. Since the ECG data contains 3,000 to 72,000 sampling points in each lead, we unified the length of the input vector to 10,000 by padding the shorter data with the wrap of sampling points and sliding the longer data with stride of 250. In this way, the training data were augmented by more than 5 times. Then signal segments from a long ECG signal were projected into 1D Resnet model. Model architecture: Resnet is a widely used model for 2D-image classification in many computer vision tasks, which cannot be directly applied in our study. Therefore, we modified and addressed this issue by proposing a 1D Resnet for automatic classification of ECG signals. Our 1D Resnet model was designed to address the degradation problem in deep neural networks by introducing a residual block. In particular, 7 residual blocks were connected while each contained two 1D convolutional layers with multiple 15-length convolution kernels. The results from the residual blocks was then flattened and received by two fully-connected layers to output a 9-dimensional tensor which indicated the probability of multiple clinical diagnoses. To mitigate the effects of overfitting, we applied a dropout layer at the end. We measured the loss of the multi-label task by cross entropy after sigmoid activator. Results: On the released data, the 5-fold cross validation results were F2:0.710, G2:0.489, Score:0.589. On the hidden test data, our team, BraveHeart400, received F2:0.656, G2:0.438, Score:0.536. Conclusion: We propose an automatic classification on ECGs using 1D ResNet. The preliminary results from our model demonstrated a good performance in cardiac abnormality identification.